3 540 47952 X

Lecture Notes in Computer Science 2347
Edited by G. Goos, J. Hartmanis, and J. van Leeuwen

3
Berlin
Heidelberg
New York
Barcelona
Hong Kong
London
Milan
Paris
Tokyo
Paul De Bra Peter Brusilovsky
Ricardo Conejo (Eds.)
Adaptive Hypermedia
and Adaptive
Web-Based Systems
Second International Conference, AH 2002
Málaga, Spain, May 29-31, 2002
Proceedings
13
Gerhard Goos, Karlsruhe University, Germany
Juris Hartmanis, Cornell University, NY, USA
Jan van Leeuwen, Utrecht University, The Netherlands
Volume Editors
Paul De Bra
Eindhoven University of Technology
Department of Computer Science
P.O. Box 513, 5600 MB Einhoven, The Netherlands
E-mail: [email protected]
Peter Brusilovsky
University of Pittsburgh, School of Information Sciences
Department of Information Science and Telecommunications
135 North Bellefield Avenue, Pittsburgh, PA 15260, USA
Ricardo Conejo
Universidad de Málaga, Escuela Tecnica Superior de Ingenieros en Informatica
Departamento de Lenguajes y Ciencias de la Computacion
Boulevard Luis Pasteur, 35, Málaga, Spain
Cataloging-in-Publication Data applied for
Die Deutsche Bibliothek - CIP-Einheitsaufnahme

Adaptive hypermedia and adaptive Web based systems : second international
conference ; proceedings / AH 2002, Malaga, Spain, May 29 - 31, 2002. Paul
de Bra ... (ed.). - Berlin ; Heidelberg ; New York ; Barcelona ; Hong Kong ;
London ; Milan ; Paris ; Tokyo : Springer, 2002
(Lecture notes in computer science ; Vol. 2347)
ISBN 3-540-43737-1
CR Subject Classification (1998): H.5.4, H.4, H.5, H.3
ISSN 0302-9743
ISBN 3-540-43737-1 Springer-Verlag Berlin Heidelberg New York
This work is subject to copyright. All rights are reserved, whether the whole or part of the material is
concerned, specifically the rights of translation, reprinting, re-use of illustrations, recitation, broadcasting,
reproduction on microfilms or in any other way, and storage in data banks. Duplication of this publication
or parts thereof is permitted only under the provisions of the German Copyright Law of September 9, 1965,
in its current version, and permission for use must always be obtained from Springer-Verlag. Violations are
liable for prosecution under the German Copyright Law.
Springer-Verlag Berlin Heidelberg New York
a member of BertelsmannSpringer Science+Business Media GmbH
http://www.springer.de
© Springer-Verlag Berlin Heidelberg 2002

Printed in Germany
Typesetting: Camera-ready by author, data conversion by Steingräber Satztechnik GmbH, Heidelberg
Printed on acid-free paper SPIN 10869993 06/3142 543210
Preface
Hypermedia as a means for accessing information was first proposed by Vannevar

Bush, in 1945, in his memorable article “As We May Think”. The first research
that resembles what we consider as hypermedia (i.e. at least using computers
with textual or graphical displays) was started much later, by great minds in-
cluding Doug Engelbart (known for many inventions including the mouse) and
Ted Nelson (known for inventing Xanadu, an architecture similar but superior
to that of the World Wide Web). After 20 years of hypermedia research the
Web started and really changed hypermedia from a small research community
and mostly small stand-alone applications to a world-wide network of intercon-
nected information sources. Hypermedia applications serving a large audience
suffer greatly from the traditional “one-size-fits-all” approach of static websites.
Starting in the 1990s, and essentially in parallel with the development of the
Web, many research teams began to investigate ways of modeling features of the
users of hypermedia systems. This has led to a number of interesting adaptation
techniques and adaptive hypermedia systems. Nowadays adaptive hypermedia
and the Web have really come together. Virtually all adaptive systems use the
Web for both the (browser) front end and the server back end.
Following a number of successful workshops on adaptive hypermedia, the
first conference on Adaptive Hypermedia and Adaptive Web-Based Systems
(AH 2000) was held in Trento, Italy, in August 2000. (See Springer Lecture Notes
in Computer Science, vol. 1892.) The research community has continued to grow.
The Second International Conference on Adaptive Hypermedia and Adaptive
Web-Based Systems (AH 2002) attracted twice as many full paper submissions,
and also a large number of short papers describing ongoing research.
AH 2002 continued the tradition of earlier workshops, and the AH 2000 con-
ference, in being a meeting place for researchers with very different backgrounds
such as hypertext, user modeling, machine learning, natural language genera-
tion, information retrieval, intelligent tutoring systems, cognitive science, and
Web-based education. In 2001 the community gathered at the User Modeling
Conference in Sonthofen, Germany and at the ACM Hypertext Conference in
Århus, Denmark for a two-session workshop on adaptive hypermedia, reported
on in LNCS vol. 2266. A lot of research that was ongoing in 2001 can now be
found in the full papers of the AH 2002 conference proceedings. A lot of new
research is presented in short papers and posters.
The success of the conference generated a tremendous amount of work for the
Program Committee members and the External Reviewers. We are immensely
greatful for the effort they put into the process of selecting the very best papers.
This has resulted in 33 full-paper presentations (selected from 109 submitted), 23
short-paper presentations, 30 posters, and 5 presentations at the Doctoral Con-
sortium. All presentations can be found in this volume (except for two keynote
talks, where we only have an abstract).
VI Preface
The success also generated an incredible amount of work for the program
chair’s team in Eindhoven (The Netherlands). The timely production of these
proceedings from 94 contributions would not have been possible without the help
of Lora Aroyo and Alexandra Cristea who worked almost round the clock to fix
formatting and English language problems.
We gratefully acknowledge the help from AH 2002 cooperative societies, in-
stitutes, and sponsors. First of all we thank the University of Málaga, and espe-
cially the School of Informatics, that kindly hosted the conference. Thanks also
for the economic contribution of Unicaja, and the support of the city institutions,
the Excm. Ayuntamiento de Málaga, and the Exma. Diputacion Provincial de
Málaga. Their sponsorship was very important. We thank doctor Chen’s family,
Springer-Verlag, and Kluwer for the best paper awards they supported. We also
thank the Organizing Committee in Málaga for their support with web pages,
scheduling, and all the arrangement they made. We would also like to thank to
the cooperative societies: the Association for Computing Machinery (ACM) and
the Special Interest Groups SIGART, SIGCHI, SIGIR, SIGWEB, SIGECOM;
the Asociación Española Para la Inteligencia Artificial (AEPIA), the European
Coordinating Committee for Artificial Intelligence (ECCAI), the EUropean Net-
work on Intelligent TEchnologies for Smart Adaptive Systems (EUNITE), the
International Artificial Intelligence in Education Society (IAED), User Modeling
Inc. (UMInc.), and the International World Wide Web Conference Committee,
that endorsed this conference.
The University of Málaga also provided an excellent local organization com-
mittee, which handled mailings, registrations, and all local arrangements.
Last but not least, we would like to thank the more than 200 authors from
over 30 different countries who contributed to this book and conference. We
thank them, not only for their valuable contribution to this research field, but
also for their effort in supplying us with all the necessary information in a timely
fashion.
May 2002 Paul De Bra

Peter Brusilovsky
Ricardo Conejo
Organization
AH 2002 was organized by the University of Málaga, Spain.
General Chairs
Peter Brusilovsky (University of Pittsburgh)
Francisco Triguero (Universidad de Málaga)
Program Chair
Paul De Bra (Technische Universiteit Eindhoven)
Local Organization Chair

Ricardo Conejo (Universidad de Málaga)
Doctoral Consortium Chairs

José-Luı́s Pérez de la Cruz (Universidad de Málaga)
Wolfgang Nejdl (Universität Hannover)
Workshops/Tutorials Chair
Carlo Strapparava (ITC-IRST, Trento)
Industry Chair
Barry Smyth (Changingworlds, and University College Dublin)
Program Committee
Elisabeth André, DFKI GmbH
Liliana Ardissono, Università degli Studi di Torino
Alexandra Cristea, Technische Universiteit Eindhoven
Fiorella De Rosis, Università di Bari
Franca Garzotto, Politecnico di Milano
Wendy Hall, University of Southampton
Lynda Hardman, CWI, Amsterdam
VIII Organization
Nicola Henze, Institut für Technische Informatik, Hannover

Kristina Höök, Swedish Institute for Computer Science
Lewis Johnson, University of Southern California
Judy Kay, University of Sydney
Alfred Kobsa, University of California at Irvine
Paul Maglio, IBM Almaden Research Center
Mark Maybury, The MITRE Corporation
Alessandro Micarelli, Università di Roma 3
Eva Millán, Universidad de Málaga
Maria Milosavljevich, Dynamic Multimedia Pty. Ltd.
Tanja Mitrovic, University of Canterbury, New Zealand
Dunja Mladenic, J. Stefan Institute, Slovenia
Marc Nanard, LIRMM
Vittorio Scarano, Università degli Studi di Salerno
Daniel Schwabe, Universidade Católica do Rio de Janeiro
Jude Shavlik, University of Wisconsin
Marcus Specht, GMD
Myra Spliopoulou, Handelshochschule Leipzig
Oliviero Stock, ITC-IRST, Trento
Carlo Tasso, Università degli Studi di Udine
Julita Vassileva, University of Saskatchewan
Gerhard Weber, Pedagogical University Freiburg
Massimo Zancanaro, ITC-IRST, Trento
External Reviewers
Rogelio Adobbati Ralph Deters Paola Rizzo
Demosthenes Josef Fink Ralph Schaefer
Akoumianakis Milos Kravcik Erin Shaw
Rudström Åsa Donato Malerba Maja Skrjanc
Janez Brank Lawrence Mandow Carlo Strapparava
Susan Bull Martin Müller Martin Svensson
Giovanna Castellano Leonid Pessine Stephan Weibelzahl
Ricardo Conejo Francesco Ricci
Berardina De Carolis Thomas Rist
Local Organization Committee

Marı́a Victoria Belmonte
David Bueno
Eduardo Guzmán
Eva Millán
Rafael Morales
José Luı́s Pérez de la Cruz
Mónica Trella
Table of Contents
Invited Papers
Is Personalization All About Technology? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
J. Aaronson
Adaptive Linking between Text and Photos Using Common Sense Reasoning 2
H. Lieberman, H. Liu
Resource-Adaptive Interfaces to Hybrid Navigation Systems . . . . . . . . . . . . . 12

W. Wahlster
Full Papers
Ubiquitous User Assistance in a Tourist Information Server . . . . . . . . . . . . . 14
L. Ardissono, A. Goy, G. Petrone, M. Segnan, P. Torasso
Automatic Extraction of Semantically-Meaningful Information

from the Web. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
J.L. Arjona, R. Corchuelo, A. Ruiz, M. Toro
Towards Open Adaptive Hypermedia . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

C. Bailey, W. Hall, D.E. Millard, M.J. Weal
GAS: Group Adaptive System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

M. Barra, P. Maglio, A. Negro, V. Scarano
TV Scout: Lowering the Entry Barrier

to Personalized TV Program Recommendation . . . . . . . . . . . . . . . . . . . . . . . . 58
P. Baudisch, L. Brueckner
Adaptivity, Adaptability, and Reading Behaviour:

Some Results from the Evaluation of a Dynamic Hypertext System . . . . . . . 69
K. Bontcheva
Towards Generic Adaptive Systems: Analysis of a Case Study . . . . . . . . . . . 79

L. Calvi, A. Cristea
A Methodology for Developing Adaptive Educational-Game Environments 90

R.M. Carro, A.M. Breda, G. Castillo, A.L. Bajuelos
Multi-model, Metadata Driven Approach to Adaptive Hypermedia Services

for Personalized eLearning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
O. Conlan, V. Wade, C. Bruen, M. Gargan
X Table of Contents
Adaptation and Personalization on Board Cars:

A Framework and Its Application to Tourist Services . . . . . . . . . . . . . . . . . . 112
L. Console, S. Gioria, I. Lombardi, V. Surano, I. Torre
Adaptive Authoring of Adaptive Educational Hypermedia . . . . . . . . . . . . . . . 122

A. Cristea, L. Aroyo
Hypermedia Presentation Adaptation on the Semantic Web . . . . . . . . . . . . . 133

F. Frasincar, G.-J. Houben
User Data Management and Usage Model Acquisition

in an Adaptive Educational Collaborative Environment . . . . . . . . . . . . . . . . . 143
E. Gaudioso, J.G. Boticario
Personalizing Assessment in Adaptive Educational Hypermedia Systems . . 153

E. Gouli, K. Papanikolaou, M. Grigoriadou
Visual Based Content Understanding towards Web Adaptation . . . . . . . . . . 164

X.-D. Gu, J. Chen, W.-Y. Ma, G.-L. Chen
Knowledge Modeling for Open Adaptive Hypermedia . . . . . . . . . . . . . . . . . . . 174

N. Henze, W. Nejdl
Adaptive Navigation for Learners in Hypermedia Is Scaffolded Navigation . 184

R. Hübscher, S. Puntambekar
Pros and Cons of Controllability: An Empirical Study . . . . . . . . . . . . . . . . . . 193

A. Jameson, E. Schwarzkopf
Personis: A Server for User Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203

J. Kay, B. Kummerfeld, P. Lauder
The Munich Reference Model for Adaptive Hypermedia Applications . . . . . 213

N. Koch, M. Wirsing
Tracking Changing User Interests through Prior-Learning of Context . . . . . 223

I. Koychev
Prediction of Navigation Profiles in a Distributed Internet Environment

through Learning of Graph Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
D. Kukulenz
Educo – A Collaborative Learning Environment

Based on Social Navigation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 242
J. Kurhila, M. Miettinen, P. Nokelainen, H. Tirri
GOOSE: A Goal-Oriented Search Engine with Commonsense . . . . . . . . . . . . 253

H. Liu, H. Lieberman, T. Selker
Table of Contents XI
On Adaptability of Web Sites for Visually Handicapped People . . . . . . . . . . 264

M. Macı́as, J. González, F. Sánchez
A Framework for Filtering and Packaging Hypermedia Documents . . . . . . . 274

L.C. Martins, T.A.S. Coelho, S.D.J. Barbosa, M.A. Casanova,
C.J.P. de Lucena
Adaptation in an Evolutionary Hypermedia System:

Using Semantic and Petri Nets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 284
N. Medina-Medina, L. Garcı́a-Cabrera, M.J. Rodrı́guez-Fortiz,
J. Parets-Llorca
Evaluating the Effects of Open Student Models on Learning . . . . . . . . . . . . . 296

A. Mitrovic, B. Martin
Ephemeral and Persistent Personalization in Adaptive Information Access

to Scholarly Publications on the Web . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 306
S. Mizzaro, C. Tasso
Fuzzy Linguistic Summaries in Rule-Based Adaptive Hypermedia Systems 317

M.-Á. Sicilia, P. Dı́az, I. Aedo, E. Garcı́a
The Plight of the Navigator:

Solving the Navigation Problem for Wireless Portals . . . . . . . . . . . . . . . . . . . . 328
B. Smyth, P. Cotter
Towards an Adaptive Web Training Environment

Based on Cognitive Style of Learning: An Empirical Approach . . . . . . . . . . . 338
M.A.M. Souto, R. Verdin, R. Wainer, M. Madeira, M. Warpechowski,
K. Beschoren, R. Zanella, J.S. Correa, R.M. Vicari, J.P.M. de Oliveira
Automated Personalization of Internet News . . . . . . . . . . . . . . . . . . . . . . . . . . 348

A.V. Sunderam
Short Papers
Conceptual Modeling of Personalized Web Applications . . . . . . . . . . . . . . . . 358
S. Abrahão, J. Fons, M. González, O. Pastor
On Evaluating Adaptive Systems for Education . . . . . . . . . . . . . . . . . . . . . . . . 363

R. Arruabarrena, T.A. Pérez, J. López-Cuadrado, J. Gutiérrez,
J.A. Vadillo
Recommending Internet-Domains Using Trails and Neural Networks . . . . . . 368

T. Berka, W. Behrendt, E. Gams, S. Reich
Learning Grammar with Adaptive Hypertexts: Reading or Searching? . . . . 372

A. Brunstein, J. Waniek, A. Naumann, J.F. Krems
XII Table of Contents
SIGUE: Making Web Courses Adaptive . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 376

C. Carmona, D. Bueno, E. Guzman, R. Conejo
An Ontology-Guided Approach to Content Adaptation in LEO:

A Learning Environment Organizer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 380
J.W. Coffey
A Scrutable Adaptive Hypertext . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 384

M. Czarkowski, J. Kay
AHA! Meets AHAM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 388

P. De Bra, A. Aerts, D. Smits, N. Stash
Adaptive Real Time Comment Generation for Sail Racing Scenarios . . . . . . 392
A. Esuli, A. Cisternino, G. Pacini, M. Simi
A Web-Based Selling Agent That Maintains Customer Loyalty

through Integrative Negotiation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 397
M. Grimsley, A. Meehan
Adaptive Content
for Device Independent Multi-modal Browser Applications . . . . . . . . . . . . . . 401
J. Healey, R. Hosn, S.H. Maes
Behavioral Sequences: A New Log-Coding Scheme

for Effective Prediction of Web User Accesses . . . . . . . . . . . . . . . . . . . . . . . . . . 406
R. Kanawati, M. Malek
A Fuzzy-Based Approach to User Model Refinement

in Adaptive Hypermedia Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 411
C. Martinovska
Towards an Authoring Coach for Adaptive Web-Based Instruction . . . . . . . 415

J. Masthoff
Generation of Personalized Web Courses Using RBAC . . . . . . . . . . . . . . . . . . 419

S. Montero, I. Aedo, P. Dı́az
An Automatic Rating Technique Based on XML Document . . . . . . . . . . . . . 424

H. Mun, S. Ok, Y. Woo
Using Effective Reading Speed

to Integrate Adaptivity into Web-Based Learning . . . . . . . . . . . . . . . . . . . . . . 428
M.H. Ng, W. Hall, P. Maier, R. Armstrong
A Non-invasive Cooperative Student Model . . . . . . . . . . . . . . . . . . . . . . . . . . . 432

T. Roselli, E. Gentile, P. Plantamura, V. Rossano, V. Saponaro
Table of Contents XIII
Category-Based Filtering in Recommender Systems

for Improved Performance in Dynamic Domains . . . . . . . . . . . . . . . . . . . . . . . 436
M. Sollenborn, P. Funk
Protecting the User from the Data:

Security and Privacy Aspects of Public Web Access . . . . . . . . . . . . . . . . . . . . 440
A. Spalka, H. Langweg
Exploiting a Progressive Access Model to Information . . . . . . . . . . . . . . . . . . 444

M. Villanova-Oliver, J. Gensel, H. Martin
Adapting to Prior Knowledge of Learners . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 448

S. Weibelzahl, G. Weber
Monitoring Partial Updates in Web Pages Using Relational Learning . . . . . 452

S. Yamada, Y. Nakai
Posters
Adaptation in the Web-Based Logic-ITA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 456
D. Abraham, K. Yacef
Collaborative Radio Community . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 462

P. Avesani, P. Massa, M. Nori, A. Susi
Improving Interactivity in e-Learning Systems

with Multi-agent Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 466
R.A. Silveira, R.M. Vicari
Adaptive Hypermedia Made Simple

with HTML/XML Style Sheet Selectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 472
F. Bry, M. Kraus
A Framework for Educational Adaptive Hypermedia Applications . . . . . . . . 476

F. Buendı́a, P. Dı́az
METIORE: A Publications Reference

for the Adaptive Hypermedia Community . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 480
D. Bueno, R. Conejo, C. Carmona, A.A. David
User-Driven Adaptation of e-Documents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 484

P. Carrara, D. Fogli, G. Fresta, P. Mussio
Server Independent Personalization for Effective Searching

in the World Wide Web . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 488
L.N. Cassel, U. Wolz, R.E. Beck
XIV Table of Contents
Preventing Misleading Presentations of XML Documents:

Some Initial Proposals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 492
A. Cawsey, E. Dempster, D. Bental, D. Pacey, H. Williams,
L. MacKinnon, D. Marwick
Adaptivity Conditions Evaluation for the User of Hypermedia Presentations

Built with AHA! . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 497
A. Cini, J. Valdeni de Lima
Development of Adaptive Web Sites

with Usability and Accessibility Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 501
M. Fernández de Arriba, J.A. López Brugos
An Adaptive e-Commerce System Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . 505

F.J. Garcı́a, F. Paternò, A.B. Gil
Web Site Personalization Using User Profile Information . . . . . . . . . . . . . . . . 510

M. Goel, S. Sarkar
Who Do You Want to Be Today?

Web Personae for Personalised Information Access . . . . . . . . . . . . . . . . . . . . . 514
J.P. McGowan, N. Kushmerick, B. Smyth
Adaptive Navigation Path Previewing for Learning on the Web . . . . . . . . . . 518

A. Kashihara, S. Hasegawa, J. Toyoda
A Case-Based Recommender System Using Implicit Rating Techniques . . . 522

Y. Kim, S. Ok, Y. Woo
IMAP - Intelligent Multimedia Authoring Tools for Electronic Publishing 527

S. Kraus, A. Kröner, L. Tsaban
A Hybrid Recommender System

Combining Collaborative Filtering with Neural Network . . . . . . . . . . . . . . . . 531
M. Lee, P. Choi, Y. Woo
An Adaptive Cooperative Web Authoring Environment . . . . . . . . . . . . . . . . . 535

A.M. Martı́nez-Enrı́quez, D. Decouchant, A.L. Morán, J. Favela
Adapting Learner Evaluation by Plans . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 539

A. Martens, A.M. Uhrmacher
WETAS: A Web-Based Authoring System for Constraint-Based ITS . . . . . . 543

B. Martin, A. Mitrovic
Open-Ended Adaptive System for Facilitating Knowledge Construction

in Web-Based Exploratory Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 547
H. Mitsuhara, Y. Ochi, Y. Yano
Table of Contents XV
Adapting Web Interfaces By WHAT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 551

W. Ng, J. Yau
Considering Sensing-Intuitive Dimension to Exposition-Exemplification

in Adaptive Sequencing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 556
P. Paredes, P. Rodriguez
PlanEdit: An Adaptive Problem Solving Tool for Design . . . . . . . . . . . . . . . . 560

M.A. Redondo, C. Bravo, M. Ortega, M.F. Verdejo
Dynamic Generation of Interactive Dialogs Based on Intelligent Agents . . . 564

M. González Rodrı́guez, B. López Pérez, M.P. Paule Ruı́z,
J.R. Pérez Pérez
A Simple Architecture for Adaptive Hypertext . . . . . . . . . . . . . . . . . . . . . . . . 568
O. Signore
Adaptive Learning Environment for Teaching and Learning in WINDS . . . . 572
M. Specht, M. Kravcik, R. Klemke, L. Pesin, R. Hüttenhain
Adaptation through Interaction Modelling in Educational Hypermedia . . . . 576

E. Toppano
The Hybrid Model for Adaptive Educational Hypermedia . . . . . . . . . . . . . . . 580

M.R. Zakaria, A. Moore, H. Ashman, C. Stewart, T. Brailsford
Doctoral Consortium
Navigation Modelling in Adaptive Hypermedia . . . . . . . . . . . . . . . . . . . . . . . . 586
P. Dolog, M. Bieliková
An Adaptive Agent Model for e-Commerce Architecture . . . . . . . . . . . . . . . . 592
A.B. Gil, F. Garcı́a, Z. Guessoum
Personalized Adaptation to Device Characteristics . . . . . . . . . . . . . . . . . . . . . 598

E. Herder, B. van Dijk
Users Modeling for Adaptive Call Centers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 603

I. Torre
The Exploratory of Personal Assistants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 608

L.T. Yong, T.E. Kong
Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 613

Is Personalization All About Technology?
(Keynote Talk)
Jack Aaronson
Jack Aaronson Consulting

666 Greenwich Street, Suite 409
New York, New York 10014
[email protected]
Abstract. A lot of research and development has pushed the limits of personal-
ization far beyond where they were just a few years ago. Yet the major goals of
personalization continue to elude us: serving customers what they want, when
they want it. Is all this research being developed with pure science in mind, or
do the new technologies actually solve real business needs? In this lecture, we
will discuss the business needs around personalization. We will explore which
technologies have been effective, and which haven't been. We will discuss the
importance of mixing technology with smart design to ensure that personalized
features are intuitive and easy-to-use. We will also talk about personalization
that doesn't require technology.
P. De Bra, P. Brusilovsky, and R. Conejo (Eds.): AH 2002, LNCS 2347, p 1.

 Springer-Verlag Berlin Heidelberg 2002
Adaptive Linking between Text and Photos
Using Common Sense Reasoning
(Keynote Talk)
Henry Lieberman and Hugo Liu
MIT Media Laboratory

20 Ames St., E15-320G
Cambridge, MA 02139, USA
{lieber,hugo}@media.mit.edu
Abstract. In a hypermedia authoring task, an author often wants to set up

meaningful connections between different media, such as text and photographs.
To facilitate this task, it is helpful to have a software agent dynamically adapt
the presentation of a media database to the user's authoring activities, and look
for opportunities for annotation and retrieval. However, potential connections
are often missed because of differences in vocabulary or semantic connections
that are "obvious" to people but that might not be explicit.
ARIA (Annotation and Retrieval Integration Agent) is a software agent that
acts an assistant to a user writing e-mail or Web pages. As the user types a
story, it does continuous retrieval and ranking on a photo database. It can use
descriptions in the story to semi-automatically annotate pictures. To improve
the associations beyond simple keyword matching, we use natural language
parsing techniques to extract important roles played by text, such as "who,
what, where, when". Since many of the photos depict common everyday situa-
tions such as weddings or recitals, we use a common sense knowledge base,
Open Mind, to fill in semantic gaps that might otherwise prevent successful as-
sociations.
1 Introduction
As digital photography becomes more popular, consumers will need better ways to
organize and search their large collections of images, perhaps collected over a life-
time. Just as people compile ordinary photos into albums and scrapbooks in order to
share stories with friends and family, people will want to share stories online. It is
popular for users to engage in the hypermedia authoring task of sharing stories both
by email and through a web page. However, there are few tools available which assist
the user in their task of selecting the pictures to use to tell stories with.
ARIA [6], the software agent presented in this paper, aims to facilitate the user’s
storytelling task by observing the user as she tells a story, and opportunistically sug-
gesting photos which may be relevant to what the user is typing. When a user incor-
porates one of the system’s photo suggestions by dragging the photo into the story,
P. De Bra, P. Brusilovsky, and R. Conejo (Eds.): AH 2002, LNCS 2347, pp 2–11.

our system will automatically associate with the photo any relevant keywords and
phrases from the story context.
1.1 ARIA
The ARIA Photo Agent combines an email client or web page editor with a database
of the user’s photos, as shown in Fig. 1.
Fig. 1. A screenshot of ARIA which combines an email panel (left) with a photo database
(right) that dynamically reorders itself in real-time, as the user types
Photos are automatically imported into ARIA when a digital camera flash card is
inserted into the computer. Rather than requiring the user to organize photos into a
directory or album structure, our system tags photos with text annotations, organized
into “who, what, where, when” for each picture. The user composes an email or web
page in the text client on the left. Whenever a photo is dragged from the photo pane
into the text pane, new annotations are automatically associated with the photo.
These annotations consist of people, places, things, and events, and are extracted from
the story text adjacent to the photo in the text. Users can also edit and add to the
annotations associated with a photo by double-clicking on that photo.
4 H. Lieberman and H. Liu
As the user goes about his authoring task, the photo agent monitors what he types,
and in real-time, the agent reorders the annotated photos in the photo pane to suggest
photos that might be relevant to the user. A photo is deemed relevant if any of its
annotations can be potentially linked with the current focus of the text, either through
explicit keywords, or through a variety of semantic connections.
1.2 Our Approach
ARIA goes beyond the naïve approach of suggesting photos by a simple match be-
tween keywords in a photo’s annotations with keywords in the story. Such an ap-
proach often misses potential connections between keywords with different vocabu-
lary, or keywords that exhibit implicit semantic connectedness. By this, we mean that
it may be obvious to a person that two different keywords are conceptually related,
such as “bride” and “wedding;” however, computer programs cannot usually under-
stand such connections. Our approach remedies the problems associated with naïve
keyword matching by applying natural language parsing techniques to the annotation
process, and commonsense reasoning to the retrieval of pictures.
To address the issue of different vocabulary, we apply natural language techniques
to the annotation process such that we extract concepts rather than keywords from the
text. Unlike keywords, concepts are not sensitive to morphological variation, such as
“bridesmaids” versus “bridesmaid,” or abbreviations or near synonyms, such as “LA”
versus “Los Angeles.” We map keywords into concepts using a morphological tool
and abbreviation and near synonym dictionary.
In cases where potential connections are missed due to keywords that are related
semantically rather than explicitly, “common sense” can help. Consider a photo an-
notated with “bride.” Knowing some everyday knowledge about the world, ARIA can
infer concepts closely related to “bride” such as “groom,” “wedding,” “flower girl,”
and “wife.” These concepts are related socially, spatially, and temporally to the
original concept. Expanding the original annotation with semantically related con-
cepts gives the software agent more opportunities to recommend relevant photos to
the user.
1.3 Shaping Authoring Behavior
ARIA assists the user in her authoring task by annotating photos with descriptions
from the story, and dynamically adapting the presentation of the photos while the user
writes the story.
Our preliminary experience with ARIA shows that the opportunistic presentation
of photos can even shape the user’s authoring behavior. While a user may begin the
authoring task with a predisposition to tell the story a certain way, she may change
her mind if the agent suggests an interesting photo at an unexpected moment. This
might cause her to recall a memory, think differently, and tell the story differently.
As the story unfolds, the presentation of photos will adapt accordingly, and if the
agent suggests another interesting photo, the user may again revise her authoring
behavior, and so on. The interaction between ARIA and the user may be able to
stimulate interesting changes in the user’s authoring behavior.
This paper is structured as follows: First, we discuss the source and nature of the
corpus of common sense knowledge used by ARIA. Second, we present how natural
language processing can incorporate commonsense knowledge in the automated an-
notation mechanism. Third, we discuss how commonsense can be used with shallow
inference in the adaptive retrieval mechanism. Fourth, we compare our approach to
related work. The paper concludes with a discussion of the benefits and limitations of
our approach, and the application of commonsense to other domains.
2 Open Mind: A Source of Common Sense
The source of the commonsense knowledge used by ARIA is the Open Mind Com-
monsense Knowledge Base (OMCS) [11] – an endeavor at the MIT Media Labora-
tory that allows a web-community of teachers to collaboratively contribute to a
knowledge base of “common sense.” OMCS contains over 400,000 semi-structured
English sentences, each of which represents a simple fact about the everyday world.
Some examples of entries in the knowledge base are as follows:
1. Something you may find in a restaurant is a waiter.
2. Something that might come after a wedding is a wedding reception.
3. People get married at weddings.
OMCS is often compared with its more famous counterpart, the CYC Knowledge
Base [5]. CYC contains over 1,000,000 hand-entered rules of common sense. The
difference between OMCS and CYC is that CYC is meant for a more formal type of
reasoning using logic, while OMCS’s English sentence representation may not be
constrained enough for formal logic. Even though OMCS is noisier than CYC and
inherits the ambiguities associated with its natural language representation, it is still
suitable to our task because we only need binary semantic relations to make adaptive
linking work. This can be achieved through shallow techniques.
3 Common Sense for Parsing
When a user drags and drops a photo into the story, the description of the photo given
in the story is used to automatically annotate the photo. The annotations extracted
from the text are the semantically important concepts of person, place, thing, and
event, which can be used to answer the “who, what, and where” questions about a
photo. For the natural language parser to correctly identify these semantic types, it
needs dictionaries of concepts falling under each type.
To recognize people’s names, we obtain a dictionary of first names from the Web,
and combine that with regular expressions to recognize full names. Geographical
places are also mined from databases on the Web and added to the parser’s semantic
lexicon. As for everyday places, thing, and events, we extract dictionaries from Open
Mind. The extraction is fairly straightforward, because many of the sentence pat-
terns, or ontological relations, found in Open Mind sufficiently constrain the semantic
types of the slots.
The result is a lexicon of words and phrases with their associated semantic type.
The natural language parser uses this to enhance a syntactic parse tree with semantic
and thematic phrasal tags. The resulting tree represents an event structure using an
ontology based on the work of Jackendoff [3]. Below is an example of a sentence
and its parse.
Sentence:
Last weekend, I went to Ken and Mary's wedding in San Francisco, and I
took gorgeous pictures of the Golden Gate Bridge.
Event Structure Parse:
(ROOT (ASSERTION (TIME ARIA_DATESPAN{03m09d2002y-
03m10d2002y} ) , (ASSERTION (PERSON I ) (ACTION
went (PROPERTY to (EVENT (THING (PERSON Ken and )
(PERSON Mary 's ) ) wedding ) ) (PROPERTY in (PLACE
San Francisco ) ) ) , and (ASSERTION (PERSON I )
(ACTION took (THING (THING gorgeous pictures )
(PROPERTY of (PLACE the Golden Gate Bridge ) ) ) )
) ) . ) )
As shown in this example, knowledge mined from Open Mind and the Web allows
a semantically meaningful parse to be produced. ARIA uses heuristics to decide
which people, places, things, and events are relevant to the photo and should be used
to annotate the photo placed adjacent to this sentence in the story.
4 Commonsense Inference for Adaptive Retrieval
ARIA uses commonsense relations mined out of Open Mind to expand annotations
with semantically connected concepts that make adaptive retrieval of photos possible.
To do this, a resource was automatically constructed from Open Mind by applying
sentence patterns to the corpus, and extracting simple predicate argument structures
(usually a binary relation). Arguments are normalized into syntactically neat con-
cepts, and these concepts, together with the predicate relations, are used to construct a
spreading activation network of nodes and directed edges. The edges between the
concept nodes in the network represent the notion of semantic connectedness. The
task of expanding an annotation with its related concepts is modeled as spreading
activation over the network. Another way to think about spreading activation is as
inference directed by the strength of relations between concepts (edge weight).
In this section, we describe how a subset of the knowledge in OMCS is extracted
and structured to be useful to annotation expansion, and how spreading activation can
return semantically connected concepts. Examples of actual runs of the concept ex-
pansion are given.
4.1 Extracting Concepts and Relations from OMCS
The first step of extracting predicate argument structures from OMCS is to apply a
fixed set of mapping rules to the sentences in Open Mind. Each mapping rule cap-
tures a different commonsense relation that may be valuable to facilitating the re-
trieval task in our problem domain. The relations of interest fall under the following
general categories of knowledge:
1. Classification: A dog is a pet
2. Spatial: San Francisco is part of California
3. Scene: Things often found together are: restaurant, food, waiters, tables,
seats
4. Purpose: A vacation is for relaxation; Pets are for companionship
5. Causality: After the wedding ceremony comes the wedding reception.
6. Emotion: A pet makes you feel happy; Rollercoasters make you feel excited
and scared.
In our extraction system, mapping rules can be found under all of these categories.
To explain mapping rules, we give an example of knowledge from the aforemen-
tioned Scene category:
somewhere THING1 can be is PLACE1

somewherecanbe
THING1, PLACE1
0.5, 0.1
This rule contains a sentence pattern with the variables THING1 and PLACE1
binding to some text blob, and the name of the predicate that this relation should map
to. Text blobs are normalized into concepts using a sieve-like grammar. The pair of
numbers on the last line represents the confidence weights given to forward relation
(left to right), and backward relation (right to left), respectively, for this predicate
relation. This also corresponds to the weights associated with the directed edges
between the nodes, THING1 and PLACE1 in the spreading activation network repre-
sentation.
It is important to distinguish the value of the forward relation on a particular rule,
as compared to a backward relation. For example, let us consider the fact, “some-
where a bride can be is at a wedding.” Given the annotation “bride,” it may be very
useful to return “wedding.” However, given the annotation “wedding,” it is arguably
not as useful to return all the things found at a wedding such as “bride,” “groom,”
“wedding cake,” “priest,” etc. For our problem domain, we will generally penalize the
direction in a relation that returns hyponymic (taxonomic child) concepts as opposed
to hypernymic ones (taxonomic parent).
Approximately 20 mapping rules are applied to all the sentences (400,000+) in the
OMCS corpus, and a set of 50,000 predicate argument relations is extracted. These
structures are compiled into a spreading activation network consisting of 30,000 con-
cept nodes and 160,000 direct edges. The average branching factor is 5.
4.2 Expansion as Spreading Activation
In spreading activation, the origin node is the annotation or concept we wish to ex-
pand and it is the first node to be activated. Next, the nodes one hop away from the
origin node are activated, then two levels away, and so on. Nodes will continue to be
activated so long as their activation score meets the activation threshold, which is a
number between 0 and 1.0. Given nodes A and B, where A has one edge pointing to
B, the activation score (AS) of B can be constructed:
AS ( B ) = AS ( A) * weight (edge ( A, B ))
When no more nodes are activated, we have found all the relevant concepts that
expand the input concept. One problem that can arise with spreading activation is
that nodes that are activated two or more hops away from the origin node may
quickly lose relevance, causing the search to lose focus. This can be due to noise.
Because concept nodes do not make distinctions between different word senses, it is
possible that a node represents many different word senses. Therefore, activating
more than one hop away risks exposure to noise. Although associating weights with
the edges provides some measure of relevance, these weights form a homogenous
class for all edges of a common predicate (recall that the weights came from mapping
rules).
We identify two opportunities to re-weight the graph to improve relevance: rein-
forcement and popularity. Both are relatively common techniques associated with
spreading activation, but we motivate their explanations in the context of common
sense.
Reinforcement
We make the observation that if the concept “bride” is connected to “groom,” both
directly, and through “wedding,” then “groom” is more semantically relevant to
“bride” because two paths connect them. This is the idea of reinforcement. Looking
at this another way, if three or more concepts are mutually connected, as all the con-
cepts about a wedding might be, they form a cluster, and any two concepts in the
cluster have enhanced relevance because the other concepts provide additional paths
for reinforcement. Applying this, we re-weight the graph by detecting clusters and
increasing the weight on edges within the cluster.
Popularity
The second observation we make is that if an origin node A has a path through node
B, and node B has 100 children, then each of node B's children are less likely to be
relevant to node A than if node B had had 10 children. This is a common notion used
in spreading activation, often referred to as “fan-out” [10].
We refer to nodes with a large branching factor as being popular. It so happens
that popular nodes in our graph tend to be very common concepts in commonsense,
or tend to have many different word senses, or word contexts. This causes its chil-
dren to be in general, less relevant.
to wedding
concepts
bride groom (bf = 129)
to beauty
to wedding concepts
concepts
bridesmaid
(bf = 2)
Fig. 2. Illustrating the negative effects of popularity
As illustrated in Figure 2, the concept bride may lead to bridesmaid and groom.
Whereas bridesmaid is a more specific concept, not appearing in many contexts,
groom is a less specific concept. In fact, different senses and contexts of the word
can mean “the groom at a wedding,” or “grooming a horse,” or “he is well-groomed.”
This causes groom to have a much larger branching factor.
Despite being a knowledge base of common sense, there seems to be more value
associated with more specific concepts than general ones. To apply this principle, we
visit each node and discount the weights on each of its edges based on the following
heuristic (• and • are constants):
newWeight = oldWeight * discount
1
discount =
log(α * branchingFactor + β )
4.3 Example
Below is actual output of the concept expansion program using an activation thresh-
old of 0.1.
>>> expand(“bride”)
('wedding', '0.3662') ('woman', '0.2023')
('ball', '0.1517') ('tree', '0.1517')
('snow covered mountain', '0.1517')
('flower', '0.1517') ('lake', '0.1517')
('cake decoration', '0.1517') ('grass', '0.1517')
('groom', '0.1517') ('tender moment', '0.1517')
('veil', '0.1517') ('tuxedo', '0.1517')
('wedding dress', '0.1517') ('sky', '0.1517')
('hair', '0.1517') ('wedding boquet', '0.1517')
5 Related Work
The state-of-the-art in image annotation for consumer photography is probably best
represented by Kuchinsky et. al. [4]. Kuchinsky does not observationally learn anno-
tations from text descriptions, but it does use some image analysis to propose annota-
tions. Budzik and Hammond’s Watson [1] is an agent that observes user actions and
automates retrieval, but does not consider annotation. Neither of the aforementioned
programs provides real-time recommendations of images or adaptively links text with
images through semantic connectedness.
The concept expansion mechanism proposed here is not necessarily a new ap-
proach, but performing concept expansion with commonsense relations is new. In the
past, other dictionary-like resources such as lexical semantic relations [12], and key-
word co-occurrence statistics [9] have been used. The limitations of these resources
have been that for the most part, they operate on a word, rather that concept level. In
addition, the size and variety of their relational ontologies have been a limiting factor.
For example, OMCS gives us numerous relations including temporal, social, and
emotion but a resource like WordNet [2] can only give us a small set of nymic rela-
tions. Represented as semi-structured English sentences, it is also relatively easy to
augment the relational ontology, and easy to update.
6 Conclusion
In this paper, we presented ARIA, a software agent that facilitates a hypermedia au-
thoring task. While the user tells a story in an email client, the agent observes the text
pane and continuously presents suggestions of photos that may be relevant to the
context of the developing story. By using a semantically enriched parsing technique
on description text, the agent is able to automatically annotate photos used in the
story with semantically important concepts like the “who, what, and where” of the
photo. Then using concepts and relations extracted from Open Mind, the photo rec-
ommendation mechanism is able to adaptively present not only photos whose annota-
tions explicitly match the text, but also photos whose annotations exhibit implicit
semantic connectedness to the text.
In user testing [6], we saw not only that ARIA adapts to the user, but that the user
adapts to ARIA. Often a user's typing will bring up some photos relevant to the user's
current text, but that also trigger the user's memory, encouraging him or her to ex-
plain related pictures in subsequent text, triggering new picture retrieval. This mutual
adaptation is an important characteristic of adaptive systems, and our users particu-
larly liked the continual interplay between their story and ARIA's suggestions.
Another example of a system that successfully integrates common sense knowl-
edge into an interactive application is Erik Mueller's Common Sense Calendar [8]. It
makes “sanity checks” such as helping you avoid situations like inviting a vegetarian
friend to a steak house for dinner. We think applications like this, and ARIA, show
that it is not necessary to find complete solutions to the common sense reasoning
problem in order to make common sense knowledge useful in an interactive applica-

tion. All you have to do is use a little common sense.
References
1. Budzik, J. and Hammond, K. J.: User Interactions with Everyday Applications as Context
for Just-in-Time Information Access, ACM Conf. Intelligent User Interfaces (IUI 2000),
ACM Press, New York, (Jan. 2000), pp.44-51.
2. Fellbaum, C. (Ed.): WordNet: An Electronic Lexical Database. MIT Press, Cambridge,
MA. (1998).
3. Jackendoff, R.: Semantic structures. Cambridge, MA: MIT Press, (1990).
4. Kuchinsky, A., Pering, C., Creech, M. L., Freeze, D., Serra, B., and Gwizdka, J.: FotoFile:
a consumer multimedia organization and retrieval system, ACM Conference on Human-
Computer Interface, (CHI-99) Pages 496 – 503, Pittsburgh, (May 1999).
5. Lenat, D.: The dimensions of context-space, Cycorp technical report, (1998),
www.cyc.com.
6. Lieberman, H., Rosenzweig, E., and Singh, P.: Aria: An Agent For Annotating And Retrie-
ing Images. IEEE Computer, (July 2001), pp. 57-61.
7. Minsky, M.: Commonsense-Based Interfaces. Communications of the ACM. Vol. 43, No. 8
(August, 2000), Pages 66-73
8. Mueller, E. T.: A calendar with common sense. In Proceedings of the 2000 International
Conference on Intelligent User Interfaces, 198-201. New York: Association for Computing
Machinery. (2000).
9. Peat, H. J. and Willett, P.: The limitations of term co-occurrence data for query expansion
in document retrieval systems. Journal of the ASIS, 42(5), (1991), 378--383.
10. Salton G. and Buckley C.: On the Use of Spreading Activation Methods in Automatic
Information Retrieval, In Proc. 11th Ann. Int. ACM SIGIR Conf. on R&D in Information
Retrieval (ACM), (1988), 147-160.
11. Singh, P.: The Public Acquisition of Commonsense Knowledge. AAAI Spring Symposium,
Stanford University, Palo Alto, CA, (2002).
12. Voorhees, E.: Query expansion using lexical-semantic relations. In Proceedings of ACM
SIGIR Intl. Conf. on Research and Development in Information Retrieval. (1994) 61-69.
Resource-Adaptive Interfaces
to Hybrid Navigation Systems
(Keynote Talk)
Wolfgang Wahlster
DFKI - German Research Center for Artificial Intelligence

Stuhlsatzenhausweg 3
D-66123 Saarbrücken, Germany
[email protected]
www.dfki.de/~wahlster/
We present the hybrid navigation systems REAL [1] and SmartKom [2] that combine
various positioning technologies to determine the user’s location in outdoor and in-
door situations and that adapt the multimodal presentation of spatial information to the
limited technical resources of various output devices and to the limited cognitive re-
sources of the user.
We distinguish three different classes of resource sensitive processes: (a) resource
adapted processes, (b) resource adaptive processes and (c) resource adapting proc-
esses [1]. Resource adapted processes have been optimized in advance for restricted
resources that are well-known and follow regular patterns. The quality of their results
remains constant for a given input. In contrast, resource adaptive and resource adapt-
ing processes can handle varying resource restrictions. Therefore, their results depend
on the available resources during runtime. Resource adaptive processes rely on a sin-
gle strategy to react to varying resources, whereas resource adapting processes select
among a number of strategies on a meta-cognitive level to comply with different re-
source situations. REAL uses both resource adaptive and resource adapting processes
to generate multimodal space descriptions.
Personal navigation systems that extend beyond today’s use in cars will play a ma-
jor role in the future. Especially pedestrians will benefit from getting localized infor-
mation on mobile devices anywhere at any time. One essential issue in this context is
the switch between various positioning technologies, e. g., from outdoor systems
based on GPS or cell-based positioning based on GSM or UMTS to indoor systems
based on infrared or Bluetooth and vice versa. Ideally, a system should conceal the
switching between these technologies from the user entirely. In addition, the presenta-
tion of spatial information has to be adapted to the specific output devices that may be
used. This includes adaptations to the screen size, resolution, and color.
In terms of cognitive resources we take into account the user’s actual travelling
speed, his familiarity with the environment, and the time pressure in the current situa-
tion. Special adaptation services are provided when navigation is not the primary goal
of the user. Additional tasks that have to be performed in parallel such as carrying
heavy luggage or answering a telephone call while walking generate additional stress
that has to be compensated by the system. Our system already attempts to achieve this
during the process of way finding. Instead of choosing the shortest route, it tries to

Resource-Adaptive Interfaces to Hybrid Navigation Systems 13
avoid complex instructions at the cost of a slightly longer route, thus minimizing the
additional cognitive load on the user.
SmartKom is a multimodal dialogue system that combines speech, gesture, and fa-
cial expressions for input and output. SmartKom-Mobile uses a PDA as a front end.
Currently, the iPAQ Pocket PC with a dual-slot PC card expansion pack is used as a
hardware platform. It can be added to a car navigation system or carried by a pedes-
trian. SmartKom-Mobile provides personalized mobile services. Examples of value-
added services include route planning and interactive navigation through a city.
SmartKom provides an anthropomorphic and affective user interface through its
personification of an interface agent. Understanding of spontaneous speech is com-
bined with recognition of natural gestures and facial expressions. One of the major
scientific goals of SmartKom is to design new computational methods for the seam-
less integration and mutual disambiguation of multimodal input and output on a se-
mantic and pragmatic level. SmartKom is based on the situated delegation-oriented
dialogue paradigm, in which the user delegates a task to a virtual communication
assistant, visualized as a life-like character on a graphical display. The life-like char-
acter designed for the SmartKom system is called “Smartakus”. The “i”- shape of
Smartakus reminds one of the “i” often used as a sign that directs people to informa-
tion kiosks. The display of the 3D character Smartakus is adapted to the user’s view-
ing angle.
SmartKom’s interaction style breaks radically with the traditional desktop meta-
phor. SmartKom is based on the situated delegation-oriented dialogue paradigm
(SDDP):The user delegates a task to a virtual communication assistant, visible on the
graphical display. Since for more complex tasks this cannot be done in a simple com-
mand-and-control style, a collaborative dialogue between the user and the agent,
visualized as a life-like character, elaborates the specification of the delegated task
and possible plans of the agent to achieve the user’s intentional goal. In contrast to
task-oriented dialogues, in which the user carries out a task with the help of the sys-
tem, with SDDP the user delegates a task to an agent and helps the agent, where nec-
essary, in the execution of the task. The interaction agent accesses various web and
navigation services on behalf of the user, collates the results, and presents them to the
user.
REAL and SmartKom represent, reason about, and exploit models of the user, his
current cognitive state, his task, the situational context, and the available technical
resources and media in order to generate situation- and user-adaptive multimodal
route descriptions.
References
1. Baus, J., Krüger, A., Wahlster, W.(2002): A Resource-Adaptive Mobile Navigation Sys-
tem. In: Proceedings of the 2002 International Conference on Intelligent User Interfaces
(IUI'02), ACM Press, pp. 15- 22, ISBN 1-58113-459-2
2. Wahlster, W.,Reithinger N., Blocher, A. (2001): SmartKom: Multimodal Communication
th
with a Life-Like Character. In: Proceedings of Eurospeech 2001, 7 European Conference
on Speech Communication and Technology, Aalborg, Denmark, September 2001, Vol. 3,
pp. 1547 – 1550.
Ubiquitous User Assistance
in a Tourist Information Server
Liliana Ardissono, Anna Goy, Giovanna Petrone,

Marino Segnan, and Pietro Torasso
Dip. Informatica, Università di Torino, Corso Svizzera 185, I-10149 Torino, Italy
{liliana, goy, giovanna, marino, torasso}@di.unito.it
http://www.di.unito.it/˜ seta
Abstract. This paper discusses the provision of ubiquitous services for

desktop and handheld interfaces. The focus of the paper is on the inter-
active agenda offered by INTRIGUE, a tourist information server which
assists the user in the organization of a tour by providing personalized
recommendations of tourist attractions and helping the user to schedule
her itinerary. The interactive agenda can be accessed both remotely, by
interacting with the central server, and locally to the user’s device.
1 Background
With the current expansion of wireless communications, several Web-based ser-

vices have been extended to the mobile phone market, with special attention to
context-awareness and to the adaptation of the user interface to the peculiarities
of the various handset devices. However, only very basic, “real-time” services,
such as the consultation of stock quotes, are accessed by mobile phone users,
because wireless connections are extremely expensive. In contrast, the applica-
tions that can be executed locally to the device, such as music and games, are
appreciated by a broader customer base. Although the situation might change
in the future, at the current stage applications should rely on wireless network
connections for the tasks that really need a remote connection [1].
In our recent research, we have addressed the design of system architectures
for the development of services accessible both from desktop environments and
handset devices, via remote connections to central servers, or local execution
of tasks within the devices themselves. In these applications, persistent interac-
tion contexts have to be managed to support the synchronization of the user’s
activities when switching from one device to the other [2]. We experimented
our ideas with INTRIGUE (INteractive TouRist Information GUidE), a proto-
type information service presenting tourist information on Web browsers and
WAP phones. This paper describes INTRIGUE’s interactive agenda, which sup-
ports the tour scheduling by generating optimized itineraries on the basis of the
user’s temporal constraints, the attractions’ opening times and their location. In
this application, the persistent interaction context is represented by the user’s
itinerary, which can be consulted and modified during different sessions. The
P. De Bra, P. Brusilovsky, and R. Conejo (Eds.): AH 2002, LNCS 2347, pp. 14–23.
c Springer-Verlag Berlin Heidelberg 2002

Ubiquitous User Assistance in a Tourist Information Server 15
Fig. 1. Recommendation of attractions generated by INTRIGUE for a Web browser.
interactive agenda is downloadable and can run both in remote modality, by

interacting with the tourist information server, and local modality, where the
user can re-schedule her itinerary, without connecting to the server. This facility
supports a quick revision of itineraries during the tour, therefore extending the
role of our system, from tourist information service to interactive tourist guide,
supporting the user during the visit of the attractions.
The paper is organized as follows: Section 2 provides an overview of IN-
TRIGUE. Sections 3 and 4 describe the interactive agenda service and the ubiq-
uitous access facilities offered by our system. Section 5 sketches the system ar-
chitecture and the dynamic generation techniques exploited to manage the user
interface. Finally, Section 6 discusses some related work and concludes the paper.
2 Overview of the System

INTRIGUE provides multilingual information about tourist attractions and ser-
vices, such as accommodation and food, in a geographical area. The main goal is
assisting the user in the organization of a tour, by taking into account the possi-
bly conflicting preferences of a group of people traveling together; for instance,
a family with adults and children. Our current prototype presents information
about the city of Torino and the surrounding Piedmont area, in Italian and in
English [5]. The system supports the search for tourist attractions by offering or-
thogonal criteria, such as category-based and geographical search: e.g., the user
can look for churches in Torino city, buildings in the area around Torino, and so
forth. The search queries are created by exploiting a graphical interface hiding
the details of the query specification language: the user browses a taxonomy of
categories and selects the desired geographical area by clicking on a map, or by
following suitable hypertextual links, depending on the device she is using.
Personalization strategies are applied to recommend the attractions best suit-
ing the preferences of the tourist group. For instance, Fig. 1 shows a recommen-
dation list generated by our system: each item is a hypertextual link, which the
16 L. Ardissono et al.
Fig. 2. Specification of constraints and itinerary presentation in the interactive agenda:

(a) desktop and (b) handheld interfaces.
user can follow to view a page presenting the most relevant information about
the attraction. Moreover, each item is associated with a pictorial representation
of its recommendation degree (stars) and, if relevant, with an explanation of
the system’s suggestion. Finally, an “add to agenda” button enables the user to
include the item in the list of attractions to be scheduled in her itinerary.
3 The Interactive Agenda

The interactive agenda can be accessed by clicking a “View Agenda” button
available in the recommendation pages. Given the set of attractions selected by
the user, the agenda supports the specification of various types of constraints:
e.g., day of visit, arrival/departure time and location, preferred time of visit.
On the basis of the user’s choices, the system tries to generate an itinerary
that includes as many selected attractions as possible, complies with the attrac-
tions’ opening times and the user preferences, and minimizes the overall transfer
time.1 The presentation of an itinerary specifies, for each attraction, the esti-
mated transfer time to reach it, the time of the visit and its expected duration.
1
The system applies constraint-solving techniques for the generation of candidate
solutions, which are then compared for the selection of the one to be suggested on
the basis of their total transfer times. The straight application of “shortest path”
algorithms, such as the sales traveler one, is not possible, as they do not guarantee
that the generated permutations of attractions satisfy temporal constraints like those
concerning the opening time of attractions.
Fig. 3. Failure in the organization of an itinerary.
Figures 2.(a) and 2.(b) present an itinerary for visiting Palazzo Reale, the Lin-
gotto building and the Sacra di San Michele, in a situation where the user asked
to start the visit from the Royal Palace and to end it at the Sacra. The presen-
tation is different for desktop and handheld interfaces: in the former case, the
system shows the user’s constraints and the itinerary in the same page. In the
latter, for space limitations, this information is compactly presented, in separate
pages.
If the system fails in scheduling some of the selected attractions, it explains
the problem and provides useful information to re-schedule the itinerary. For in-
stance, the failure shown in Fig. 3 is due to the fact that most tourist attractions
are closed on Mondays. The user can solve the conflict by changing her selec-
tions. For example, she could try to visit other attractions, such as the parks,
or she could modify her constraints, e.g., by changing the day of the tour. The
interactive agenda enables the user to save an itinerary for further consideration
(“save schedule” button in Fig. 3).
4 Ubiquitous Interaction with INTRIGUE

As our system can be accessed by exploiting desktop computers and mobile
phones, the user can use it during the organization of a tour and during the tour
itself, in order to view the tourist catalog and generate itineraries for the visit of
Fig. 4. Local interaction with INTRIGUE.
attractions. Moreover, the possibility to store the user’s itinerary supports the
any time revision of previously generated itineraries. However, the peculiarity of
our system is the possibility to access some of its services locally to the user’s
device, which enables the access to the interactive agenda without a continuous
connection to the central server. In this way, the user can consult her itinerary
and possibly re-schedule it, by exploiting the mobile phone in a similar fashion
to the way people use to run games on such devices.
The remote interaction with INTRIGUE is similar for desktop and handset
devices: the user browses the catalog, selects tourist attractions to be visited and
schedules itineraries. The central server stores the interaction context, which can
include the user’s itinerary, her last search query and maintains the information
about the list of attractions satisfying the query (e.g., buildings around Torino,
corresponding to the recommendation list shown in Fig. 1). Moreover, the con-
nection to the tourist information database supports the execution of further
queries to retrieve new items that can be added to the agenda.
The local access to the services deserves further discussion. A possible sce-
nario of a local use of the agenda is shown in Fig. 4: the user connects to the
remote server, downloads a previously generated itinerary on her handset and
closes the interaction. Then, she opens the agenda to view and possibly resched-
ule her itinerary on the handset, in local modality. When the user later on
connects again to the server, local changes to the itinerary are sent back to the
networked data repository for permanent storage.2 In local access modality, the
interactive agenda exploits its own information resources: the itinerary genera-
tion and scheduling capabilities are offered, but the user cannot add new items to
the agenda, as only the information about the previous selections is available on
the device. Therefore, if the user wants to consider new tourist attractions, she
2
In the figure, the currently used agenda is shown by a medium size icon; the possibly
out of date one is shown as a small icon.
has to connect to the central server, by clicking a “modify → extend itinerary”

link available in the interface of the handheld interactive agenda.
The support of local and remote connections to the interactive agenda relies
on the storage of an interaction context aimed at managing subsequent accesses
to the system as portions of the same session.
– On the one hand, when saving an itinerary, the system also saves the infor-
mation about the attractions selected by the user for further consultation.
Moreover, the user’s search query, representing the last focus of attention
during the interaction, is saved. This information enables the user to con-
nect to the central server for retrieving new items without browsing the
catalog from the beginning. For instance, if the user initially searched for
the buildings around Torino and wants to visit more places than those avail-
able in the agenda, she can directly access the recommendation page where
the items are available; see Fig. 1.
– On the other hand, when the user downloads an itinerary, the network con-
nection is needed to download the itinerary and the information about the
selected tourist attractions, which permanently reside in the database of the
INTRIGUE server. After this synchronization phase, the agenda runs locally,
without requiring further connections to the central server.
5 Technical Aspects
5.1 Architecture
INTRIGUE is based on the architecture described in [4], which has been ex-
tended to manage the interaction with mobile phones by exploiting the WAP
protocol: we used the Nokia mobile phone simulator for our experiments. The
generation of the device-dependent user interface relies on the exploitation of
standard XML-based techniques; the local execution of the interactive agenda
on handset devices is based on the exploitation of Java Micro Edition and of the
synchronization facilities offered by such environment (SynchML [14]).
5.2 Generation of the User Interface
The User Interface (UI) of INTRIGUE can be viewed as a sequence of dynam-

ically generated pages. The generation of each page is performed according to
four distinct steps, listed in the following. While the first three steps are in-
dependent of the device the UI is implemented on, the last one generates the
appropriate UI on the basis of the user’s device.
– Selection of the information to be presented. Different information is selected

for the various page types generated by the system. In a recommendation
page (see Fig. 1) the selected information consists of a list of attractions,
sorted on the basis of the tourist group’s preferences. In a page describing
an attraction, such as the one in Fig. 5, this information includes one or
Fig. 5. Portion of a page prefenting a tourist attraction for desktop interface.
more sets of features describing the attraction from different points of view.
Finally, the information selected for the agenda includes the list of attractions
chosen by the user and the itinerary calculated by the system.
– Generation of linguistic descriptions. The internal representation of the se-
lected information is translated to natural language by exploiting efficient
template-based NLG techniques supporting the production of simple sen-
tences in Italian and English [3].
– Generation of the content structure of the page. A DTD defines the logical
structure of the various types of page the system can produce; navigation
pages, presentation pages, agenda pages, and so forth. An XML object (i.e.,
an XML instance document) representing the personalized content of the
page, independent of the actual UI, is produced.
– Surface generation. The UI is generated by transforming the XML object to
a standard HTML, or WML page, depending on the user’s device (desktop
or mobile phone). In both cases, XSL stylesheets are exploited for the trans-
formation. In contrast, when the user accesses the agenda in local modality,
XSL stylesheets cannot be used, since the Java Micro Edition, running on
mobiles, does not currently support standard tools for XSL processing such
as Xalan. Taking also performance limitation of handset devices into ac-

count, we implemented a low-level WML generation module, based on light
XML parsing (basically relying on SAX 1.0 APIs).
In the transformation from XML to HTML/WML, the surface generator per-

forms two choices:
– It selects the subset of content stored in the XML object to presented, given
the space constraints of the display. For instance, the page describing the
tourist attraction in Fig. 5 is generated for a Web browser, where there
are no serious space constraints: thus, all types of features are displayed.
When generating the same type of page for a mobile, the surface gener-
ator omits the picture and selects only the most important types of fea-
tures, leaving the rest available via “more info” links. For instance, both the
“DESCRIPTION” paragraph (containing qualitative information about the
item), and the “CHARACTERISTICS” section (reporting detailed informa-
tion) are hidden, although they can be accessed via a link.
– It provides the layout information, deciding where and how, on the display,
the various pieces of content should be placed.
6 Related Work and Discussion

The dynamic, context-aware organization of a tour was initially proposed in the
GUIDE system to support users in the real-time organization of their visit to the
Lancaster town [12]. However, the architecture of such system is based on the
exploitation of specialized hardware, physically located in the place to be visited,
and on the use of special handset devices. In contrast, INTRIGUE is based on
standard wireless connections and on the use of (Java-enabled) mobile phones.
Moreover, the interactive agenda maintains persistent itineraries that can be
consulted and updated by exploiting wired and wireless connections. Therefore,
our system offers an effective ubiquitous service accessible from any place.
Ubiquitous computing has been interpreted by several researchers as a way to
offer, on wireless connections, the same services already developed for internet-
based browsers. As discussed in [8], the user interfaces offered by desktops, televi-
sions and handset devices differ in size, interaction tools (e.g., keyboard, mouse,
etc.) and usage environment (e.g., desktop, handheld device). In order to face
this heterogeneity, the Adaptive Hypermedia community has developed adap-
tation techniques to tailor the interaction to the specific devices. For instance,
interesting examples of services that the user can access both from desktop and
WAP phone are provided in [10,11,15,17]. Although these examples deeply differ
from one another in the offered functionalities and underlying technologies, all
of them exploit the user’s device as a lightweight client, continuously interacting
with a remote server, where the logic of the application resides. Moreover, they
mainly interpret context-awareness as taking the user’s physical (or temporal)
coordinates into account.
Our proposal differs from these approaches in the following two main aspects.
– In addition to the basic service provision, which has to rely on a server-based

management of the interaction with the user, we offer an interactive agenda
and other services that the user can access locally to the device.
– We interpret context awareness and ubiquitous computing as a persistence
property of the interaction with the user. In particular, the interactive agenda
is modeled as a persistent object, that the user can create, update and carry
with herself at any time, and along different interactions with the system,
regardless the device she is using.
The first aspect is particularly relevant to the exploitation of handset devices,

which are subject to important limitations: first, at the current stage of techno-
logical development, wireless connections are not reliable; therefore, long-lasting
interactions, typically performed in the wired case, are problematic. Similarly,
bandwidth is limited, so that there are clear demands for traffic minimization.
Finally, in most countries wireless communications are very expensive; therefore,
in order for users to appreciate the service, such connections should be limited
to the services that cannot be offered off-line: e.g., the search for continuously
updated information, such as stock quotations, and the management of business
interactions, such as reservations and payment transactions. Although the first
two issues might be overcome in few years, it is not clear whether the third one
will be solved in the medium term. In order to develop successful applications,
such constraints should be taken into serious account.
We have performed an initial testing of our system with users who helped
us to improve its interface and functionalities. The users were happy about the
system’s explanation facilities, both regarding group recommendations and fail-
ures in itinerary generation, because they clarify the system’s behavior, therefore
increasing the trust in the system’s recommendations. Moreover, users appreci-
ated the possibility of saving itineraries and retrieving (and modifying) them in
subsequent interactions, as this facility would enable them to use the interactive
agenda during the tour. In contrast, there was some negative feedback in the
organization of the device-dependent interaction with the user. For instance, the
registration form for the specification of heterogeneous tourist groups, which was
well-accepted during the desktop interaction, was considered too time consum-
ing for a minibrowser. To address this issue, we have extended the system to
support logging without registering: if the user logs in this way, she receives non
personalized recommendations, but she can access the other facilities offered by
the system, such as category and geographical search, itinerary scheduling and
presentation of tourist attractions. Moreover, we will investigate the application
of unobtrusive user modeling techniques aimed at the identification of the user’s
preferences on the basis of the observation of her browsing behavior [6,9].
Our future work will include further testing of the system, in order to collect
more usage information and hopefully to try the system with real mobile phones:
the mobile functionalities of our system are currently tested by exploiting a the
WAP simulator. Moreover, we will exploit further contextual information to
enhance the itinerary generation capabilities; e.g., to take location and temporal
information into account, as done in other systems [2,7,13,16].
References
1. Special issue: The mobile internet. The Economist, (October 11, 2001) 27–54
2. G.D. Abowd and E.D. Mynatt. Charting past, present and future research in
ubiquitous computing. ACM Transactions on Computer-Human Interaction, Spe-
cial Issue on HCI in the new Millennium, 7(1) (2000) 29–58
3. L. Ardissono and A. Goy. Dynamic generation of adaptive Web catalogs. In
Lecture Notes in Computer Science, Vol. 1892: Adaptive Hypermedia and Adaptive
Web-Based Systems. Springer-Verlag, Berlin (2000) 5–16
4. L. Ardissono, A. Goy, G. Petrone, and M. Segnan. A software architecture for
dynamically generated adaptive Web stores. In Proc. 17th IJCAI, Seattle (2001)
1109–1114
5. L. Ardissono, A. Goy, G. Petrone, M. Segnan, and P. Torasso. Tailoring the recom-
mendation of tourist information to heterogeneous user groups. In Third Workshop
on Adaptive Hypertext and Hypermedia, Aarhus, Denmark (2001) 93–107
6. L. Ardissono and P. Torasso. Dynamic user modeling in a Web store shell. In
Proc. 14th Conf. ECAI, Berlin (2000) 621–625
7. B.B. Bederson. Audio augmented reality: a prototype automated tour guide. In
R. Mack J. Miller, I. Katz and L. Marks, editors, Conf. companion for the ACM
Conf. on Human Factors in Computing Systems, Denver (1995) 210–211
8. E. Bergman, editor. Information Appliances and Beyond. Interaction design for
consumer products. Morgan Kaufmann Publishers (1999)
9. D. Billsus and M. Pazzani. A personal news agent that talks, learns and explains.
In Proc. 3rd Int. Conf. on Autonomous Agents, Seattle (1999) 268–275
10. D. Billsus, M. Pazzani, and J. Chen. A learning agent for wireless news access. In
Proc. 2000 Int. Conf. on Intelligent User Interfaces, New Orleans (2000) 33–36
11. B. De Carolis, F. de Rosis, and S. Pizzutilo. Context-sensitive information presen-
tation. In Proc. UM’2001 Workshop on User Modelling in Context-Aware Appli-
cations, Sonthofen, Germany (2001)
12. K. Cheverest, N. Davies, K. Mitchell, A. Friday, and C. Efstratiou. Developing
a context-aware electronic tourist guide: some issues and experiences. In Proc.
CHI’2000: Human factors in computing systems, The Hague, NL (2000) 17–24
13. A.K. Dey and D. Abowd. Towards a better understanding of context and context-
awareness. In Proc. CHI2000 Workshop on the What, Who, Where, When and
How of Context-Awareness, The Hague, NL (2000)
14. Y. Feng and J. Zhu. Wireless Java Programming with J2ME. SAMS (2001)
15. C. Paris, S. Wan, R. Wilkinson, and M. Wu. Generating personal travel guides -
and who wants them? In Proc. 8th Int. Conf. on User Modeling Sonthofen, DE
(2001) 251–253
16. D. Petrelli, A.De Angeli, and G. Convertino. A user centered approach to user
modelling. In Proc. 7th Int. Conf. on User Modeling, Banff (1999) 255–264
17. D. Riecken, editor. Special Issue on Personalization, 43. Communications of the
ACM (2000)
Automatic Extraction
of Semantically-Meaningful Information
from the Web.
J.L. Arjona, R. Corchuelo, A. Ruiz, and M. Toro
Escuela Técnica Superior de Ingeniería Informática de la Universidad de Sevilla

Departamento de Lenguajes y Sistemas Informáticos
Avda. de la Reina Mercedes, s/n, Sevilla (SPAIN)
{arjona,corchu,aruiz,mtoro}@lsi.us.es
Abstract. The semantic Web will bring meaning to the Internet, mak-
ing it possible for web agents to understand the information it contains.
However, current trends seem to suggest that the semantic web is not
likely to be adopted in the forthcoming years. In this sense, meaning-
ful information extraction from the web becomes a handicap for web
agents. In this article, we present a framework for automatic extraction
of semantically-meaningful information from the current web. Separat-
ing the extraction process from the business logic of an agent enhances
modularity, adaptability, and maintainability. Our approach is novel in
that it combines different technologies to extract information, surf the
web and automatically adapt to web changes.
Keywords: Web agents, information extraction, wrappers, and ontolo-

gies.
1 Introduction
In recent years, the web has consolidated as one of the most important knowl-
edge repositories. Furthermore, the technology has evolved to a point in which
sophisticated new generation web agents proliferate. They enable efficient, pre-
cise, and comprehensive retrieval and extraction of information from the vast
web information repository. They can also circumvent some problems related to
slow Internet access, and free up prohibitively expensive surf time by operating
in the background. It is thus not surprising that web agents are becoming so
popular.
A major challenge for web agents has become sifting through an unwieldy
amount of data to extract meaningful information. Two important factors con-
tribute to these difficulties: first, the information on the web is mostly available
in human-readable forms that lack formalised semantics that would help agents

The work reported in this article was supported by the Spanish Inter-ministerial
Commission on Science and Technology under grant TIC2000-1106-C02-01

Extraction of Semantically-Meaningful Information for the Web 25
use it [3]; second, the information sources are likely to change their structure,
which usually has an impact on their presentation but not on their semantics.
Thus, if we want to succeed in the development of web agents, we need a
framework in which they can be separated from the information sources or the
way to extract semantically-meaningful information from them. This way we
enhance modularity, adaptability and maintainability, as well as agent interop-
erability. In this article, we present such a framework. It is organised as follows:
Section 2 goes into details about our motivation and some related work; Sec-
tion 3 presents our proposal and a case study; finally, Section 5 reports on our
main conclusions and future research directions.
2 Motivation and Related Work

The incredible successfulness of the Internet world has paved the way for tech-
nologies whose goal is to enhance the way humans and computers interact on the
web. Unfortunately, the information a human user can easy interpret is usually
difficult to be extracted and interpreted by a web agent.
Figure 1 shows two views of a web page picked from Amazon.com. On the left,
we present a shot of the page a human sees when he or she searches for informa-
tion about a book; on the right, we present a portion of HTML that describes
how to render this page. If we were interested in extracting the information
automatically, the following issues would arise immediately:
– The implied meaning of the terms that appear in this page can be eas-
ily interpreted by humans, but there is not a reference to the ontology that
describes them precisely, which complicates the communication interface be-
tween user and agent, and the interoperability amongst agents. For instance,
...
<b class="sans">Beginning Xml</b><br> <font
face=verdana,arial,helvetica size=-1>
by <a href="/exec/obidos/search-handle-url/index=books&field-
author=Cagle%2C%20Kurt/107-9979303-0464511">Kurt Cagle</
a>, <a href="/exec/obidos/search-handle-url/index=books&field-
author=Gibbons%2C%20Dave/107-9979303-0464511">Dave
Gibbons</a>, <a href="/exec/obidos/search-handle-url/
index=books&field-author=Hunter%2C%20David/107-9979303-
0464511">David Hunter</a>, <a href="/exec/obidos/search-handle-
url/index=books&field-author=Ozu%2C%20Nikola/107-9979303-
0464511">Nikola Ozu</a>, <a href="/exec/obidos/search-handle-url/
index=books&field-author=Pinnock%2C%20Jon/107-9979303-
0464511">Jon Pinnock</a><br>
</font>
<br>
<a href=/https/www.scribd.com/exec/obidos/tg/stores/detail/-/books/1861003412/reader/1/
107-9979303-0464511#reader-link><img src="http://
images.amazon.com/images/P/1861003412.01.MZZZZZZZ.jpg"
width=111 height=140 align=left border=0></a>
<font face=verdana,arial,helvetica size=-1> <span class="small">
<span class="small">
<b>List Price:</b> <span class=listprice>$39.99</span><br>
<b>Our Price: <font color=#990000>$27.99</font></b><br>
<b>You Save:</b> <font color=#990000><b>$12.00 </b>
(30%)</font><br>
...
Fig. 1. A web page that shows information about a book.

26 J.L. Arjona et al.
if we do not consider semantics and we interpret prices as real numbers, a

US librarian agent that deals with dollars would be likely to misinterpret a
European one that deals with euros.
– The layout and the aspect of a web page may change unexpectedly. For
instance, it is usual that web sites incorporate Christmas stuff in December,
which does not affect the information they provide, but may invalidate the
automatic extraction methods used so far.
– The access to the page that contains the information in which we are inter-
ested may involve navigating through a series of intermediate pages, such as
login or index pages. Furthermore, this information may spread over a set
of pages.
On sight of these issues, several researchers began working on proposals whose

goal is to achieve a clear separation between presentation concerns and data
themselves. XML [12] is one of the most popular languages aimed at representing
structured data, but, unfortunately, it lacks a standardised way to link them with
an abstract description of their semantics. There are many proposals that aim at
solving this problem. They usually rely on annotating web pages with instances
of ontologies that is usually written in a language such as DAML [1] or SHOE
[19].
Most authors agree in that it would be desirable for a web in which pages
are annotated with semantics to exist, because this would help web agents un-
derstand their contents, and would enhance semantic interoperability. Unfor-
tunately, there are very little annotated pages if we compare them with non-
annotated pages. As of the time of writing this article, the DAML crawler
(www.daml.org/crawler) reports 17,019 annotated web pages, which is a negli-
gible figure if we compare it with 2,110 millions, which is the estimated number
of web pages (Cyveillance.com). Current trends seem to suggest that the seman-
tic web is not likely to be adopted in the forthcoming years [14]. This argues for
an automatic solution to extract semantically-meaningful information from the
web that is clearly separated from the business logic so as to enhance modularity,
adaptability, and maintainability.
Several authors have worked on techniques for extracting information from
the web, and inductive wrappers are amongst the most popular ones [17,23,15,4],
[21,2]. They are components that use a number of extraction rules generated by
means of automated learning techniques such as inductive logic programming,
statistical methods, and inductive grammars. These techniques use a number of
web pages as samples that feed an algorithm that uses induction to generalise a
set of rules that allows to extract information from similar pages automatically.
Recently, researchers have put a great deal of effort to deal with changes, so
that extraction rules can be regenerated on the fly if the layout of the web page
changes [18,16]. Although induction wrappers are suited to extract information
from the web, they do not associate semantics with the extracted data, This
being their major drawback.
There are also some related proposals in the field of databases, e.g., TSIM-
MIS [11] and ARANEUS [20]. Their goal is to integrate heterogeneous informa-
tion sources such as traditional databases and web pages so that the user can
work on them as if they were a homogeneous information source. However, these
proposals lack a systematic way to extract information from the web because ex-
traction rules need to be implemented manually, which makes them not scalable
and unable to recover from unexpected changes on the web.
3 Our Proposal
Our proposal aims at providing agent developers with a framework in which

they can have access to semantically-meaningful data that resides on web pages
not annotated with semantics. It relies on using a number of agents that we call
information channels or IC for short. They allow to separate the extraction of
information from the logic of an agent, and offer agent developers a good degree
of flexibility. In order to allow for semantic interoperability, the information they
extract references a number of concepts in a given application domain that are
described by means of ontologies.
Before going into details, it is important to say that our notion of agent
was drawn from [24]: “Agents have their own will (autonomy), they are able to
interact with each other (social ability), they respond to stimulus (reactivity),
and they take initiative (pro-activity).” In our proposal, web agents are software
agents that interact with the web to retrieve, extract or manage information.
3.1 The Architecture
Figure 2 sketches the architecture of our proposal. As we mentioned above, in-

formation channels are at its core agents because they specialise in extracting
information from different sources, and are able to react to information inquiries
(reactivity) from other agents (social ability). They act in the background proac-
tively according to a predefined schedule to extract information and to maintain
the extraction rules updated.
Each information channel uses several inductive wrappers to extract informa-
tion so that they can detect inconsistencies amongst the data each one extracts.
If such inconsistencies are found, they then use a voting algorithm to decide
whether to use the data extracted by most wrappers or regenerate the set of
extraction rules on the fly. This may happen because of an unexpected change
to the structure of the web page that invalidates the extraction rules used so far.
There is also an agent broker for information extraction that acts as a trader
between the agents that need information from the web and the set of available
information channels. When an agent needs some information, it contacts the
broker, which redirects the request to the appropriate information channel, if
possible. This way, agents need not be aware of the existence of different ICs,
which can thus be adapted, created or removed from the system transparently.
However, every time an IC is created or destroyed, it must be registered or
unregistered so that the broker knows it. It therefore has a catalogue with the
description of every IC in the system (yellow pages).
Information
channel 1
Agent
Broker
society
Information Internet
channel 2
Information
channel n
Agent platform (FIPA)
Fig. 2. Proposed architecture.
We use ACL [10] as a transport language to send messages from an agent to

another. The content of the messages describes how an agent wants to interact
with another, and it is written in DAML. Figure3 shows the brokering protocol
[9] to communicate user agents with the ICs using the notation AUML [13,22].
When an initiator agent sends a message with the performative proxy to the bro-
ker, it then replies with one of the following standard messages: not-understood,
refuse or agree. If the broker agrees on the inquiry, it then searches for an ade-
quate IC to serve it. If not found, it then sends a failure-no-match message to
the initiator; otherwise, it tries to contact the IC and passes the inquiry onto
it. If the broker succeeds in communicating with the IC, this will later send the
requested information to the initiator; otherwise, a failure-com-IC message is
sent back to the initiator, which indicates that an appropriate IC exists, but
cannot respond.
3.2 A Case Study
We illustrate our proposal by means of a simple example in which we are in-

terested in extracting information about books from Amazon.com. We first need
to define an information channel, i.e., we need to construct an expert system
whose goal is to extract information about books from the web. Such a channel
is characterised by the following features: a set of ontologies, a set of extraction
rules, and a navigational document.
The Ontologies. The ontologies [6] associated with an IC describes the con-
cepts that define the semantics associated with the information we are going
Information
Initiator Broker
Channel
proxy
not-understood
refuse
agree
failure-no-match
request
[Broker cannot find any Inf
channel for the request]
agree
refuse
not-understood
failure-com-IC
inform-result-IC
[Broker, find an Inf. channel
for the request]
Fig. 3. Broker Interaction Protocol in AUML.
to extract from a high-level, abstract point of view. An ontology allows to de-

fine a common vocabulary by means of which different agents may interoperate
semantically.
Figure 4 shows a part of the ontology that describes books using DAML. We
selected this language because it is one of the most important contributions to
the semantic web, it is being supported by many researchers, and the repository
of available tools is quite rich, c.f. http://www.daml.org/tools.
The Extraction Rules. The extraction rules allow to define how to have
access to the information in which we are interested. To generate them, we need
a set of sample pages containing test data on which we use inductive techniques.
To endow the sample pages with semantics, we also need to annotate them
with DAML tags that allow to associate the concepts they contains with their
corresponding ontologies. Figure 5 shows a piece of DAML code that we can use
to annotate the web page in Fig. 1.
Once the sample pages have been annotated, we can generate the extraction
rules. The input to the wrapper generator is a tuple of the following form:
({O1 , O2 , ..., On }, {(P1 , D1 ), (P2 , D2 ), ..., (Pk , Dk )}); n ≥ 1, k ≥ 1
The first element of the tuple denotes the set of ontologies under considera-
tion, and the second element is a set of pairs of the form (P, D), where P denotes
a web page containing sample data, and D its corresponding annotation. With
this information, we apply several induction algorithms [17,23,15,4] to generate
<rdf:RDF xmlns:rdf="http://www.w3.org/1999/02/22-rdf-syntax-ns#"
xmlns:rdfs="http://www.w3.org/2000/01/rdf-schema#" UML Model for the Representation of
xmlns:daml="http://www.daml.org/2001/03/daml+oil#" the DAML Book Ontology.
xmlns:xsd ="http://www.w3.org/2000/10/XMLSchema#">
(According to OMG Agent SIG)
<daml:Ontology about="">
<daml:versionInfo>10/30/2001</daml:versionInfo>
<rdfs:comment>Book Ontology</rdfs:comment>
<daml:imports rdf:resource="http://www.daml.org/2001/03/daml+oil"/>
</daml:Ontology>
Book
<daml:Class rdf:ID="Book">
<rdfs:comment>
isbn : String
A set of written sheets of skin or paper or tablets of wood or
title : String
ivory. Websters Dictionary.
author : String
</rdfs:comment>
editorial : String
</daml:Class>
pages : Integer
<daml:DatatypeProperty rdf:ID="isbn">
...
</rdf:RDF>
Fig. 4. An ontology that describes books from an abstract point of view.
<Book rdf:ID="A book">

<rdfs:label>A book</rdfs:label>
<rdfs:comment>Instance of Book.</rdfs:comment>
<title>Beginning Xml</title>
<isbn>1861003412</isbn>
<editorial>Wrox Press, Inc</editorial>
<author>Kurt Cagle</author>
<author>Dave Gibbons</author>
<author>David Hunter</author>
<author>Nikola Ozu</author>
<author>Jon Pinnock</author>
<pages>823</pages>
</Book>
Fig. 5. Annotations to a web page with information about a book.
a set of extraction rules R1, R2, ..., Rm. Their exact form depend on the algo-
rithm used to produce them, and may range from simple regular expressions to
search procedures over a DOM Tree [5] or even XPointers [8]. hereafter, we refer
to this set of rules as BookRules.
The Navigational Document. A navigational document defines the path of

intermediate pages, such as login or index, we need to visit in order to have
access to the page that contains the information in which we are interested.
Figure 6 shows a navigational document that describes how to have access to
the web page that contains information about a book with a given ISBN. This
document indicates that there exists a web page at www.amazon.com called d1.
This page has quite a complex structure, but it contains a form called search-
form with two fields a human user may use to select the kind of product he or
she is searching for (index ) and some key words that describe it (field-keywords).
Thus, d2 denotes the web page that we obtain when the server processes this
form using “index” := “Books” and “field-keywords” := “:Book#isbn”. Books is
the value that indicates the server that we are looking for books, and :Book#isbn
denotes the ISBN code associated with the book for which we are searching.
index=Books
field-keywords=181...
Navigational document NDBook

Using ontology http://arjona.lsi.us.es/Book#;
Document d1:="http://www.amazon.com/";
Document d2:=d1{searchform, "index" := "Books", "field-keywords" := :Book#isbn};
Extract Info from d2 using BookRules;

End.
Fig. 6. A sample navigational document.
Roughly speaking, a navigational document may be viewed as a state machine

in which each state is a web page, and the transitions between states may be
viewed as the navigation from the page associated with the initial state to the
page associated with the destination state.
3.3 Dealing with Pagers and Indexers

In this section, we present a slightly modified version of the previous case study in
which we are interested in extracting information about the set of books written
by a given author. If we write the name of an author in the above-mentioned
search form, we can get a result page similar to the one presented in Fig. 1. In this
case, the initial page does not lead directly to the requested information. Instead,
it leads to an index page in which we can find some links that point to pages
that describe books written by that author, and a pager that allows to search
for more results. Although the wrapper we generated previously keeps working
well, the way to have access to the information has changed substantially.
The new navigational document we need is presented in Fig. 7. There, we
define a new document called d3 that represents the web page to which the
initial search form leads. d4[] references the set of web pages that we can get if
we dereference the link labelled with “More results” by means of a sequential
pager. This is an artefact that allows to navigate through the complete set of
Navigational document NDBook

Using ontology http://arjona.lsi.us.es/Book#;
Document d1 := "http://www.amazon.com/";
Document d2 := d1{"searchform", "index" := "Books", "field-keywords" := :Book#isbn};
Document d3 := d1{"searchform", "index" := "Books", "field-keywords" := :Book#author};
Document d4[] := SequentialPager(d3, "More Results");
Document d5[] := Indexer(d4[], AmazonIndexer);
Extract Info from d2, d5[] using BookRules;

End.
Fig. 7. Modified version of the navigational document.
pages that contain the information about the books written by an author. d5[]
references the set of web pages we obtain by dereferencing the links on books
that appear in the set of pages d4[]. This set is obtained by means of an indexer,
which is an artefact that analyses a set of pages and extracts the links to the
pages in which we are interested. In this case, we use an inductive wrapper called
AmazonIndexer, but new indexers can be easily generated.
3.4 Using the Information Channel
Once we have set up an abstract channel, we can send messages to it in order to

retrieve information about a given book by means of the broker. The content of
the messages in DAML is based on an ontology that defines the communication
[7]. This ontology is illustrated in Fig. 8.
ExtractInfo
*
attributeRestrictions
Attribute ExtractInfoResponse
label : String
comment : String name : String
about : String value : String
Warning Error Ok
wcode : Integer ecode : Integer NumberRecords : Integer

wdescription : String edescription : String
Records
*
Thing
Fig. 8. Ontology for the content language.

Information requests are expressed as instances of class ExtractInfo. The

reply from the abstract channel is an instance of ExtractInfoResponse, an error
message (for instance, the channel can not have access the web page that contains
the information), a warning message (for instance, 0 records have been found) or
the information requested by the agent (as instances of the ontology class that
defines the channel).
Figure 9 shows two example messages, the first one relative to an information
request from an agent called Agent-1 to the broker agent; the second one is the
reply from the IC to Agent-1. In this case a book instance is returned.
(inform :sender BookChannel

:receiver Agent-1
:content (
(request :sender Agent-1
:receiver broker
xmlns:rdfs="http://www.w3.org/2000/01/rdf-schema#"
:content (
xmlns:daml="http://www.daml.org/2001/03/daml+oil#"
xmlns:xsd="http://www.w3.org/2000/10/XMLSchema#"
xmlns:rdfs="http://www.w3.org/2000/01/rdf-schema#"
xmlns:ie="http://arjona.lsi.us.es/ie#">
xmlns:daml="http://www.daml.org/2001/03/daml+oil#"
xmlns="http://arjona.lsi.us.es/Book#">
xmlns:xsd="http://www.w3.org/2000/10/XMLSchema#"
<ie:Ok rdf:ID="Request 123332">
xmlns:libro="http://arjona.lsi.us.es/Book#"
<ie:numberRecords>1</ie:numberRecords>
xmlns="http://arjona.lsi.us.es/ie#">
<ie:records>
<ExtractInfo rdf:ID="EI1">
<Book rdf:ID="A-Book">
<rdfs:label>EI1</rdfs:label>
<title>Beginning Xml</title>
<about>http://arjona.lsi.us.es/Book#Book</about>
<isbn>1861003412</isbn>
<attributeRestrictions>
<editorial>Wrox Press Inc</editorial>
<attribute>
<author>Kurt Cagle</author>
<name>http://arjona.lsi.us.es/Book#isbn</name>
<author>Dave Gibbons</author>
<value>1861003412</value>
<author>David Hunter</author>
</attribute>
<author>Nikola Ozu</author>
</attributeRestrictions>
<author>Jon Pinnock</author>
</ExtractInfo>
<pages>823</pages>
</rdf:RDF>
</Book>
)
</ie:records>
:language daml)
</ie:Ok>
)
:language daml)
Fig. 9. Example of messages.
4 Benefits of the Framework

for the Adaptative Hypermedia Community
The Adaptative Hypermedia Community can use the framework presented in
this paper to personalise web sites based on users’ profiles. In this respect, the
information from the web that is interesting to a user can be extracted auto-
matically and displayed in a suitable form.
We illustrate this idea with a fictitious case study whose aim is the person-
alisation of an e-commerce portal that sells VHS, DVDs, and so on. Using the
information infered from the buys by some customers, we can identify who the
the preferred actors are. Once the actors have been indentified, we can develop
an agent that uses the proposed framework. It queries an abstract channel that
is able to extract relevant information about this actor (birth name, location, fil-
mography, ...) from several web sites, for example The Internet Movie Database
(IMDb)1 .
1
http://www.imdb.com/
5 Conclusions and Future Work

In this article, we have shown that the process of extracting information from
the Internet can be separated from the business logic of a web agent by means of
abstract channels. They rely on inductive wrappers and can analyse web pages
to get information with its associated semantics automatically.
Our proposal shows that there is no need to annotate every web page, to
extract information with associated semantics. In this sense we are contributing
to bring together the community of agents programmers and the semantic web.
In the future, we are going to work on an implementation of our framework
in which data sources can be more heterogeneous (databases, news servers, mail
servers, and so on). Extraction of information from multimedia sources such as
videos, images, or sound files will be also paid much attention.
References
1. DARPA (Defense Advanced Research Projects Agency). The darpa agent mark
up language (daml). http://www.daml.org, 2000.
2. R. Baumgartner, S. Flesca, and G. Gottlob. Visual web information extraction
with lixto. In 27th VLDB Conference, 2001.
3. T. Berners-Lee, J. Hendler, and O. Lassila. The semantic web. Scientific American,
May 2001.
4. W. W. Cohen and L. S. Jensen. A structured wrapper induction system for ex-
tracting information from semi-structured documents. In Workshop on Adaptive
Text Extraction and Mining (IJCAI-2001), 2001.
5. W3C (The World Wide Web Consortium). Document object model.
http://www.w3.org/DOM/, 2000.
6. O. Corcho and A. Gómez-Pérez. A road map on ontology specification languages.
In Workshop on Applications of Ontologies and Problem solving methods. 14th
European Conference on Artificial Intelligence (ECAI’00), 2000.
7. S. Cranefield and M. Purvis. Generating ontology-specific content languages. In
Proceedings of Ontologies in Agent Systems Workshop (Agents 2001),, pages 29–35,
2000.
8. S.J. DeRose. Xml linking. ACM Computing Surveys, 1999.
9. Finin, T. Labrou, and Y. Mayfield. Kqml as an agent communication language.
Software Agents, MIT Press, 1997.
10. FIPA (The Fundation for Intelligent Physical Agents). Fipa specifications.
http://www.fipa.org/specifications/index.html.
11. H. Garcı́a-Molina, J. Hammer, K. Ireland, Y. Papakonstantinou, J. Ullman, and
J. Widom. Integrating and accessing heterogeneous information sources in tsimmis.
In The AAAI Symposium on Information Gathering, pages 61–64, March 1995.
12. C.F. Goldfarb and P. Prescod. The XML Handbook. Prentice-Hall, 2nd edition,
2000.
13. OMG (Object Management Group). Unified modelling language version 2.0.
http://www.omg.org/uml/, 2001.
14. J. Hendler. Agents and the semantic web. IEEE Intelligent Systems Journal, 2001.
15. C. A. Knoblock. Accurately and reliably extracting data from the web: A machine
learning approach. Bulletin of the IEEE Computer Society Technical Com-mittee
on Data Engineering, 2000.
16. N. Kushmerick. Regression testing for wrapper maintenance. In Proceedings of

the 16th National Conference on Artificial Intelligence (AAAI-1999), pages 74–79,
1999.
17. N. Kushmerick. Wrapper induction: Efficiency and expressiveness. Artificial Intel-
ligence, 118(2000):15–68, 1999.
18. N. Kushmerick. Wrapper verification. World Wide Web Journal, 2000.
19. S. Luke, L. Spector, D. Rager, and J. Hendler. Ontology-based web agents. In
First International Conference on Autonomous Agents, 1997.
20. G. Mecca, P. Merialdo, and P. Atzeni. Araneus in the era of xml. Data Engineering
Bullettin, Special Issue on XML, September 1999.
21. I. Muslea, S. Minton, and C. Knoblock. Wrapper induction for semistructured,
web-based information sources. In Proceedings of the Conference on Automated
Learning and Discovery (CONALD), 1998.
22. J. Odell, H. Van Dyke, and B. Bauer. Extending uml for agents. In AOIS Worshop
(AAAI), pages 3–17, 2000.
23. S. Soderland. Learning information extraction rules for semi-structured and free
text. Machine Learning, pages 1–44, 1999.
24. M. J. Wooldridge and M. R. Jennings. Intelligent agents: Theory and practice.
The Knowledge Engineering Review, 10(2):115–152, 1995.
Towards Open Adaptive Hypermedia
Christopher Bailey, Wendy Hall, David E. Millard, and Mark J. Weal
IAM Group, ECS, University of Southampton, UK

{cpb99r,wh,dem,mjw}@ecs.soton.ac.uk
Abstract. Research at the University of Southampton has extended generalised

Open Hypermedia (OH) models to include concepts of context and behaviour,
both traditionally very important to the Adaptive Hypermedia (AH) domain. In
this paper we re-evaluate Brusilovsky’s pragmatic taxonomy of AH techniques
from a structural perspective. A variety of OH structures are discussed that can
be used to implement the techniques found in the taxonomy. By identifying
common structures we gain a new perspective on the relationship between
different AH techniques.
Keywords. Open Hypermedia, Adaptive Techniques, Hypermedia Structure,

FOHM, taxonomy
1 Introduction
In the late 1980’s, the hypermedia research community developed, amongst others,
two separate research threads; one focusing on Open Hypermedia systems (OH) and
one on Adaptive Hypermedia systems (AH). The AH community arose partly from
the extensive work that had already been conducted into artificial intelligence and
partly from Intelligent Tutoring Systems (ITS). AH researchers are primarily
concerned with using pre-existing methods and techniques found in the fields of AI,
ITS and User Modelling, and extending, combining and merging these ideas to create
complete systems that understand and aid the user in knowledge acquisition.
ITS promoted the development of educational server-side adaptive web-based
systems such as MANIC [19], INTERBOOK [5] and more recently AHA [8]. Other
such server-side systems index web sites [18] or provide personalized interfaces to
large hypermedia systems [9].
AH has also seen the development of client-side adaptive systems that follow users
as they browse the WWW. Examples of these systems include WebMate [6], Letizia
[14] and LiveInfo [15].
A second thread of hypermedia research focused on the Open Hypermedia (OH)
field. Open Hypermedia Systems (OHSs) such as Microcosm [10], Chimera [1] and
DHM [11] separate links from documents, allowing hyperstructure to be processed
separately from the media it relates to.
In recent years the OH community have tackled the issue of interoperability
between different OHSs, in particular the development of the Open Hypermedia
Protocol (OHP) [7]. The scope of the OHP project evolved into an attempt to create a
reference model and architecture for Open Hypermedia Systems in general.

Towards Open Adaptive Hypermedia 37
This change has focused the OH community on the fundamental structures that
such systems deal with, which has resulted in the promotion of structure to a first-
class status and the consideration of how context might affect that structure. In
particular the development of the Fundamental Open Hypermedia Model (FOHM)
[17] deals directly with issues of context, and behaviour that can modify context.
The most recent OHS developed at Southampton, Auld Linky [16], is a contextual
structure server that stores and serves structures expressed in FOHM.
1.1 Motivation
Our involvement in OH began with the Microcosm system in 1989 [12]. Although it
was not designed as an AH system, Microcosm did provide a framework for building
adaptive hypermedia systems. We first reported the use of Microcosm to develop an
educational adaptive hypermedia application in [13]. More recently we have
developed an adaptive hypermedia system that uses a collection of agents to process
the contents of pages visited by the user and augment the current page with contextual
links to related information [2].
It is our early experiments with using Linky to support AH [16] that have led us to
the observation that many of the current AH techniques can be described and
supported with a simple set of contextual structures.
The purpose of this paper is to produce an alternative view of the AH domain from
a contextually aware OH perspective, to this end we will be using Brusilovsky’s
taxonomy of adaptive techniques [4]. We will show that many of the techniques in the
taxonomy can be implemented with a small key group of hypermedia structures. This
structural perspective is used to analyse and critique the taxonomy and inform the
continued development of our contextual link server.
2 Techniques of Adaptive Hypermedia
Figure 1 shows Brusilovsky’s taxonomy [4], updated from [3]. This diagram will be
used as a basis for a structural comparison of OH and AH techniques.
The taxonomy focuses on the interface and user interaction, and has been divided
into two distinct areas: ‘Adaptive presentation’ and ‘Adaptive navigation support’.
Adaptive navigation support focuses on aspects of navigational hyperlinks such as
generation, appearance, spatial placement and functionality.
Adaptive presentation systems rely on information chunks (or fragments) that can
be processed and rendered in a variety of ways depending on the user preferences. In
broad terms, Adaptive navigation support is about links and Adaptive presentation
about content.
While the distinctions of the taxonomy are important for identification and
classification of adaptive systems, the implementation of these techniques can be
achieved using a small selection of fundamental data structures that can be combined
to create powerful AH systems.
38 C. Bailey et al.
Adaptive Multimedia Natural Language Inserting/removing

Presentation adaptation fragments
Adaptive Adaptive text

Altering fragments
Presentation presentation
Adaptation of Canned text

Stretchtext
modality adaptation
Sorting fragments
Adaptive Hypermedia
Technologies
Direct guidance
Dimming fragments
Adaptive link sorting Hiding
Adaptive navigation
Adaptive link hiding Disabling
support
Adaptive link
Removal
annotation
Adaptive link
generation
Map adaptation
Fig. 1. Updated taxonomy of adaptive hypermedia technologies [4]
3 Structures for Adaptive Hypermedia
In this section we present a selection of hypermedia structures that could be used to

represent the underlying information within an adaptive system. It is not our intention
that these structures be exhaustive, more that they provide enough richness to support
Brusilovsky’s taxonomy. The structures are represented using the Fundamental Open
Hypermedia Model (FOHM).
3.1 Data
The Data object (shown in Fig. 2) is the basic building block of FOHM, it can
represent any piece of media (text, graphic, video stream, etc.). A Context object can
be attached to the data and describes the context in which the data item is visible (the
precise format of a context object is not specified in FOHM although one might
imagine a set of key/value pairs as a basic format). When used in an AH system, the
context of the data item would be matched against a user’s profile to see whether the
data item should be hidden from the user.
Data
Context
Behaviour
Fig. 2. FOHM Data structure
In addition to context, Behaviour objects can be attached to a Data object. These

describe an action that occurs as a result of an event associated with the object they
are attached to. In an AH system this can be used to modify the user profile, for
example, in the case of data, the behaviour event might be ‘on display’ with an action
of setting a flag in the user profile stating that the information in that data item has
been read. Any technique from Brusilovsky’s taxonomy that manipulates content
might utilise data objects, while context and behaviour can be used for modelling pre-
requisite information in the system.
3.2 Link
Navigational link structures are represented in FOHM as Associations, bound to these

are Reference objects that are used to address Data objects or other Associations.
Figure 3 shows a Link with three Data members, the first is bound as a source and the
other two as destinations.
Association Link
SRC DEST DEST
Binding
Reference
Context
Behaviour
Data text text text
Fig. 3. FOHM Navigational Link
Context objects can be placed on Associations in a number of positions. When

attached to the link, it can be used to scope the link (i.e. the link is only visible to the
user if their profile matches the link’s context). If the context is placed on the
destination bindings instead, then while the link will always be visible, the
destinations will change depending on the profile.
Behaviour objects can also be attached to Associations. For example, if we attach a
behaviour with the event ‘on traversal’ then, when the user follows the link, the action
specified in the behaviour object would be enacted by the system.
Contextual link structures could be used to implement any of the ‘Adaptive Link
Hiding’ techniques described by Brusilovsky. For example a link could be authored
with a context object attached that specified that it was only visible to an expert.
When a novice queries the system for links, their user profile does not match against
the link’s context and the link is effectively hidden.
3.3 Tour
A Tour is an association that represents a set of objects designed to be viewed in

sequence. These objects might be data items representing a sequence of pages,
association objects representing a sequence of sub structures or a mixture of both.
40 C. Bailey et al.
Tour
1 2 3 4 5
frag- frag- frag- frag- frag-

ment ment ment ment ment
Fig. 4. Pruning the hyperstructure using context. The black contexts have failed to match
leading to the pruning of the grey sub-structures.
A tour of data objects could be used to implement Brusilovsky’s

‘Inserting/Removing Fragments’. The client would query the link server to extract the
structure shown in Fig. 4. Before being returned by the link server, it is pruned to
remove sub-structures whose context does not match the user profile. The client
inserts the remaining sub-structures, in the order they appear in the tour, into the
document.
3.4 Concept
A Concept association can be used to collect together multiple objects that represent
the same conceptual entity (for example the same piece of text in different languages).
Brusilovsky’s ‘Adaptation of Modality’ could be implemented with such a structure.
Different media representations of the same information (e.g. video, audio, text) could
be gathered together in a concept and then selected via querying with a particular user
profile. When the concept structure is returned it will contain only those members that
fit the query profile. The concept might be structured to be mutually exclusive, or to
allow for multiple representations to be applicable at any given time.
3.5 Level Of Detail (LoD)
Like the concept, the LoD structure associates multiple representations of the same
object. Unlike the concept structure, these representations are ordered within the
association, from the simplest representation to the most complex. When queried in
context the members of the LoD will be filtered. The application can then choose
which of the remaining bindings to display according to what level of detail the
current user would prefer.
For example, a system could support ‘Stretchtext’ from the taxonomy by storing a
LoD structure containing data objects representing text at increasingly advanced
levels. Initially the lowest positioned data object in the LoD would be displayed.
When the user selects the text it could be replaced with the next available data object
in the LoD.
3.6 Equator “City” Example
It is possible to combine these structures to produce more advanced effects (for

example a tour of concepts). Before we go any further it is useful to see how they
have been used in a real-world application.
City is a museum-based sub-project of the Equator Interdisciplinary Research
Collaboration (IRC). The premise of the City project is that data objects represent real
world locations so that adaptive structure can be authored about them. These
structures can then be queried contextually to obtain personal views of the physical
space (interpreted for the user via a device such as a PDA). Figure 5 shows the
structures that we are using.
Tour
1 2 3
Transition Transition Transition

src exp dst src exp dst src exp dst
location location location location

A B C D
text Concept LoD

1 2 3
video audio text audio audio audio
Fig. 5. FOHM Tour, Concept, LoD and Transition structures
The fundamental structure used in City is a Transition. This is very similar to a link
in that it indicates a navigational move from one Data object (src) to another (dst).
Unlike a link it also includes an explanation (exp) member that explains why that
move is relevant. For example, when moving from A to B, the explanation might
describe in what ways the objects at location B are related to those at location A.
We have used the structures previously described along with these transitions to
provide AH functionality. Curators of the museum space can organise sets of these
transitions into tours that describe a number of steps that a visitor might like to take
around the museum space. At any particular location the visitors’ PDAs would
present the available transitions that are available given the users current context,
including any tours that progress from that point.
We can also use the concept and LoD structures to tailor the explanations of
transitions. For example on the B to C transition there are explanations available in
several formats, while on the C to D transition there are several audio explanations of
different lengths and complexity. When retrieving these structures, the users context
will determine which explanation they will actually see.
42 C. Bailey et al.
4 Revisiting the Taxonomy
Having discussed the fundamental structures that can be used to provide adaptivity
and seen how they can be used in a novel AH application, Fig. 6 illustrates how these
structures could be used to implement the various adaptive techniques described in
Brusilovsky’s taxonomy.
The legend in Fig. 6 is divided up between the two fundamental objects (data and
links) and the extended FOHM structures. In addition, some of the techniques in the
taxonomy require extra resources such as System Information (e.g. user access logs),
or a Weight Metric that describes not only which parts of the structure match in a
particular context but also how well they match.
Fig. 6. A structural view of Brusilovsky’s taxonomy that shows how FOHM structures can be
combined to implement a range of adaptive hypermedia techniques.
Each technique in the taxonomy is built up using the structures or resources

required to implement that technique. For example, ‘Adaptive link annotation’ would
require explicit link structures to store each hyperlink and collections of alternative
annotations for each link would be stored in a concept structure. Finally the adaptive
system would analyse the system information (in the form of user history) when
choosing which annotation to present to the user.
The only visual change to the taxonomy that was required, involves duplicating
‘Direct Guidance’ (labelled a on Fig. 6) as it was felt that this technique has two
possible implementations.
The first approach could use a tour of the information domain; the adaptive ‘next’
button would therefore point to the next appropriate destination on the tour. An
alternative approach would be to apply a link sorting routine (requiring a weight on
each link) and then automatically select the link with the highest weight as the next
destination.
One of the philosophies behind OHSs is that there is a general view of data that
covers all media types. From this point of view, it can be seen that any of the
‘Adaptive text presentation’ techniques described in the taxonomy can also be applied
equally well to ‘Adaptive multimedia presentation’. To improve clarity, both
techniques could be combined and re-labelled ‘Adaptive Media Presentation’. This
issue is highlighted as b in Fig. 6.
The same perspective leads us to question the way that Natural Language (NL)
Adaptation (c in Fig. 6) is placed under text adaptation. Because the same techniques
can be used with any media type, we believe that NL adaptation falls under a larger
umbrella of ‘Adaptive Sequencing’, which might use canned or constructed
information fragments. Moreover, NL adaptation influences nearly every single
technique in the taxonomy. Whenever information fragments (or pages) are
sequenced together, there is a need to conserve the progression of the narrative flow.
In these situations, NL techniques, and wider sequencing methods in general, can be
used as the glue to join fragments together.
Another consideration is that the various subcategories of ‘Adaptive link hiding’
(hiding, disabling and removal, labelled d) are all structurally equivalent. This means
that in a system that uses navigational link objects to implement a particular
technology, for example link hiding, it has in place all the requirements needed to
implement any of the other adaptive link techniques.
The diagram also shows the apparent similarity between ‘Adaptation of modality’
and ‘Altering fragments’ labelled e in Fig. 6. These two techniques are functionally
identical if one considers that fragments can contain multiple media representations of
the same data objects. In such cases, choosing the best media type to display
(adaptation of modality) is a process of selecting one fragment from a set of
fragments (altering fragments).
4.1 Reflections on Brusilovsky’s Taxonomy
Brusilovsky’s taxonomy originally provided a mechanism for classifying the various

AH systems at the time. Since then, more systems have been developed, some of
which fit in to the existing taxonomy, and others that have forced extensions to the
taxonomy making it more lengthy and complex.
With our focus on structure we propose that many techniques described by the
taxonomy as textual are in fact applicable to a wide range of media.
It initially seems clear that the taxonomy divides structurally into two halves. The
upper half is concerned with content, and requires a contextual data representation;
the lower is concerned with navigation, and requires links with contextual
membership. However, we have argued that Natural Language Adaptation impacts on
both halves of the taxonomy and therefore blurs the division of the taxonomy into
presentation and navigation. In our opinion both adaptive presentation and adaptive
navigation are about adaptive sequencing. The first is concerned with intra-document
44 C. Bailey et al.
adaptive sequencing, the second in inter-document adaptive sequencing. The

difference concerns the presentation of the sequence i.e. as a linearisation or as a
hyperstructure.
Other aspects of the taxonomy can also become blurred when viewed with a
structural eye. In particular Map Adaptation seems a vague term that is concerned
with the visualisation of many other aspects of the taxonomy (such as link hiding,
altering fragments etc.) rather than being a category in its own right.
4.2 Reflections on Auld Linky
Auld Linky was designed as a contextual link server to deal with all the structures
discussed in this paper. However, it is unable to totally support the taxonomy of
techniques.
Linky is designed only to store and serve structure and for any particular problem
clients need to be written that interpret the structures served by Linky and present
them to the user. On its own Linky supports the first five criteria in the legend of
Fig. 6 (the contextual structures). If a client were written that maintains a basic user
history then it would also be able to support the sixth criteria, System Information.
We feel that this division of responsibility is correct.
The reason Linky cannot support the entire taxonomy of techniques is because it
lacks support for the seventh criteria, Weight Metrics. Currently Linky’s matching
process is binary (if a structural element does not match the users current context then
it is culled from the context) and therefore provides no feedback on the quality of the
match. The client is left unable to sort the resulting structure; a process needed to
implement two of the techniques in the taxonomy, ‘Adaptive Link Sorting’ and
‘Sorting Fragments’.
5 Conclusions
While both the OH and AH communities exist largely independently of each other,
there are many areas where crossover could yield new research directions and offer
solutions to shared problems. In this paper we have used Brusilovsky’s taxonomy as a
means to show how the structural OH approach could inform AH research, but also to
explore what our own OH research, particularly the development of our contextual
structure server, Auld Linky, could learn from established AH techniques.
OH’s emphasis on structure provides the means to implement a wide range of
adaptive technologies and helps clarify the taxonomy. However it neither completely
covers the taxonomy, nor provides the best programming solution in every situation.
As we have shown, to implement every AH feature would require additional
information such as system information and weight metric’s.
We believe that AH systems that acknowledge the structural equivalence of many
AH techniques have an advantage in that they may handle adaptation consistently
across different techniques and media.
We also believe that any contextual OH server needs to provide for the problem
domains being explored in AH research. To this end we are exploring the possibility
of adding contextual weight metrics to the structures served by Auld Linky to allow it
to support all the techniques described by Brusilovsky.
Acknowledgements
This research is funded in part by EPSRC Interdisciplinary Research Collaboration

(IRC) project "EQUATOR'' GR/N15986/01. Thanks to the Equator City Project,
particularly Matthew Chalmers and Ian McColl of the University of Glasgow.
References
1. Anderson, K.M., Taylor, R.N. and Whitehead, E.J. (1994). “Chimera: Hypertext
for Heterogeneous Software Environments”. In ECHT ’94. Proceedings of the
ACM European conference on Hypermedia technology, Sept. 18-23, Edinburgh,
Scotland, UK, pp 94-197, 1994.
2. Bailey C., El-Beltagy, S.R. and Hall, W. (2001). “Link Augmentation: A Context-
Based Approach to Support Adaptive Hypermedia”. In Proceedings of the 3rd
Workshop on Adaptive Hypertext and Hypermedia, August 14-18, Århus,
Denmark, LNCS 2266, pp. 239-251.
3. Brusilovsky, P. (1996). “Methods and Techniques of Adaptive Hypermedia”.
Journal of User Modelling and User-Adaptive Interaction 6, n 2-3, pp 87-129.
4. Brusilovsky, P. (2001). “Adaptive hypermedia”. User Modeling and User-Adapted
Interaction, Ten Year Anniversary Issue 11 (Alfred Kobsa, ed.), pp 87-110.
5. Brusilovsky, P., Eklund, J. and Schwarz, E. (1998). “Web-based education for all:
A tool for developing adaptive courseware”. Computer Networks and ISDN
Systems. Proceedings of 7th International World Wide Web Conference, April 14-
18, 30 (1-7), pp 291-300, 1998.
6. Chen, L. and Sycara, K. (1998). “WebMate: Personal Agent for Browsing and
Searching”. In Proceedings of the 2nd International Conference on Autonomous
Agents, St. Paul, MN, May. ACM Press, New York, NY. pp. 132-139.
7. Davis, H.C., Rizk A. and Lewis A.J. (1996). “OHP: A Draft Proposal for a
Standard Open Hypermedia Protocol”. In Uffe Kock Wiil and Serge Demeyer,
editors, Proceedings of the 2nd Workshop on Open Hypermedia Systems, ACM
Hypertext ’96, Washington, D.C., March 16-20, 1996.
8. De Bra, P. and Calvi, L. (1998). “AHA! An open Adaptive Hypermedia
Architecture”. The New Review of Hypertext and Multimedia 4, 115-139.
9. Espinoza, F. and Höök, K. (1997). “A WWW Interface to an Adaptive Hypermedia
System”. Presented at PAAM’96, London, UK, April 1996.
10. Fountain, A.M., Hall, W., Heath, I. and Davis, H.C. (1990). “MICROCOSM: An
Open Model for Hypermedia With Dynamic Linking”. In A. Rizk, N. Streitz, and
J. Andre, editors, Hypertext: Concepts, Systems and Applications (Proceedings of
ECHT’90), pp 298-311. Cambridge University Press, 1990.
11. Grønbæk, K. and Trigg, R.H. (1994). “Design issues for a Dexter-based
hypermedia system”. Communications of the ACM, 37(3) pp 40-49, Feb. 1994.
46 C. Bailey et al.
12. Hall, W., Davis, H., Hutchings, G. (1996). “Rethinking Hypermedia: The
Microcosm Approach” Electronic Publishing Series, No 4. Kluwer Academic
Pub; ISBN: 0792396790.
13. Hothi, J. & Hall, W. (1998). “An Evaluation of Adapted Hypermedia Techniques
using Static User Modelling”. In Proceedings of the 2nd Workshop on Adaptive
Hypertext and Hypermedia, Pittsburgh, USA, June 20-24, pp 45-50, 1998.
14. Lieberman, H. (1995). “Letizia: An Agent That Assists Web Browsing”. In
Proceedings of the International Joint Conference on Artificial Intelligence,
Montreal, August 1995.
15. Maglio, P.P. and Farrell, S. (2000). “LiveInfo: Adapting web experience by
customization and annotation”. In Proceedings of the 1st International Conference
on Adaptive Hypermedia and Adaptive Web-based Systems. AH2000, Trento,
Italy, August 2000.
16. Michaelides, D.T., Millard, D.E., Weal, M.J. and De Roure, D., C. (2001). “Auld
Leaky: A Contextual Open Hypermedia Link Server”. In Proceedings of the 7th
Workshop on Open Hypermedia Systems, ACM Hypertext 2001 Conference.
Aarhus, Denmark 2001.
17. Millard, D.E., Moreau, L, Davis, H.C. and Reich, S. (2000). “FOHM: A
Fundamental Open Hypertext Model for Investigating Interoperability Between
Hypertext Domains”. In Proceedings of the ’00 ACM Conference on Hypertext,
May 30 - June 3, San Antonio, TX, pp 93-102, 2000.
18. Perkowitz, M. and Etzioni, O. (1999). “Towards adaptive web sites: Conceptual
cluster mining”. In Proceedings of the 17th International Joint Conference on
Artificial Intelligence, 1999.
19. Stern, M.K. & Woolf, B.P. (1998). “Curriculum Sequencing in a Web-Based
Tutor”. In the Proceedings of Intelligent Tutoring Systems, 1998.
GAS: Group Adaptive System
Maria Barra1 , Paul Maglio2 , Alberto Negro1 , and Vittorio Scarano1

1
Dipartimento di Informatica ed Applicazioni
Università di Salerno
Baronissi (Salerno), 84081, Italy
{marbar,alberto,vitsca}@dia.unisa.it
2
IBM Almaden Research Center
San Jose, California, USA
[email protected]
Abstract. This paper describes an ongoing research project to design a

Group Adaptive System (GAS) for collaborative navigation on the web.
Our objective is to provide information that adapts to web users based
on automatically determined, dynamic user groups. To do this, our ap-
proach takes into account the structure of the web (using the CLEVER
algorithm [16]) and user interactions when navigating the web (gathered
by our collaborative environment) to offer users in a consistent infor-
mation space determined by their interests and activities. In the end,
our system, GAS, enables users to perform asynchronous collaborative
navigation by combining resources discovered by a group of users and
suggestions discovered by the CLEVER algorithm to provide recommen-
dations to the group.
1 Introduction
Collaborative navigation on the web can improve information-finding by making
users aware of others’ information activities. Collaborative navigation can be
used, for example, to provide users with relevant resources others have found
automatically, and to enable users to share information deliberately.
We all know that finding information on the web can be time consuming
and frustrating given the vast number of resources available. Though people use
the web to find information all the time, their experiences are rarely captured
and used to guide or inform others. Of course, in most other information-finding
contexts, people routinely rely on the experiences of others to help make choices
and find desired information. For instance, librarians guide library users and
friends ask one another what books to read. Relying on the experience of others
is commonplace. Put simply, the world is a social place and people often rely on
social interactions to find information.
Making the web more of a social place enables users to (1) take advantage
of the other people’s activity, (2) save the time they spent searching for relevant
information on the web (using others people navigation and judgments), (3)
become aware of others’ interests and knowledge, (4) get recommendations of
relevant information based on others’ opinions, (5) use web navigation as a kind

48 M. Barra et al.
of communication channel. Social navigation can be direct (people give recom-

mendations or guide each other) or it can be indirect (the system creates a feeling
of awareness for other people’s activities in an information space [12]). Informa-
tion on preferences and judgments aggregated over a community of users, is a
valuable tool to make navigation decisions for users who have shared interests
(e.g., discovering new resources is a useful group activity to share and to learn
from). Thus, collaborative navigation can promote knowledge and information
exchange among users as well as creating awareness of others on the web.
Recommendation systems on the web often suggest products to customers
based on their explicit and implicit preferences, similarity of user interests, and
consumer and product attributes (e.g., user skills preferences, titles, authors,
brands). These systems have been widely studied and constitute a large number
of web-based applications, including on-line stores that sell books, movies, or
music, and on-line newspapers, newsgroups, or scientific articles services that
tailor content to individual users. Though such domains are huge, the space of
elements are enumerable, and therefore can easily be classified to provide reason-
able customization. Applying the approach of such specialized recommendation
systems to the more general problem of web browsing or searching (i.e., recom-
mending web pages based on many users’ histories of web page navigation) is
not straightforward. One major obstacle is that the information on the web is
not fixed or under the control of a system designer, which makes it difficult to
classify items into domains of interest. In addition, users navigate the web in
many different contexts and for many different reasons, making it difficult even
to know what content area might be of interest.
1.1 Group Adaptive System

Our Group Adaptive System offers users a collaborative navigation environment
on the web, enabling group- and web-resource adaptation to users based on their
interactions and behavior. Rather than classify documents and rely on static or
manually developed profiles for users and groups, we chose to rely on (1) user
interactions with pages and with other users, and (2) the underlying structure
of the portions of the web that users interact with. This choice was motivated
by the fact that reasons for web navigation can be as different as documents on
the web itself, making stereotypes limited relative to users’ interests.
In our system, users navigate independently on the web but belong to groups
joined by common interests, which are determined through resource access and
other interactions. Group membership depends on (1) resources accessed and
the information clusters these resources belong to, (2) the group related to these
information clusters, (3) the interactions among users in each group, and (4)
the group the user currently belongs to. To make such a system work, we are
studying users’ interactions with web resources to measure user interest in the
information they see. We intend to correlate user interest with group interest
and use this correlation to determine group membership.
In this paper, we report on the status of our project. Our prototype Group
Adaptive System (GAS) has been tested with a group of users. The system allows
GAS: Group Adaptive System 49
users to interact through web resources (message exchanges, rating, providing

shortcuts, etc.) and presents a common information space or “Collaborative
View” that is, determined by a rough classification on the number of interactions.
In the end, the GAS is meant to provide users (1) awareness of others and (2)
authoritative recommendations focused on group interest.
In the rest of this section, we describe the model we are currently testing to
provide recommendations. In Sect. 2, we describe related work. In Sect. 3, we
briefly describe the prototype architecture and its implementation. In Sect. 4,
we report on a first validation study of the model based on data collected during
a usability test. Finally, future work is outlined in Sect. 5.
1.2 CLEVER-Based Model

Our model for determining group adaption of web resources with respect to
group interest is based heavily on Kleinberg’s HITS [16] and CLEVER [10] al-
gorithms for discovering authoritative sources of information (authorities and
hubs) leveraging on the linked structure of the web. In this algorithm, a query
(standard result by text-based search engines) is a subgraph of nodes (HTML
pages) and edges (HTML links) augmented with several “close” nodes. In partic-
ular, the algorithm evaluates a starting set of a few (200) pages (called “core”)
by submitting a text query to a standard search engine, enlarging the core by
adding pages that point to or are pointed to by pages into the core.
Initial authority and hub weights are set to 1 for each node. Then an itera-
tive algorithm is run on the graph, alternatively evaluating authority and hub
weights depending on neighbors’ weights. The algorithm is shown to converge
in identifying good authorities (nodes well recognized as good sources) and good
hubs (nodes well recognized as repository of links to good authorities). This al-
gorithm takes into account the structure of the underlying portion of the web,
relying on the authors of the pages or nodes to help determine “good pages” on
a topic. The key point is to leverage the hundreds of authors of pages that know
about the topic: placing a link from a page toward another infers authority on
a resource that is believed (by the author) to be relevant.
Recommender and adaptive models usually do not take such things into ac-
count to provide users with suggestions or to cut down the amount of information
available on a topic (filtering). Thus, our approach adds this dimension to the
common information space created by the group participants’ navigation. We
consider documents navigated by the group as the result of a broad-topic query
by the users to the search engine, and apply the CLEVER algorithm to these,
obtaining recommendations that are based not only on interactions of the group
with pages but also on the pages own structure. In fact, a link X ⇒ Y is meant
as “inferring authority from X on Y ” and is explicitely chosen (suggested) by
X’s author that knows both resources. In this way, if user U1 and user U2 access
page P1 and page P2 , (authored, respectively, by A1 and A2 ) and both pages
point to a page P3 (which was not actually visited by either user) the algorithm
can “discover” P3 and propose it as a possibly useful resource to U1 and U2 (or to
any other similar user). Since this suggestion derives, in a certain way, both from
50 M. Barra et al.
U1 and U2 accesses and from A1 and A2 choices, it is widely supported by an

authoritative chain that begins at users, passes through page authors and iden-
tifies a useful resource. Put another way, we consider the HTML pages accessed
by the group (representing group experience and current interests) as the core
(i.e., the results of the standard text-based query in CLEVER), enlarge it with
the pages that point to the set itself and those that are pointed to by it (using
authors’ “experience” in terms of links established), and then run the algorithm.
This can help to identify authoritative resources that might be relevant for the
group.
Now, the cooperation among the group members steps in. In fact, in every-
day life, people rely on recommendations from others, such as movie and book
reviews printed in newspapers, or general surveys such as restaurant guides [22].
The Group Adaptive System’s goal is to use the structure of users’ behavior to
present group members with a shared space relevant to the group. The common
information space is a view of group knowledge adapted to participants’ behavior
(i.e. the authorities and hubs previously discovered), enhancing the awareness of
the community’s activity and providing participants a place for interaction by
allowing them to create new relationships among people and resources. Users’
interactions, such as adding links, exchanging messages or explicitely ranking a
resource (or even the simple navigation of a suggested page), add information
to the model and provide feedback that is, then, taken into account in the next
evaluation of authorities and hubs. The initial authority and hub weights, that in
CLEVER algorithm are set to 1, can, in fact, be changed according to a weighted
measure of all the interactions taken.
As described in Sect. 4, we tested the first part of the model (i.e., running
CLEVER by using as core a set of pages navigated by the group) and the results
are encouraging. The next step will be to further refine the already significant
suggestions by determining an initial weighted measure of interactions for au-
thority and hub weights. Finally, the model will be tested in a real case study.
2 Related Work
Social navigation in information space is a relatively new research direction that

examines user behavior and user interaction in information-seeking. The objec-
tive of social navigation work is to use information about other people’s behavior
and suggestions to decide what to do. In short, the hope is to support people
taking advantage of the navigation of others who have similar tastes or interests.
Recommender systems are a branch of social navigation systems [12,18,22].
In a typical system, people provide recommendations as inputs, which the system
then aggregates and directs to appropriate recipients. In this case, the main step
is in the aggregation. User evaluations may be expressed by a single bit (good or
bad) or by unstructured textual annotations. Recommendations may be entered
explicitly or may be gathered by implicit evaluations, and even may be anony-
mous. Evaluations can be aggregated through content analysis and weighted
voting. Aggregated evaluations may be used to select items, or to filter out neg-
ative recommendations, or may accompany items in a display. In collaborative

filtering, the goal is to predict the preferences of one user based on the prefer-
ences of a group of users. Here, the main problem is the ability to make good
matches between those who recommend and those seeking recommendations.
GroupLens [15] is a collaborative filtering system for net news that gathers,
disseminates, and uses ratings from some users to predict other users’ interest in
articles. It is based on the simple idea that people who agreed in their subjective
evaluation of past articles are likely to agree again in the future. GroupLens
asks its users to rate the articles they read, and uses ratings in two different
ways: (1) correlates them to determine which users’ ratings are most similar
to each other, and (2) predicts how well a user will like new articles based on
rating from similar users. Another system, Fab [1,2], combines collaborative and
content-based filtering by maintaining users profiles based on content analysis
and directly comparing these profiles to determine similar users. With Fab, users
receive items both that score highly against their own profile and that are rated
highly by a user with a similar profile.
Mathe and Chen [17] propose a method to adapt information access to in-
dividual user profiles based on contextual information relevance. They rely on
a descriptive user model containing tasks and user names to describe profiles,
and on an information relevance network incrementally built from user feedback.
The model records the relevance of references based on user feedback for spe-
cific queries and profiles, and derives relevant references for similar queries and
profiles.
Swiki [14] is a collaborative web server in which every user can easily modify
and extend pages by simply clicking on a link labelled “Edit this page” and then
modifying the page in a web form. Swiki maintains a list of recently modified
pages accessible by authorized users. This server has also been modified to add
functionality to enhance the awareness of the community’s activities through
annotations posted on the documents list. These annotations are interpreted as
footprints pointing out links the lead to pages that have been used recently and
that give the feeling of the activity of a virtual community. These tools have
been shown to be most useful within a small close group of users, such as a work
group, where people are already know of each other’s interests.
Plastic [21] is a new web site in the Automatic Media Network. It is a live
collaboration between the web readers and editors. Operating somewhere be-
tween anarchy and hierarchy, it is a place to suggest and discuss news, opinions,
rumors, humor, and anecdotes online. Drawing from the best material posted
to Plastic’s sections by editors, the Plastic homepage gives a summary view of
what people at Plastic are talking about at any given moment. The goal is to
allow people to share information, with the hope that the stories and reader
comments will engage people in conversation with readers and editors.
52 M. Barra et al.
3 Group Adaptive System: Prototype
This section describes GAS, the prototype for the Group Adaptive System, as
it was developed to support our user study (described in the next section) to
validate its interface and, later, the model described in Sect. 1.1. It must be
also noted that the prototype tested (and described here) did not include a
sophisticated model for automatically determining suggestions.
The prototype is a proxy server infrastructure that lies between user clients
and web servers. GAS offers collaborative and adaptive web navigation to users
through a collaborative interface added to pages (viewed in the browser). The
collaborative interface enables information exchange in the form of documents
and comments among participants through (1 ) a “common information space”
adapted and created based on participants’ interactions (annotations) and (2 )
a “communication channel” obtained by annotating the resources navigated to.
In particular, GAS allows group participants to navigate the web indepen-
dently, using the interaction tools to augment the page contents with shortcuts
to other resources, comments (ideas, opinions, clues), scores, messages, as well as
to observe other people’s behavior by accessing trails left on the pages. Because
the number of resources involved in navigation is incremental, these are dynam-
ically filtered and tailored to represent the state of the group navigation (new
resources, new users, new annotations). When users access to web, the “data
collector” stores their interactions, the “adaptive engine” uses these to deter-
mine those resources that are “relevant” for the group, and the “presentation
module” builds the common information space to be delivered to participants.
GAS is intended to make participants aware of each other and also to sug-
gest relevant information and provide an access point for future navigation. In
particular, the common information space or Collaborative View is a source of
information based on group interests and, it is a place from which participants
can gather others’ ideas and knowledge. The accessed documents provide aware-
ness of others through footprints1 that GAS embeds into pages.
In order to determine group relevant resources, and elaborate the common
information space, GAS (in this preliminary version we are describing) considers
only the number of interactions group members taken on the resource: the higher
the number of these interactions, the more relevant the resource, and once the
number crosses some threshold, the resource becomes eligible to be part of the
common space.
GAS Architecture. GAS’ s distributed architecture is dictated by its charac-

teristics (see also [4]). People must be able to access different servers (we provide
group adaptation of web resources) with different clients (laptop, as phone, as
pc). To achieve this, the system exhibits a three level architecture and uses
a programmable proxy as its middle layer. The proxy knows the client-server
protocol and can manipulate information during transactions. The proxy is not
1
Footprints signal the presence of the group interactions and allow users to access the
contents.
constrained to run at a specified location, but can be placed anywhere [6] along
the information stream. GAS consists of three parts: a data collector, an adap-
tive engine, and a data presentation module. Using GAS is easy: First, the user
sets the browser’s HTTP proxy, registers with the system, and logs in. Once
logged in, the user can access any web page, but the pages will have additional
information embedded into them, namely, the GAS interface and the footprints
of other users.
GAS User Interface. The access point to the system is the GAS registration
and log-in page. Once the user is logged in, he or she can choose to access any
web site (the login phase is required only when the user accesses from a new
browser instance). The GAS user interface consists of the 3 elements a Tools
bar, an Annotations bar, and a Collaborative View. Tools and Annotations bars
are embedded in the pages by including HTML and JavaScript into the sources
without modifying the “original” page display. These consist of a set of links
to the tools to annotate pages and the group footprints (annotations posted by
the participants). The Collaborative View consists of an HTML page (shown on
a separate window) with links to resources that are considered relevant for the
group by the adaptive engine that uses the modified CLEVER algorithm.
GAS Implementation. GAS is implemented using the Web Intermediaries

(WBI) framework [6,7]. WBI is an HTTP request processor that receives the
user’s requests and produces or filters the responses to these requests. More sim-
ply, WBI is a programmable HTTP proxy. The basic programmable unit in WBI
is the MEG (request editor, generator, monitor end document editor), which are
combined together into applications called plugins. GAS is implemented by a
set of WBI plugins: (1) the web server/proxy plugin provides the access to the
system, (2) the authorization plugin provides the users registration and their
identification, and (3) the GAS plugin that provides the collaborative environ-
ment.
The GAS MEGS are triggered when users interact with a web page: col-
lecting and analyzing data, and generating of dynamic contents are their main
functionalities. Users interactions, on documents accessed, are stored in a per-
sistent store mechanism provided by WBI, a hierarchical database much like a
file-system, which contains arbitrary key/value pairs. Pages accessed are stored
in the filesystem by URL. A GAS MEG chain constructs the user interface by
embedding HTML and JavaScript in the pages loaded by the browser. When the
footprints contents are requested by a user, the presentation module MEGs are
triggered, displaying the interface for that query. Different HTML and JavaScript
implement the user interface properly in different browsers (e.g., Netscape and
Internet Explorer).
4 Preliminary Results
During the development of the prototype GAS user interface, we tested usability
with a group of 10 users (faculties, Ph.D and Master degree students) at the
54 M. Barra et al.
University of Salerno, all conducting research in our research lab. This usability
test allowed us to improve the system and better define the user interface based
on feedback and user requirements.
To preliminarly validate the model used to provide suggestions to the users,
we decided to use the data collected during the Salerno students’ navigation
(model validation test). It must be said that almost 5 months passed between
the usability test of the system and our testing of the model with these data.
At the time of the usability test, none of the users knew of the existence of the
model to be used.
Data were collected during the usability test and users navigated for about
a month. First, we examined data subdivided by week. From the results, it
appears that the authorities and hubs the algorithm determines for each week
are different (new resources can always be added as suggestions with the time),
and also that many of the authorities and hubs are not from the core (the pages
determined are unknown, i.e. not visited, to users).
During the week that users navigated most2 , 498 URLs were accessed. A
manual screening of the results showed that many of the hubs and authorities
with higher scores were rather heterogeneous (as well as the pages navigated by
the group), making it difficult to fully establish their validity as suggestions for
the group. Nevertheless, some of the resources suggested (but which were not
navigated by the users) were both meaningful and useful . For example, the first
authority was a site for nVidia Developers about an OpenGL demo on Fluid
Visualization3 , as a group of students was actually working on a project about
OpenGL. The second authority was a soccer site from an Italian provider4 , many
students accessed resources (on other sites) about Italian soccer league results.
Hubs were also rather significant. Although the first hub was a project site in the
Gnome Developer’s site (and almost all of the students were developing under
Linux though none were directly involved with Gnome), the second was the URL
of the FAQ for the newsgroup comp.graphics.algorithms5 , a relevant resource,
for the group that had the graphics project. Other hubs included a description
of Enterprise JavaBeans at CalTech, a very good resource of information about
the web (a personal home page at the CS Dept. in University of Milan, Italy)
and a very good Italian hub for resources about thesis writing (including style
manuals and software manuals on text- and word-processors).
To get a more precise estimation of the validity of suggestions, we decided
to select the two group members who showed the highest level of activity in
the system (number of pages navigated and interactions with GAS), and re-run
the algorithm on only the resources accessed by them during the entire month.
The two users accessed a total of 761 URLs. From the manual analysis of the
results, we were able to determine that many of the hubs and authorities were
2
Note that the system was down intermittently for modifications to improve the
visualization of bars on particularly intricate pages.
3
http://developer.nvidia.com/view.asp?IO=ogl fluid viz.
4
http://calcio.iol.it
5
http://www.faqs.org/faqs/graphics/algorithms-faq/.
indeed fully meaningful, and therefore useful as suggestions given by the system.
For example, analyzing the core of pages navigated, it was clear the two users
were mainly looking for conferences on Cooperative and Adaptive systems, La-
tex manuals (one of the two was writing a thesis), a specific conference to be
held in Finland (and consequent search of travel information and accommoda-
tion). Among the highest ranked hubs and authorities obtained were links to
conference lists and conferences sites on the same topics or related topics, links
to previously unvisited resources to Latex manuals, and a link to an interesting
and particularly useful resource about Finland: a student hostel in Helsinki.
A final comment is necessary: The dynamic nature of some sites (news, ad-
vertisements, etc.), makes some of the suggestions useless after a while. This was
one of the facts that limited the effectiveness of the validation test for the whole
group of users, many of whom accessed news sites (e.g., newspapers) whose out-
going links and backlinks dramatically changed over time and were often broken.
Of course, in a real setting, this is not going to be a problem, as backlinks and
outgoing links can be computed immediately after the resource is accessed .
In the next experiment, we plan to plug users’ interactions and explicit rank-
ing into the CLEVER model. For example, the initial authority weight given
to a node can be incremented proportionally to the number of interactions by
different users and to the ranking given by the users. Moreover, by evaluating
the interactions of users with the resources suggested, we can further provide re-
fined suggestions. Though we believe that the validation of the model provided
us with positive feedback, we also believe that these refinements can improve
the quality of the system’s suggestions.
In this paper we presented the status of the GAS project. We described the model
and prototype, and reported encouraging results from an initial validation test.
We plan to conduct more complete tests to determine the best way to combine
user interactions and the explicit ranking of the resources with a CLEVER-like
model. In addition, we would like to study how best to determine subgroups
of users with more closely related interestes. Known methods of coclustering
information in bipartite graphs (users and resources) can probably be helpful in
this. Subgroup identification will be very useful, as we then can (1) further refine
the suggestions given (e.g., by re-running the CLEVER-based algorithm on the
resources navigated by the subgroup and subgroup member’s interactions), and
(2) allow multiple membership for each user in such a way that suggestions
will relate to the “current” interests of the user as shown by his or her current
navigation pattern.
We are well aware of the implications that a widely used GAS can have on
performance, and thus the need for distributing the load of such a system among
different machines, either on a cluster that team up to provide the service or on a
truly distributed architecture that incorporates several cooperating GASs. One
intriguing scenario is a distributed hierarchy of GASs that act as a proxy chain,
56 M. Barra et al.
each providing suggestions to a larger group of users (in subgroups) before finally
gathering the resources requested by a single user.
References
1. Balabanovic M., Shoham Y.: “Fab: Content-Based, Collaborative Recommenda-

tion”. In Communications of the ACM. Vol. 40, N. 3. March 1997.
2. Balabanovic M.: “An Adaptive Web Page Recommendation Service”. In The First
International Conference on Autonomous Agents. Marina del Rey, California.
February 1997.
3. Barra M., Cattaneo G., Negro A., Scarano V.: “Symmetric Adaptive Customer
Modeling in a Electronic Store”. In Proc. of 3th IEEE Symposium on Computers
and Communications (ISCC). Greece 1998.
4. Barra M.: “Distributed Systems for Group Adaptivity on the Web”. In the Proc. of
AH2000 Inter. Conf. on Adaptive Hypermedia and Adaptive Web-based Systems.
ITC/IRST Trento, Italy. August 2000.
5. Barra M.: “Group Adaptation and Collaboration on the Web”. In Proc. of Inter.
Conference of Computer Supported Cooperative Work (CSCW2000). Philadelphia,
PA. December 2000.
6. Barrett R., Maglio P.: “Intermediaries: New Places for Producing and Manipulating
Web Content”. In Proc. of 7th International WWW Conference. 1998.
7. Maglio P., Barrett R.: “WebPlaces: Adding people to the Web”. In Poster Proc. of
8th International WWW Conference. Canada 1999.
8. Breese J., Heckerman D, Kadie C.: “Empirical Analysis of Predictive Algorithms
for Collaborative Filtering’”. In Proc. of the 14th Conference on Uncertainty in
Artificial Intelligence, Madison, WI, July, 1998.
9. Brusilovsky P.: “Methods and techniques of adaptive hypermedia”. User Modeling
and User-Adapted Interaction, 6 (2-3), 87-129, 1996.
10. Chakrabarti, S. Dom, B., Gibson, D., Kleinberg, J., Raghavan R., and Rajagopalan
S.: “Automatic Resource Compilation by Analyzing Hyperlink Structure and Asso-
ciated Text.” Proceedings of the 7th World-Wide Web Conference, 1998.
11. Dieberger A.: “Where did all the people go? A collaborative Web space with social
navigation information”, Poster at WWW9, Amsterdam, The Netherlands, May
2000.
12. Dieberger, A., Höök K.: “Applying principles of social navigation to the design of
shared virtual spaces’”, presented at WebNet’99, Honolulu, Hawaii, 1999.
13. Erickson T., Laff M. R.: “The Design of the ’Babble’ Timeline: A Social Proxy
for Visualizing Group Activity over Time”. In Proc. of CHI2001. Conference on
Human Factors in Computing Systems. Seattle, Washington, 31 March-5 April
2001.
14. Guzdial M.: “Swiki” Georgia Tech.
http://www.cc.gatech.edu/mark.guzdial/squeak/
15. Konstan J. A., Miller B. N., Maltz D., Herlocker J. L., Gordon L. R., Riedl J.:
“GroupLens: Applying Collaborative Filtering to Usenet News”. In Communica-
tions of the ACM. March 1997.
16. Kleinberg J.: “Authoritative sources in a hyperlinked environment”. In Proc. of 9th
ACM-SIAM Symposium on Discrete Algorithms. San Francisco, California January
25-27, 1998. Extended version in Journal of the ACM 46(1999).
17. Mathe N., Chen J.: “A User-Centered Approach to Adaptive Hypertext based on
a Information Relevance Model”. In the Proc. of 4th International Conference on
User Modeling, UM94. Hyannis, MA 1994.
18. Murno A. J., Hook K., Benyon D.: “Computer Supported Cooperative Work. Social
Navigation of Information Space”. Springer-Verlag. G. Britain 1999.
19. O’ Connor M., Herlocker J.: “Clustering Items for Collaborative Filtering”. In Proc.
of the ACM SIGIR ’99 Workshop on Recommender Systems: Algorithms and Eval-
uation. Berkeley. August 1999.
20. Billsus D., Pazzani M.: “A Hybrid User Model for News Story Classification”. In
Proc of the Seventh International Conference on User Modeling (UM ’99), Banff,
Canada 1999.
21. Plastic: http://www.plastic.com.
22. Resnick P., Varian H.R.: “Recommender Systems”. In Communications of the
ACM,40,3. March 1997.
23. Ungar L. H., Foster D. P.: “Clustering Methods for Collaborative Filtering”. In
AAAI Workshop on Recommendation Systems, 1998.
TV Scout: Lowering the Entry Barrier
to Personalized TV Program Recommendation
1,* 2,*
Patrick Baudisch and Lars Brueckner
1
Inf. Sciences and Technologies Lab.
Xerox Palo Alto Research Center
Palo Alto, CA 94304, U.S.A.
+1 (650) 812 4656
[email protected]
2
IT Transfer Office (ITO)
Darmstadt University of Technology
64283 Darmstadt, Germany
+49 (6151) 16-6217
[email protected]
Abstract. In this paper, we present TV Scout, a recommendation system provid-

ing users with personalized TV schedules. The TV Scout architecture addresses
the “cold-start” problem of information filtering systems, i.e. that filtering sys-
tems have to gather information about the user’s interests before they can com-
pute personalized recommendations. Traditionally, gathering this information
involves upfront user effort, resulting in a substantial entry barrier. TV Scout is
designed to avoid this problem by presenting itself to new users not as a filter-
ing system, but as a retrieval system where all user effort leads to an immediate
result. While users are dealing with this retrieval functionality, the system con-
tinuously and unobtrusively gathers information about the user’s interests from
implicit feedback and gradually evolves into a filtering system. An analysis of
log file data gathered with over 10,000 registered online users shows that over
85% of all first-time users logged in again, suggesting that the described archi-
tecture is successful in lowering the entry barrier.1
1 Introduction
Information filtering systems [7] suffer from a bootstrapping problem. Before they
can give personalized recommendations to a user, they have to find out what the
user’s interests are. Only then can filtering systems build user profiles and compute
personalize recommendations. The problems resulting from this undesirable order of
required user effort and delayed benefit is a well-known phenomenon in collaborative
filtering, the so-called cold start problem [17]. Users are reluctant to invest effort,
especially if they don’t know whether the offered service will be worth the effort.
This approach bears the risk that users will avoid the gamble and stick with a system
offering more immediate benefit, such as a retrieval-oriented system. Users making
this decision, however, will never come to discover the long-term benefits the filter-
* The work presented in this paper was carried out during the authors’ affiliation at GMD-IPSI

TV Scout: Lowering the Entry Barrier to Personalized TV Program Recommendation 59
ing system would have offered. For additional studies on incentive structures and the
results of the lack of incentives see [11].
In this paper, we describe an architecture designed to address this incentive prob-
lem and we will demonstrate this architecture at the example of our TV program
recommendation system TV Scout. We will begin by briefly introducing the field of
TV recommendation. We will then discuss TV Scout and its user interface and dis-
cuss the underlying filtering architecture. Finally, we will report results of an analysis
of TV Scout online usage data, discuss our findings, and present conclusions and
future work.
2 Recommending TV Programs
In 1992, Belkin and Croft wrote “In particular, applications such as the recreational
use of television programming pose special problems and opportunities for research in
filtering” [7, p.37]. Several current trends make TV an interesting application area for
information filtering. TV viewers are facing an information overload situation [10]. A
number of technical improvements, such as cable, satellite, and digital TV technology
have resulted in an increasing number of available TV channels. Today, hundreds of
channels broadcast thousands of programs every day. Since the amount of content that
is of interest for a given viewer has not increased proportionally, planning ones TV
consumption has become a challenge. The amount of TV programs will soon exceed
the limits of what can reasonably be printed and channel surfing is no longer fast
enough to allow getting an overview of all channels [11]. Attempting to meet
the changing requirements, web-based TV program guides (e.g. TV Guide,
http://www.tvguide.com), set-top boxes with electronic program guides (EPGs, [20]),
and digital VCRs (e.g. Tivo http://www.tivo.com) have emerged in the past few years.
There have been several research projects around TV recommendation in the past
[11, 9], but most of them focused on set-top boxes and on the technical possibilities
for monitoring user behavior rather then on web-based systems and usability. Current
research in personalized TV evolves still around personalized EPGs [1], but also
around new concepts, such as multi-agent recommender systems [14]. A more thor-
ough overview of current research in the field of personalized TV recommendation
can be found in [18].
3 TV Scout
TV Scout [3, 4] is a web-based TV recommendation system. Its goal is to support

users in planning their personal TV consumption.
In order to understand the design requirements for such a system, we began our re-
search with an informal survey among students [3]. The survey indicated that expecta-
tions about the functionality of an ideal TV recommendation system were dominated
by experiences with printed TV program guides. While our goal was to eventually
provide users with a personalized TV program at a single mouse click, our survey
indicated that only a minority of the users we had interviewed would be willing to
invest the required effort. We concluded that in order to attract users, a successful TV
60 P. Baudisch and L. Brueckner
recommendation system would first have to emulate the expected print-like function-
ality, as well as the straightforward usage of printed guides: pick up the TV guide,
find today’s listing, pick a program, and watch TV. The challenge was to provide a
seamless transition from this scenario to the filtering functionality we had in mind. To
prevent the filtering functionality from conflicting with the user expectations and
system learnability, we decided to create a system that would progressively disclose
its filtering features to users.
TV Scout with starting page

TV listing
& table retention tools
!
Fig. 1. How TV Scout presents itself to first-time users (screenshots partially translated from
German)
3.1 Implementation
The TV Scout project was conducted in cooperation with the TV program guide pub-
lisher TV TODAY. While this resulted in TV Scout getting implemented as a web-
based system, we see no architectural problems in porting the resulting architecture to
a set-top box. To allow maintaining personal user profile data, first-time users have to
create an account, which they access using a self-selected login name and password.
The web-based TV Scout front end is implemented in HTML, Java, and JavaScript.
3.2 Retrieving Program Descriptions
To users entering TV Scout for the first time, the system presents itself as a retrieval
system. Its functionality at this stage restricts itself to the functionality of a printed TV
program guide, with a graphical user interface. Users specify a query (or simply hit a
button for the default “what’s on now”), sort through the resulting list and select pro-
grams to watch. Users can also print the list of selected programs for later use.
Fig. 1 shows how users accomplish that using the TV Scout user interface. The in-
terface consists of the menu frame on the left and the content frame on the right. The
menu frame provides users with access to all retrieval and filtering functions and is
permanently visible. The content frame is used to display various types of TV listings
and all profile editing tools.
The system is used as follows. Users execute a query by picking a query from the
query menu. Fig.2 shows several close-ups of this menu. In its current version, TV
Scout offers four query groups: text search, genres, user tips, and TV TODAY tips,
plus a favorites group that we will explain later. Text search allows users to search for
keywords using optional Boolean syntax. The other three submenus are executed by
picking the corresponding menu entry. To provide more precise queries, theses query
c
d
e
Fig. 2. The query menu offers four groups of queries

groups contain hierarchies of submenus that can be browsed in a file system explorer-
like fashion. Genres contains a historically grown genre classification of TV pro-
grams, such as sports, comedy, and series [14]. User tips contains recommendations
volunteered by users who serve as self-proclaimed editors, so-called opinion leaders
[4]. Finally, TV TODAY tips, are recommendations provided by the editors of
TV Scout’s printed counter part.
By default, all queries are restricted to the programs starting within the current
hour, but TV Scout provides customized controls that allow specifying arbitrary time
and date intervals using mouse drag interactions (Fig. 4a). Channels can be selected
from two predefined sets or can be selected in detail using a paintable interface (Fig.
4b) [6].
When a query is executed, the resulting set of TV program descriptions (Fig. 1 bot-
tom left) is displayed in the content area. Descriptions consist of the program title, a
rating describing how well the program matches the query, an extract of the program
description, and links to a more detailed description. Users can choose between the
display styles ranked list and table.
Two toggle switches per program description allow users to retain programs they
plan to watch in the so-called retention tool (Fig. 1 bottom left, circled). The retention
tool laundry list can be used to print a list of programs; video labels are designed to
retain and print programs to be videotaped. The retention menu allows users to dis-
play the content of their retention tools for reviewing or printing. The printed list can
be used to remind users of the programs that they plan to watch.
3.3 Filtering Functionality: Creating “Bookmarks”
Using the functionality described so far, the effort for repeated usage is the same each
time the service is used. The next step therefore is for the system to reduce the effort
required of the user when querying, since the primary purpose of IF systems is to be
“time-saving devices” [2].
When a user enters a query that is broader than necessary, the user is forced to sort
through an unnecessarily long listing when trying to find desired programs. When the
system detects that the user has used such a sub-optimal query repeatedly while an-
other query with better precision exists, it makes a suggestion. Fig. 3 shows an exam-
ple. Let’s assume that the user has repeatedly used the query “movies” to exclusively
find and retain comedies and horror movies. By computing the overlap between the
retained programs and all available queries [3], the system detects that the retained
programs can also be covered by the more specific queries “horror movies” and
“comedies”. A dialog box opens and suggests using these queries instead. The user
can execute the suggested queries like any other query, i.e. by clicking their names.
The more important function of the dialog box, with respect to our filtering con-
cept, is that it also suggests retaining these queries as bookmarks. Users can do this by
clicking the toggle switch that accompanies each query (a folder symbol with a check
mark, see Fig. 3a). Retained queries pop up in the user’s favorites (Fig. 3b). The favor-
ites folder is collocated with the other query groups and can be executed the same
way. Retained queries are listed in a flat hierarchy, thereby providing the users with
convenient access to queries that would otherwise be hidden in multiple different
submenus. This functionality corresponds to the bookmark folder in a web browser.
Unlike web bookmarks these bookmarks are stored on the TV Scout server, allowing
TV Scout to use them as input for additional computation.
Retention check boxes accompany all queries in the system (see Fig. 3b), so users
can bookmark queries anytime, independent of suggestions. The primary purpose of
query suggestions is to inform users about the bookmaking concept and to encourage
its usage.
Note the special importance of the retention tools. Although the declared purpose
of the retention tools is to allow users to memorize programs and print schedules,
their primary purpose from the system’s point of view is to serve as an information
source about the user’s interests. The content of the retention tools is considered an
implicit positive rating for the retained programs, making the retention tools serve as
a source of implicit retention feedback [16]. Although implicit feedback is commonly
agreed to be a less reliable source of rating information than explicit feedback, it has
the benefit of being unobtrusive, which we considered essential for this type of filter-
ing system. See [3, 4] for how TV Scout uses the same implicit input for various types
of filtering functionality based on collaborative filtering.
b
a
Fig. 3. By clicking a checkmark-shaped button, queries can be retained in All favorites.
3.4 Filtering Functionality: One-Click Personalized TV Schedules
To provide a container for bookmarked queries is not the only purpose of the favorites
folder. The real value of this folder lies in the fact that users can execute it as a whole
by clicking the top menu entry labeled all favorites. This executes all retained queries
at once. The result listings of the individual queries, however, are not appended to
each other—they are merged into a single relevance-ordered result list. This is the
most powerful function of the TV Scout system—it fulfills the initial goal of generat-
ing personalized TV schedule with a single mouse click.
How are the individual query results merged in order to obtain a useful result?
When the query profile all favorites is executed, a script running inside the TV Scout
server executes all contained queries. This is done by delegating each query to the
corresponding subsystem; text search, for example, is executed by FreeWAIS, while
genre queries are executed by a relational database. As a result, the subsystems de-
liver sets of pairs (program, rating). The task of the query profile script is to merge all
these results into a single ranked list. This requires transforming the individual ratings
such that they include the user’s perceived importance of the interest represented by
the query. In order to express this perceived importance, the query profile stores a
linear function (i.e. a factor and an offset) for each retained query. The resulting rat-
ings are computed by transforming the ratings returned by the subsystem using this
function. If a TV program is returned by multiple queries its ratings are summed up.
Finally, programs are sorted by their result rating and returned to the user.
The critical factor is the parameters of the linear transformation. The system ac-
quires these parameters through initialization, learning, and manual updating. When
queries are bookmarked, their functions are initialized using Zipf’s law [19, p. 60].
This means that more specific queries are given positive offsets, propagating the re-
sults of these queries towards the top ranks of the resulting listings, thus preventing
them from being buried inside the large result sets of less specific queries.
After initialization, the parameters of the rating transformations can be improved
by two means. First, TV Scout continuously optimizes the query profile based on the
same implicit retention feedback that was already used for suggesting queries. See [3]
for a description of the algorithm. Second, interested users are allowed to manually
inspect and update their profile. Clicking the “>>details” link in the all favorites menu
invokes a profile editor. The simplest version of this editor provides users with a
single pull-down menu per query (Fig. 4c), allowing users to assign a symbolic rating
to each query, such as “Action movies are [very important] to me” [3, 5].
Through the use of relevance feedback the query profile improves continuously, so
that the quality of the rankings obtained by clicking all favorites increases over time.
b c
Fig. 4. The TV Scout profile editing tools (a) viewing time profile editor, (b) channel profile
editor, and (c) query profile editor.
3.5 Summary
Fig. 5 summarizes how the usage of TV Scout by a given user can evolve over time.
Each transition to a more personalized phase can be suggested by the system (T1-T3)
or initiated by the user (U1-U3). However, users are not forced through these phases
and may equally well settle with the functionality of one of the earlier phases.
1. Query phase (S1): Users can pick predefined queries (T1) or can formulate
queries, such as text searches, manually (U1).
2. Bookmark/reuse phase (S2): If the system detects reoccurring or sub-optimal
queries it proposes better-suited queries and suggests retaining them as favorites
(U2). Independent of suggestion, users can bookmark queries anytime (T2).
Profile creation (T∗∗): The user’s query profile is created automatically when the
first query is bookmarked.
3. Profile phase (S3): Initially, the query profile provides users with a convenient
way of executing all their bookmarks with a single click. Continuous supply of
relevance feedback (T3) or manual profile manipulation (U3) improves the profile.
T1 system provides T2 system suggests ∗

T∗ system compiles T3 system learns
S1 S2 S3
start queries bookmarks query profile
queries bookmarks query profile
start (one shot state) (reuse state) (filtering state)
(one shot state) (reuse state) (filtering state)
U1 user writes U2 user defines U3 user updates
Fig. 5. Evolving usage of a proposed filtering architecture
4 TV Scout Usage Data
The purpose of the TV Scout design is to reduce the entry barrier for new users by
using a progressive disclosure of the filtering functionality. How can we verify the
success of our interaction design? A controlled experimental comparison with a com-
peting system would be problematic because of the vast amount of interface variables
that would be difficult to control. In addition, modeling a realistic web-usage scenario
in a lab setting is challenging. Alternatively, a naturalistic study of web use would
provide more realistic data, but we would be unable to measure factors such as
subjective satisfaction. Ultimately, we decided to conduct an informal analysis of log
file data from actual web usage.
When we conducted our data analysis April 20, 2000, TV Scout had been publicly
available at for 18 months. The entire 18 months of log file data are included in this
study. All usage data was extracted from the web server log files and the system’s
database. With respect to the filtering functionality, this data was slightly biased, in
that the suggestion feature became available later. Because of this, we expected that
the usage of bookmarking would be underrepresented.
The main purpose of the analysis was to verify whether our filtering system design
fulfilled the primary goal, namely to provide a low entry barrier. If our design was
appropriate, then TV Scout would meet the expectations of first-time users and would
not overwhelm them. Repeated usage would indicate that users had taken the entry
hurdle; one-shot users would suggest the opposite.
We were also interested in learning more about the users’ demand for the offered
filtering functionality. How many users would adopt bookmarking functionality; how
many would make use of their personal query profiles? Based on our informal survey,
we expected the majority to be satisfied with the initial retrieval functionality, but we
had no clear expectations about the percentages. Finally, we were interested in seeing
how useful users would find the query profile. Once they had created one, would they
continue to use it or would they abandon it rapidly?
4.1 Results
At the day we examined the log data, TV Scout had 10,676 registered users. In total,
users had executed 48,956 queries. 53% of all queries (25,736 queries) were specific
queries different from the default query.
Repeated log-ins: We found that 9,190 of the 10,676 registered users had logged in
repeatedly, i.e. twice or more. This corresponds to a percentage or 86% repeated us-
ers. The most active user with 580 logins had logged in almost daily.
Bookmarks: 1770 users had bookmarked one or more queries. Together, these us-
ers had bookmarked 4383 queries, mostly genres. The most frequently executed que-
ries were the genres movies (736 times) and information (364 times), and TV TO-
DAY Movie tips (369 times). Over 300 text searches were bookmarked.
Query profiles: Out of the 1770 users who had bookmarked at least one query, 270
users (about 15%) executed their query profile at least once to obtain personalized
listings. These users executed their query profiles a total of 5851 times, which corre-
sponds to an average of 21 times per user. These users manually fine-tune their pro-
files a total of 1213 times, with an average of 4.5 times per user. These results indi-
cate that query profiles were highly appreciated by those who used them.
5 Conclusions
We interpret the measured percentage of repeated users as a confirmation of our de-

sign. 86% of all first time users logged in repeatedly; we consider this to be a very
high percentage for a web-based system. This indicates that presenting first-time users
with a retrieval setting is a successful approach to keeping the entry barrier for first-
time users low.
Only 17% of users made use of the bookmark feature; out of these, only 15% made
used of the query profile. These numbers seem low even taking into account that the
suggestion feature was not available most of the logged time. Does this result indicate
that the filtering functionality is inappropriate or difficult to learn? Why did the ma-
jority of the users not reach the “goal” of the system?
This is not how we interpret these results. In an earlier TV usage survey we con-
ducted [3] we found TV users to plan their TV consumption for very different time-
frames. Most of these users only planned a TV schedule for the following day or they
did not plan at all. Many users only used a guide to determine what was currently on
TV. Only 12% of the users planned a TV schedule for the entire week. Considering
that the filtering functionality of TV Scout addresses the relatively small subgroup of
users who plan their TV consumption, the observed results seem appropriate. The
majority of users who only used the retrieval functionality may have found the re-
trieval functionality of TV Scout to be the appropriate support for their information
seeking strategy. An online survey as well as an experimental study should help to
verify this interpretation.
Acknowledgements
We would like to thank Dieter Böcker, Joe Konstan, Marcus Frühwein, Michael
Brückner, Gerrit Voss, Andreas Brügelmann, Claudia Perlich, Tom Stölting, and
Diane Kelly.
References
1. L. Ardissono, F. Portis, P. Torasso. F. Bellifemine, A. Chiarotto and A. Difino Architec-

ture of a system for the generation of personalized Electronic Program Guides. In Pro-
ceedings of the UM 2001 workshop “Personalization in Future TV”, July 13 to July 17,
2001, Sonthofen, Germany.
2. P.E. Baclace. Information intake filtering. In Proceedings of Bellcore Workshop on High-
Performance Information Filtering, Morristown, NJ, November 1991.
3. P. Baudisch. Dynamic Information Filtering. Ph.D. Thesis. GMD Research Series 2001,
No. 16. GMD Forschungszentrum Informationstechnik GmbH, Sankt Augustin. ISSN
1435-2699, ISBN 3-88457-399-3. Also at http://www.patrickbaudisch.com/publications
4. P. Baudisch. Recommending TV Programs on the Web: how far can we get at zero user
effort? In Recommender Systems, Papers from the 1998 Workshop, Technical Report WS-
98-08, pages 16-18, Madison, WI. Menlo Park, CA: AAAI Press, 1998.
5. P. Baudisch. The Profile Editor: designing a direct manipulative tool for assembling pro-
files. In Proceedings of Fifth DELOS Workshop on Filtering and Collaborative Filtering,
pages 11-17, Budapest, November 1997. ERCIM Report ERCIM-98-W001.
6. P. Baudisch. Using a painting metaphor to rate large numbers of objects. In Ergonomics
and User Interfaces, Proceeding of the HCI '99 Conference, pages 266-270, Munich, Ger-
many, August 1999. Mahwah: NJ: Erlbaum, 1999.
7. N.J. Belkin and W.B. Croft. Information filtering and information retrieval: two sides of
the same coin? CACM, 35(12):29-37, Dec. 1992.
8. A. Borchers, J. Herlocker, J. Konstan, and J. Riedl. Ganging up on information overload.
Computer, 31(4):106-108, April 1998.
9. D. Das and H. ter Horst. Recommender systems for TV. In Recommender Systems, Papers
from the 1998 Workshop, Technical Report WS-98-08, pages 35-36, Madison, WI. Menlo
Park, CA: AAAI Press, 1998.
10. P. Denning. Electronic junk. CACM, 23(3):163-165, March 1982.
11. M. Ehrmantraut, T. Härder, H. Wittig, and R. Steinmetz. The Personal Electronic Program
Guide—towards the pre-selection of individual TV Programs. In Proc. of CIKM'96, pages
243-250, Rockville, MD, 1996.
12. L. Gravano and H. García-Molina. Merging ranks from heterogeneous Internet sources. In
Proceedings of the 23rd VLDB Conference, pages 196-205, Athens, Greece, 1997.
13. J. Grudin. Social evaluation of the user interface: who does the work and who gets the
BENEFIT? In Proc. of INTERACT’87: pages 805-811, 1987.
14. M. Kuhn. The New European Digital Video Broadcast (DVB) Standard.
ftp://ftp.informatik.uni-erlangen.de/local/cip/mskuhn/tv-crypt/dvb.txt.
15. K. Kurapati, S. Gutta, D. Schaffer, J. Martino and J. Zimmerman. A multi-agent TV
recommender. In Proceedings of the UM 2001 workshop “Personalization in Future TV”,
July 13 to July 17, 2001.
16. D.M. Nichols. Implicit ratings and filtering. In Proceedings of Fifth DELOS Workshop on
Filtering and Collaborative Filtering, pages 31-36, Budapest, November 1997. ERCIM
Report ERCIM-98-W001. Le Chesnay Cedex, France, European Research Consortium for
Informatics and Mathematics, 1998.
17. P. Resnick and H. Varian (Eds.). Special issue on Recommender Systems. Communica-
tions of the ACM, 40(3):56-89, March 1997.
18. Proceedings of the UM 2001 workshop “Personalization in Future TV”, July 13 to July 17,
2001, Sonthofen, Germany. Online at http://www.di.unito.it/~liliana/UM01/TV.html
19. G. Salton and M.J. McGill. Introduction to Modern Information Retrieval. New York:
McGraw-Hill, 1983.
20. H. Wittig and C. Griwodz. Intelligent media agents in interactive television systems. In
Proc. of the International Conference on Multimedia Computing and Systems ’95, pages
182-189, Boston, May 1995. Los Alamitos, CA: IEEE Computer Science Press, 1995.
Adaptivity, Adaptability, and Reading
Behaviour: Some Results from the Evaluation
of a Dynamic Hypertext System
Kalina Bontcheva
University of Sheffield,
Regent Court, 211 Portobello St., Sheffield S1 4DP, UK
[email protected]
Abstract. This paper presents the evaluation of a dynamic hypertext

system that generates personalised explanations. The main focus is on
some of the results, which showed that future experiments need to control
for users’ reading behaviour, as well as other factors
1 Introduction
hylite+ [2,3] is a dynamic hypertext system1 that generates encyclopaedia-
style explanations of terms in two specialised domains: chemistry and computers.
The user interacts with the system in a Web browser by specifying a term she
wants to look up. The system generates a hypertext explanation of the term;
further information can be obtained by following hypertext links or specifying
another query (see Fig. 1). The system is based on applied nlg techniques, a re-
usable user modelling component (ViewGen), and a flexible architecture with
module feedback. The adaptivity is implemented on the basis of a user and a
discourse models which are used to determine, for example, which concepts are
unknown, so clarifying information can be included for them. The user model is
updated dynamically, based on the user’s interaction with the system. When a
user registers with the system for the first time, her model is initialised from a
set of stereotypes. The system determines which stereotypes apply on the basis
of information provided by the user herself. If no such information is provided,
the system assumes a novice user.
In this paper we discuss the results of the system’s task-based evaluation.
In a nutshell, the participants found the system easy to use and the hypertext
understandable and intuitive to navigate. The generation techniques were suf-
ficiently fast, so users did not have to wait for the system response. Similar to
other evaluated adaptive hypertext systems (e.g., [5,6]), the majority of the users
expressed a preference for the adaptive version.
More importantly, the evaluation results showed that future experiments
which attempt to measure quantitatively the benefit or otherwise of the adap-
tivity techniques, need to control for users’ reading behaviour, as well as other
1
In dynamic hypertext page content and links are created on demand and are often
adapted to the user and the previous interaction.

70 K. Bontcheva
factors like users’ domain knowledge and experience with the application. In
addition, the users’ feedback showed that preference for adaptivity can be im-
proved, if users are allowed some control over the system’s decisions, e.g., if they
can disable some features they are not happy with. In other words, as argued
previously in [8,2], adaptability and adaptivity can benefit each other.
The paper is structured as follows. Section 2 describes the goals and the
methodology of this evaluation. The results are discussed at length in Section 3.
The paper concludes by summarising the results and outlining some future work.
2 HYLITE+ Evaluation Setup

The acceptability and utility of the adaptive features were evaluated by users
who interacted with two versions of our system: a baseline one and the adaptive
one. The two versions were, in fact, the same system with the user model and
adaptivity features switched off to form the non-adaptive baseline. In this way,
we ensured that the same information was available in both systems. Also, this
approach minimises the potential influence of different systems’ response times
on the experiment results, because both versions generate hypertext on the fly
and have similar performance.2
The initial system page, which was identical in the two versions, contained
an alphabetical list of topics, relevant to the experimental tasks (see Fig. 1 left).
The users requested an explanation on a given topic by selecting it and clicking
Explain. After reading the generated page, the participants could obtain further
information by following links, navigating back and forward with the browser
buttons, or by selecting a new term in the topic list (see Fig. 1 right).
The explanations generated by the adaptive version were enriched with par-
enthetical definitions of unknown terms, examples, and links to other more gen-
eral, specific, and related terms (see Fig. 1 right).
2.1 Main Goals

Hypermedia applications are often evaluated with respect to: interface look and
feel, representation of the information structure, and application-specific infor-
mation [9]. The information structure is concerned with the hypertext network
(nodes and links) and navigation aids (e.g., site maps, links to related material,
index). The application-specific information concerns the hypermedia content –
text, images, etc. For our system there is no need to evaluate the interface, since
hylite+ generates simple HTML and uses Web browsers as rendering tools.
Therefore, the evaluation efforts were concentrated on the information content
and navigational structure of the generated hypertext. In addition, observation,
questionnaires and informal interviews were used to elicit problems and assess
the acceptability of the implemented adaptivity techniques.
2
ilex, for example, used pre-generated static pages as a baseline and the study re-
ported that the difference in the two systems’ response times might have influenced
some of the results [4].
Adaptivity, Adaptability, and Reading Behaviour 71
Fig. 1. The system home page (left) and a generated explanation by the adaptive
system (right)
For the purpose of this experiment, information content was compared on

the basis of:
– average time to complete each task;
– average number of pages visited per task;
– average number of distinct pages visited per task;
– percent of correctly answered questions per task;
– questionnaire results about content and comprehension of the generated
pages;
– user preference for any of the systems.
The navigational structure was measured by the following metrics:
– average time per page visited;
– average number of pages visited;
– total number of pages visited;
– number of links followed;
72 K. Bontcheva
– usage of the browser Back button;

– usage of the system’s topic list to find information;
– observation and subjective opinion on orientation;
– subjective opinion on navigation and ease of finding information.
2.2 Methodology
The experiment has a repeated measures, task-based design (also called within-
subjects design), i.e., the same users interacted with the two versions of the
system, in order to complete a given set of tasks. Prior to the experiment, the
participants were asked to provide some background information (e.g., computing
experience, familiarity with Web browsers, and electronic encyclopaedia) and fill
in a multiple choice pre-test, that diagnosed their domain knowledge.
After completing all the tasks, the participants were asked to fill in a question-
naire and participate in a semi-structured interview discussing their experience
with the two systems.
2.3 Subjects
The tasks were completed by eight participants – three male and five female.
They were assigned randomly to the system they used first: four started with
the adaptive and four - with the non-adaptive system. After completing the
first three tasks, the users swapped systems for the other three tasks. None of
the users had any prior experience with the system, but they all had extensive
experience with Web browsing and hypertext. All users had previously used on-
line dictionaries and encyclopaedia, which are the types of systems most similar
to ours. The participants also had similar computing background, i.e., they all
had computer science degrees and were either postgraduate students or research
staff at the Computer Science department.
With respect to knowledge of the computer hardware concepts, used in the
evaluation tasks, the pre-test differentiated two groups of participants: complete
novices, who scored between 2 and 4 out of the maximum 15 points on the test;
and users with basic knowledge, who scored 7 to 8. The two types of users were
split equally between the two groups: adaptive-first and non-adaptive first.
2.4 Tasks
Each of the participants was first given a set of three tasks – each set contained
one browsing, one problem-solving, and one information location task3 . The
order was not randomised, because the browsing task was also intended as a
task that would allow users to familiarise themselves with the system and the
available information. The participants performed the first set of tasks with the
non-adaptive/adaptive system and then swapped systems for the second set.
3
The design of the tasks follows the design used in the evaluation of two other adaptive
hypermedia applications – PUSH [6] and [9].
The participants were not given a specific time limit for performing these
tasks because we wanted to provide an environment in which they would interact
as they normally would with online reference systems.
2.5 User Recording and Logging

The system’s user interaction logs provide a detailed record of: the user query
in the list of topics, the hypertext generated, links followed, state of the user
model before and after the explanation, and time stamps for each page. This
information was used to derive the quantitative statistics needed for measuring
the information content and the navigational structure of the hypertext.
In addition, all browser actions (e.g., use of the Back button) and viewed
pages were logged using Astra Quick Test – a tool for testing Web applications
with extensive recording facilities. These logs were needed, first, as a backup
in case some data got lost in the system logs, and second, in order to derive
statistics of the usage of the browser’s Back, Forward and History buttons.4
3 Results
Due to the small sample size (8 subjects) and the differences in users’ prior
domain knowledge and browsing styles, the results obtained could not be used
to derive a statistically reliable comparison between the measures obtained for
the adaptive and the non-adaptive versions, but the quantitative results and user
feedback are sufficiently encouraging to suggest that hylite+ adaptivity is of
benefit to the user.
3.1 Importance of Users’ Reading Behaviour: Skimmers vs Readers

As discussed in Sect.2.1, the information content provided in the two system
versions was evaluated using quantitative measures, derived from the user inter-
action logs.
The mean time per task results showed that the users who completed the
first three tasks with the non-adaptive system took longer on average, with the
biggest differences for the browsing and information location tasks. Since the
participants swapped systems for the next three tasks and were already familiar
with the hyperspace, the average times with both the adaptive and non-adaptive
versions are naturally lower.
Unfortunately the results for this particular element of the evaluation cannot
be taken as an indication that, for example, the adaptive system might be more
helpful when the user is not familiar with the hyperspace, because we discovered
4
Their use cannot be logged by our application, since these actions do not generate
requests to it; the browser uses a cached copy instead. Other adaptive systems, e.g.,
ilex [4] used and logged their own navigation buttons. However, users expect to be
able to use the browser’s default buttons [7] and such non-conventional behaviour
might influence their attitude to the system.
74 K. Bontcheva
post-hoc that the experiment failed to control for an important user charac-
teristic which influenced the task completion times. The problem is that even
when asked to locate particular information, some users still read thoroughly the
hypertext pages, instead of just skimming them to find only the facts relevant
to the task.5 Although we were aware that previous studies of people browsing
hypertext (e.g.,[7]) have differentiated two types, skimmers and readers, we did
not expect that this distinction would be important when the users perform in-
formation location tasks. Unfortunately, not only does the difference appear to
remain, regardless of the task, but it also happened that when the participants
were randomly assigned to the systems, all three ‘readers’ were assigned to the
same systems, i.e., the non-adaptive for the first three tasks, then the adaptive
one for the remaining three. This explains why the results obtained for the two
groups are so different.
Consequently, the results for mean time per task need to be interpreted by
taking into account this group effect. As can be seen in Fig. 2,6 both groups
were faster when they interacted with the second system (adaptive for the first
group and non-adaptive for the second group), because they were familiar with
the hyperspace. The group effect also explains why the mean times per task
are consistently higher for the non-adaptive system in the first three tasks and
higher for the adaptive system in the second set of three tasks. The readers
group (non-adaptive first) was always much slower than the other group, which
consisted only of skimmers.
This group effect will become apparent again in other measures discussed
below. Its impact on the evaluation results show that it is very important to
control for the differences in user characteristics and behaviour, which influence
the subjects’ task performance much more than the two different experimen-
tal conditions, i.e., adaptive versus non-adaptive system. In order to avoid this
problem in future experiments, we intend to control for users’ reading behaviour
prior to the experiment, just as we did for domain knowledge.
3.2 Other Information Content Measures

On average subjects visited more pages per task in the non-adaptive, than in the
adaptive system. In comparison, the average number of distinct pages visited
per task is about the same for the two versions, i.e., in the non-adaptive version
users tend to have more visits to the same pages.
We also computed the average user score per information location task7
for the two system versions. On average the subjects had higher task success rate
5
This readers/non-readers problem was discovered by comparing the mean time each
participant spent per visited hypertext page.
6
The rows of the table show the mean and standard deviation for each of the six tasks
and the total. The rows called ’Browse...’ are for the browsing tasks, ’Choose...’ are
the problem-solving tasks, and ’Locate...’ are for the information location tasks.
7
The experiment did not measure task score for the browsing tasks, because its role
was to allow the participants to familiarise themselves with the system. Also, due to
the unconstrained, goal-free nature of the task, task success would have been difficult
Fig. 2. Mean and standard deviation of task completion time (in seconds)
with the adaptive system. The novice users exhibited a bigger difference between
the task scores with the two systems: an average 86.8% correct with the non-
adaptive and 94.3% correct with the adaptive system, while those in the medium
group performed equally well with both systems, with an average of 95.8%.
Here again the results might have been also influenced by the readers/skimmers
group effect, because the skimmers (adaptive first) group did less well on the
location task than the readers (non-adaptive first) group. In a larger-scale future
experiment it will also be worthwhile checking whether there is a correlation
between task success and time on task, because the skimmers group was always
faster than the other one which might account for their lower location task
success.
Finally, the questionnaire showed that the users did not have problems lo-
cating information, neither did they find the texts too long. The results also
showed that most users (75%) found the additional information in the adaptive
system useful, while the rest were neutral. Overall, the participants did not have
problems understanding the generated explanations.
3.3 Navigation Measures
Apart from the statistics of average time per page visited and mean number of
pages visited which were discussed above, the other measures related to navi-
gation are the use of links, navigation browser buttons and the topics list (see
Fig. 1). The statistics of the use of these navigation aids in the non-adaptive and
adaptive versions are shown in Fig. 3.
to measure. The score for the other tasks was computed as percentage of correctly
answered questions.
76 K. Bontcheva
Fig. 3. Statistics of the use of navigation features by task
In both systems, the subjects used links as their primary way to navigate the
hyperspace. The browser Back button and the topics list, however, were used
much more frequently in the non-adaptive system, e.g., the topic list was used,
on average, twice as often. The topic list was used mostly in the two ‘locate
mouse details’ tasks to access directly information about particular types of
mouse (e.g. wheel, optical) – see Fig. 3.
The difference in the use of the browser Back button between the two versions,
seems to be due mostly to the different ways in which the participants navigated
the hyperspace. In the non-adaptive version, they would often start from a topic,
e.g., tape drive, then follow a link to the first subtype, e.g., DLT, then return
to the previous page, explore the next subtype, etc. In other words, the most
common navigation pattern was similar to a depth-first traversal of the domain
taxonomy. On the other hand, the additional information and links provided by
the adaptive system changed the subjects’ navigation pattern. The interaction
logs showed an increased use of ‘horizontal’ taxonomic links, i.e., users would
often jump from one subtype straight to the next, without going back to the
‘parent’ node. This change of behaviour was observed for more than half of the
participants and accounts for the reduced number of repeated visits to the same
page, and hence the reduced number of pages visited per topic in the adaptive
system.
If we look into those results in more detail, we see again the difference between
the two groups: readers (non-adaptive first) and skimmers (adaptive first). The
use of the button by the readers group changes substantially when they move
to the adaptive system (e.g., from 3.25 for task 1 to 1.25 for task 4). Unlike
them, the behaviour of the skimmers group hardly changes between the two
systems (e.g., 5.25 for task 2 and 5.00 for task 5). Here again we suspect that
the difference might be due to the readers/skimmers effect, rather than due to the
different treatment they received, i.e., which system they interacted with first.
The reason for this is that possibly the people who read the pages more carefully
could remember better which of the other terms present on the previous page
were worth exploring and jump straight to them, i.e., benefit from the extra
links. Unlike them, the skimmers needed to go back to the previous page to
remind themselves of the other terms, so they could choose where they want to
go next. This conjecture is also supported by the difference in the total number
of pages visited by the two types of users.
The questionnaire results showed that none of the users felt disoriented in
any of the systems and the majority had no problems finding information. When
deciding which links to follow in the non-adaptive system, some of the novice
users reported problems with unfamiliar terminology (25% of all users). In addi-
tion, half of the participants responded that there were not enough links in the
non-adaptive pages. 37.5% of the users also felt that they had to visit too many
pages in the non-adaptive system in order to find the information needed. For
the adaptive system this number was down to just one and all the other users
disagreed with that statement. The majority of the users also felt that the extra
information and links provided in the adaptive version were making it easier for
them to choose which link to follow next.
3.4 Subjective Satisfaction
The remaining statements in the questionnaire and the topics in the interview
were aimed at collecting feedback on user satisfaction and impression of the
two systems, as well as collecting opinions on ways of improving the generated
hypertext.
All participants found the system intuitive to use and had no problems in-
teracting with it without any prior experience. They were also satisfied with the
system response time, both for the adaptive and non-adaptive versions. The ma-
jority of users were also positive about their experience with the adaptive system
and did not find it confusing. 75% agreed that working with the adaptive system
was more enjoyable than with the non-adaptive one, while only 12.5% disagreed.
The evaluation also showed that the preference for the adaptive system could
be improved if its interface provided users with a way of changing the default
system behaviour. For example, only one of the users did not like the links to
related information, included at the bottom of the page, while she liked the rest
of the adaptivity features.
4 Conclusion
In this paper we discussed the evaluation of the dynamic hypertext generation

system hylite+ and the lessons learned from it. Probably the most important
outcome of this small-scale evaluation was that it showed the need to control not
just for user’s prior knowledge (e.g., novice, advanced), but also for hypertext
reading style. Although previous studies of people browsing hypertext (e.g., [7])
have distinguished two types: skimmers and readers, in this experiment we did
not control for that, because the tasks were concerned with locating information,
not browsing. Still, the results obtained showed the need to control for this
78 K. Bontcheva
variable, regardless of the task type, because reading style influences some of
the quantitative measures (e.g., task performance, mean time per task, number
of visited pages, use of browser navigation buttons).
The other important outcome from this evaluation was that it showed that
users’ acceptance of adaptive hypertext systems could be improved if they are
also made adaptable, i.e., allow users to control their behaviour, so unwanted
features could be disabled (for further details see [1]). Since our earlier empirical
studies [2] had already indicated that adaptability could be beneficial to the
users, hylite+ was designed to allow user control of the personalisation. Our
next step will be to evaluate an adaptable adaptive version of the system against
the other two versions discussed here.
References
1. Kalina Bontcheva. Generating Adaptive Hypertext Explanations with a Nested
Agent Model. PhD thesis, University of Sheffield, 2001.
2. Kalina Bontcheva. The impact of empirical studies on the design of an adaptive
hypertext generation system. In S. Reich, M. Tzagarakis, and P. De Bra, editors,
Hypermedia: Openness, Structural Awareness, and Adaptivity, volume 2266 of Lec-
ture Notes in Artificial Intelligence. Springer Verlag, Berling Heidelberg, 2001.
3. Kalina Bontcheva. Tailoring the content of dynamically generated explanations.
In M. Bauer, P. Gmytrasiewicz, and J. Vassileva, editors, User Modelling 2001,
volume 2109 of Lecture Notes in Artificial Intelligence. Springer Verlag, Berling
Heidelberg, 2001.
4. Richard Cox, Mick O’Donnell, and Jon Oberlander. Dynamic versus static hy-
permedia in museum education: an evaluation of ILEX, the intelligent labelling
explorer. In S.P. Lajoie and M. Vivet, editors, Artificial Intelligence in Educa-
tion: Open Learning Environment: New Computational Technologies to Support
Learning, Exploration and Collaboration, pages 181–188. IOS Press, Amsterdam,
1999.
5. John Eklund and Peter Brusilovsky. The value of adaptivity in hypermedia learn-
ing environments: A short review of empirical evidence. In Proceedings of the 2nd
Workshop on Adaptive Hypertext and Hypermedia, pages 13 – 20, Pittsburgh, PA,
June 1998. A workshop held in conjunction with the Nineth ACM International Hy-
pertext Conference (Hypertext’98), http://wwwis.win.tue.nl/ah98/Eklund.htm.
6. Kristina Höök. Evaluating the utility and usability of an adaptive hypermedia
system. Knowledge-Based Systems, 10:311—319, 1998.
7. Jakob Nielsen. Designing Web Usability: The Practice of Simplicity. New Riders
Publishing, 2000.
8. Reinhard Opperman. Adaptively supported adaptability. International Journal of
Human-Computer Studies, 40:455 – 472, 1994.
9. G. B. Wills, I. Heath, R.M. Crowder, and W. Hall. User evaluation of an industrial
hypermedia application. Technical report, M99/2, University of Southampton,
1999. http://www.bib.ecs.soton.ac.uk/data/1444/html/html/.
Towards Generic Adaptive Systems:
Analysis of a Case Study
Licia Calvi and Alexandra Cristea
Information Systems Department

Eindhoven, The Netherlands
{a.i.cristea, l.calvi}@tue.nl
http://wwwis.win.tue.nl/~{calvi, alex}
Abstract. This paper analyses one of the most well-known general purpose
adaptive hypermedia systems, AHA!, and, based on this analysis, make some
general observations about adaptive hypermedia systems and some improve-
ment suggestions for the AHA! system. We suggest here a concept-based ap-
proach to the structuring of adaptive hypermedia systems, as well as an exten-
sion of the well-known rule-based overlay method for user-adaptation. This ap-
proach is another step towards flexible generic-purpose adaptive hypermedia.
1 Introduction
Adaptation in hypermedia has been traditionally divided [3] into adaptive navigation
support (link level adaptation) and adaptive presentation (content level adaptation).
However, this division causes problems at a conceptual level, which can lead, from an
authoring point of view, to difficulties in the definition of concept relationships.
Defining which links to show to users and which concept granularity to associate
to those links presupposes a correlation between concepts and links that is not directly
matched onto the simple link visibility function adopted by most adaptive systems.
Learning is indeed always situated, it never occurs in the vacuum. In a situated learn-
ing scenario, concepts interact with one another to the extent that one concept may
assume a meaning on the basis of the context it belongs to, i.e., of the concepts that
surround it. The influence of contextualization on the learning process is more evi-
dent if we think of the adaptive system as a tool to enhance knowledge acquisition.,
In this view, concepts and links need to be intertwined in order to allow authors to
distinguish between the events producing the knowledge (the actions, i.e., add-
ing/deleting links or concept fragments, triggering the acquisition of knowledge, what
Idinopulos [13] calls causally mediated knowledge) and the inferential process at its
basis, i.e., how to construct the “new” evidence the information that is presented to
the user may lead to once clicking on a link (epistemic mediated knowledge [13]).
From the adaptation engine point of view [24], indeed, it should not make any dif-
ference if the adaptation concerns what links to show to the user, or what text to show

80 L. Calvi and A. Cristea
to the user: if the specific prerequisites are satisfied, the respective action (of adding
links, or text/ multimedia content, etc.) is triggered. Actually, some systems (e.g.,
AHA! [8]) do not make this distinction in their adaptation model totally explicit.
We argue however that, for an adaptive hypermedia author, it is difficult to sepa-
rate the two notions (links versus concepts; adaptive navigation versus adaptive pres-
entation) and at the same time to carefully design the whole system so that adaptive
navigation support actions, triggered directly by the adaptive engine, and adaptive
presentation actions, triggered by in-page (or in-concept/ content) rules, are synchro-
nous.
In the following, we propose a better way to look at the whole authoring problem
in adaptive hypermedia. This approach consists of a combination of the concept map-
ping paradigm to construct the course narrative and of several new adaptation rules.
We also show how the two formalisms may be integrated on an example version of
AHA!.
2 The AHA! System
AHA! [1] is a well-known system, one of the pioneers of adaptive hypermedia (with
its first version developed in 1996/97), which became almost a benchmark for the
domain. One of the co-authors was involved in the research and development from
the very beginning [7], while the other is involved in supervising the project towards
new developments since the support received from the NL foundation.
The power and popularity of AHA! lies on the fact that it is very simple. However,
this simplicity can have drawbacks, as shown in Fig. 1: if the system complexity is
low, the authoring efficiency cannot be very high – unless the author puts a great
effort into creative authoring. This is about the point where the AHA! system is now.
Fig. 1. Relation between system complexity and authoring efficiency
In its present form, AHA! offers the following adaptation methods:

1. each page is considered a concept, and is presented or not according to some
conditions (on some variables) present in an XML file called “requirement list”;
Towards Generic Adaptive Systems: Analysis of a Case Study 81
2. the variables changing rules are relatively simple, and are recorded in another
XML file called “generate list”;
3. AHA! uses an XML based AHA! tag language for conditional fragments within
pages.
Items 1-2 permit adaptive navigation support at page level, and item 3 permits adap-
tive presentation. Recently, AHA! was extended with the following authoring tools:
1. an editor to connect requirements to pages;
2. an editor for the generate rules;
3. forms to make changes to the user model. The most important one is a form that
allows the adaptive hypermedia user to modify knowledge attributes associated
to page-concepts.
All these features unfortunately rely on an inadequate definition of knowledge and of
knowledge acquisition because they do not seem to take into account the importance
of context in concept meaning attribution (as discussed in Sect. 3). In particular, the
possibility of altering knowledge attributes in the user model seems dangerous if this
is not coupled with a redefinition of knowledge within the whole system, with the
inevitable consequences that such an action has in determining concept presentation.
AHA! now is moving towards database-base multiple-attribute concepts (while still
trying to keep complete compatibility with the so fashionable XML format). Next, we
will present some suggestions on the conceptual structure and on adaptation tech-
niques, also pointing to possible problems that AHA! will have to face and deal with.
3 The Concept-Mapping Paradigm
A quite intuitive classification is to divide the source material into concepts [6], as
derived from the concept mapping paradigm [18]. In such a structure, each piece has
an independent semantics - in the sense of the semantic Web [21]: starting with low
level, atomic concepts, to collections of concepts (composite concepts), that together
form a concept hierarchy. Concepts can relate at any level of the hierarchy. The crea-
tion of these building bricks is the role of the adaptive hypermedia designer [6].
This hierarchy represents the primitive building blocks of the hypermedia. Putting
these building blocks together with different sequences generates different presenta-
tions at a relatively high granularity level (concept level). At this level, indeed, we
would be only speaking of adaptive navigation support. Normally, adaptive presenta-
tion is at a lower, concepts fractions level. A simple example is the construction of
text introduction. This construct can be used together with other introductory frag-
ments in an introductory chapter, or dropped at later browsing, etc. However, such a
construct has usually no independent meaning. A common solution to this is to divide
concepts into sub-concepts, without sometimes caring about the loss of semantics.
Such sub-concepts cannot be easily further used (in the context of collaborative au-
thoring), because they cannot be semantically annotated, and therefore will not be
significant for searching mechanisms.
A more appropriate solution, introduced in [6], is to sub-divide the concept into its
attributes. These can be a concept name, alternative contents, fragments, etc. By map-
ping the course content on a concept hierarchy, and describing the concepts with a set
of attributes, the adaptation has only to deal with concept-level adaptation and attrib-
ute adaptation. The advantage is that it can all be performed (and viewed) from a
high level, and does not need separate consideration of conditional fragments written
within the text, which are more difficult to re-use by other authors. In this way, the
content authoring and the adaptative engine rules authoring is clearly separated, mak-
ing also automatic checks easier. Adaptation is here only a matter of combining con-
cept attributes into pages (pieces of information that can be show at a time). Naviga-
tion, in this context, is dependent on the presentation format (Fig. 2). (e.g., a hand-
held device with shorter pages (than the regular browser) will display the “next” but-
ton more often within the same lesson(check SMIL [22], for presentation issues).
Such a model is compatible with the RDF [20] standard, where the RDF resources
become concepts, the properties attributes and the literals attribute values. AHA! has
partially implemented a similar structure, the main important difference being that
concepts are at the granulation of pages, and can have a single attribute in the current
implementation. Constructs within pages (such as conditional fragments) are not
concept attributes in AHA!, and cannot be independently used with other concepts or
concept attributes. New under development versions of AHA! consider multiple at-
tributes, and a database structure, that will allow more flexibility – including hope-
fully dropping of the artificial separation of conditional fragments (such as in adap-
tive presentation) and concept linking (such as in adaptive navigation support.
Fig. 2. A unit X is composed of, e.g., parts of 3 concepts (A, B and C) which have some attrib-
utes (attr-k, attr-l, attr-m). The presentation order is represented by directed connections be-
tween concept attributes. The unit is formed of 2 chapters that contain (parts of) the concepts.
The information in a chapter is presented (by a browser) in pages (which may be shorter than a
chapter). “Next” buttons at page level are navigation support of presentation nature, and have
nothing to do with the user-model related adaptation.
Figure 2 shows how connections are done at concept attribute level. Concept at-
tributes can appear more than once, within one or more unit(s), and there is no restric-
tion for attribute contents: text, video, audio, flash technology, etc. Next, we will
explain the benefits of this simplification by presenting new adaptation types based
on it.
4 New Adaptation Rules: How to Augment the Adaptation Engine
Different rules can be conceived to augment the functionality of the adaptation en-
gine. As they trigger when some quality measurement is reached, these rules, al-
though apparently retraceable to standard commands in traditional programming
languages, can be also viewed as deriving from the Genetic Graph modeled by Gold-
stein [14]: both paradigms indeed explain how the user’s knowledge can evolve dur-
ing learning.
Most adaptive systems are rule-based. Adaptation is mainly triggered by condi-
tional rules, which take the form:
IF <PREREQUISITE> THEN <ACTION>
We propose some alternative forms of conditional rules that allow more freedom,
both in authoring as well as in studying with this type of environment.
1. A level rule [6]:
IF ENOUGH (<PREREQUISITES>) THEN <ACTION>
where ENOUGH = function of number and quality of prerequisites; true when, e.g.,
a given number of prerequisites from a PREREQUISITES set is fulfilled. This
type of relaxation of the prerequisites is intuitive, in the sense that it allows the au-
thor to write simplified rules, instead of writing a great number of complex ones;
the idea is derived from game levels.
Example: PREREQUISITES = time_spent; ACTION = “go to next level”
The rule becomes:
IF ENOUGH (time_was spent on current level) THEN “go to next level”
Where ENOUGH is defined, e.g., as follows:
ENOUGH (time) = 30 time units;
time (advanced topic) = 10 (time units per topic);
ENOUGH (medium topic) = 5 (time units per topic);
ENOUGH (beginner topic) = 2 (time units per topic);
2. A temporal rule: a certain action is repeated as long as one or more conditions
hold. To capture unbound minimization, we add therefore the WHILE construct:
WHILE <CONDITION> DO <ACTION>
According to the concept mapping paradigm, concepts are not canned but are as-
sembled depending on the user model based on their attributes (see Sect. 3).
Therefore this operation requires more than the mere addition/deletion of links.
3. A repetition rule: a certain action (simple / composed) is repeated for a number of
times that is predefined by the author. This rule forces the user to reread a concept
(as described in [15]) that is presented to her invariably. In non-educational con-

texts, this rule might have a lyrical effect1 – such as a refrain in a song.
The temporal rule could be expressed as:
FOR <i=1…n> DO <ACTION>
This rule therefore describes the time this action has to last before the reader can
move on to another one.
4. An interruption command: the action the user is performing is interrupted (bro-

ken) and she is forced to undertake a different one abruptly.
BREAK <ACTION>
The adoption of this rule might be due to lyrical reasons. This is the case in A life
set for two [11], where the reader is forced to reach the end of the fiction once she
has read a predefined percentage of it In educational contexts, this rule might be
used by the author to constraint the user’s explorative behavior within predefined
boundaries, i.e., those corresponding to the pedagogical rules implemented in the
system. In this sense, the implementation of this rule represents an exacerbation of
the traditional behavior of AH systems: here, indeed, the user is “punished” if she
does not stick to the learning pathways provided by the system.
5. A generalization command: the new concept the reader has come across is com-
pared with the more general ones it refers to. As a result of this inductive action,
the reader is pointed to the related concepts she may be interested in reading. So,
the reader has selected a node that describes a certain concept in specific, individ-
ual terms. The system “interprets” this behavior as an interest from the reader
about that particular notion and therefore performs an inductive action to point her
to the more general ones it refers to.
GENERALIZE (COND, COND1, …, CONDn)
6. A specialization command: conversely, if the concept is general, the system de-
ductively points the reader to its more specific instantiations.
SPECIALIZE (COND, COND1, …, CONDn)
So, for instance, if a student is reading a page about the concept of the “Model
Reader” in a course about postmodern literature, she could be further pointed to an
extract from Calvino’s novel ‘Se una notte’, where this notion is exemplified, or to
a further theoretical elaboration on the same topic (by, for example, Genette), or,
again, to a description of how this idea is realized in hyperfiction (see [4]).
Other commands that could appear are comparison (concept analogy search) and
difference – both instances of generalization; or duration – a rule related to repetition.
The appropriateness of these rules directly depends on the context and on the con-
cepts to be modeled. Certainly, educational material may be simpler and more
straightforward to model than a less formal one, like, for instance literary material. In
an educational setting, Indeed, the assumption is often to guide the student to evolve
from novice to expert in the content domain [5]: this is why the content structure has
to be well defined and normally hierarchically organized. Literature, however, exem-
plifies a case where not only the reader’s goal is not clear (contrary to the learning
1 The lyrical use of repetitions in hyperfiction has given rise to a particular design pattern as
described in [2].
goal of educational systems2) and, as a consequence, the way readers use information
to construct meaning [12], but, also, it enacts a different approach: not a “guided
pulling” approach [19] like in educational systems, but one based on a “suspension of
disbelieving” [9], on the importance of rereading in constructing meaning [15], on
suspense, on playing with the reader3 .
Some of these principles would however be effective also in educational systems:
we think, for instance, of the notion of rereading (the old say “repetita iuvant”); of
suspension, of disbelieving as a way of addressing the question of authority and au-
thenticity by assigning it to the author of the content; of a sort of suspense in the way
information is presented to readers as to encourage and to motivate them to read fur-
ther. Moreover, the possibility of expressing these sorts of rules is precisely what
guarantees the general-purpose character of the underlying formalism.
5 Implementing New Rules in the Current AHA!
Following the present syntax used by AHA! to express rules, we could extend it to
represent the above mentioned functions in the following way :
1. the level rule: 2. the temporal rule:
<if expr=”enough”> <while expr=“wrong_search>50”>

<enough> <block>
Here definition of enough
</enough> Here a conditional fragment
<block> </block>
Here a conditional fragment </while>
</block>
<block> In this case, a certain action is repeated as
Here an optional alternative frag- long as a certain condition holds. For instance,
ment a warning is repeated that the search per-
</block> formed by the user is in the wrong direction,
while the system is still performing the search.
Another condition then could trigger a service
denial response if the above threshold would
be passed.
2 For an overview of the possible goals and their related reading strategies adopted by readers
in hyperfiction see, for instance, [4].
3 We refer again to the illuminating paper in [12], where they describe one of the first
systems developed to deal with adaptivity in literature.
Another example is:

<while expr=“art>70 and
not culture>80”>
<block>
Here a predefined
sequence of events
</block>
<block>
Here an alternative event (action)
</block>
</while>
3. the repetition rule: 4. the interruption command:
<for expr=”question&sm;5”> <break>

<block> <block>
Here a conditional fragment Here a conditional fragment
</block>
</for> </block>
In a particular context, this action may make </break>

sense. For the example above, some expla-
nation is given as long as the number of
questions is not larger than 5. After that, a
different strategy has to be taken into con-
sideration – such as the suggestion of con-
sulting different material, etc.
5. the generalization command: 6. the specialization command:
<generalize concept=”myconcept”> <specialize concept=”myconcept”>

Here details of generalization (levels, etc.) Here details of specialization (levels, etc.)
</generalize> </specialize>
Example: Example:
<generalize concept= “double_code_theory”> <specialize concept=“Model Reader”>
Here details of generalization (levels, etc.) Here details of specialization (levels, etc.)
</generalize> </specialize>
Such a rule can be used to jump 1 or more The application is similar as the above com-
levels in the hierarchy of concepts. Extra mand, with the difference that the direction of
processing can be done in the body of the the processing in the concept hierarchy is top-
above command, such as giving comments down instead of bottom-up.
on the level to be visited and the reason why.
6 Problems and the Need of Checking Mechanisms
As shown in Fig. 1, by increasing the system complexity, the authoring efficiency

grows for a while, and then drops. The problem with the current AHA! system is that
is it somewhere at the beginning of the slope. Adding to many more features and
flexibilities can increase the authoring efficiency for a while, but it is necessary to
stop before the down-curve. When authors have to deal with complex unit graphs
with many concepts and many concept attributes, it is easy to leave something out by
mistake. AHAM [24] tries to deal with such problems as:
!"termination ( avoiding of loops) and
!"confluence (equivalence of order of rule execution - for rule-based adapta-
tion engines).
Their suggestion is to use activation graphs (from the static analysis of active data-
base rules), therefore constructing the whole graph of possible states that is deter-
mined by the concepts, their links, the attributes, their values (especially, initial val-
ues and possible ranges – to eliminate unnecessary branches and optimize the search
tree with the help of constrains) and the rule sets. If such a graph has no loops, the
system will always terminate. For confluence, a difficult procedure of checking the
possibility of commutation between each rule pair (so their order equivalence) is
proposed. AHA! at the moment ensures termination by allowing only monotonic
increases of the attributes (per concept), which will be more difficult for the following
version with multiple attributes. As for confluence, AHA! doesn’t deal with it at all.
Other problems that can appear are:
!"concepts (or concept fragments) never reached;
!"rules (or other adaptation mechanisms) that generate attributes with out of
range (or domain) values.
The good news is that the added rules that we have proposed don’t require extra
checking mechanisms than the ones studied previously in the AHAM context [24].
Basically, loops existing in the regular rules will also be present (and noticeable) in
level rules, temporal rules or repetition rules. (and vice-versa). Non-equivalent rules
that can be executed at a given time but are not commutable will pose the same prob-
lems on regular rules as on the extended set. The extended commands of generaliza-
tion and specialization can be treated the same as regular links (or rules). The inter-
ruption command can actually help in breaking infinite loops, or signaling problems,
similar to the catch-throw mechanism of exception handling in Java.
The bad news is that such a mechanism can be rather time - and space -
consuming.
A better way of dealing with this problem is by means of various simplifications
and complexity decreasing assumptions. One such simplification could be by means
of a belief revision technique to check inconsistencies among knowledge attributes to
concepts and the consequent knowledge acquisition problem. Belief revision consists
in the introduction of a sort of case-based heuristics that:
1. recalls a previous concept with the same features and its associated attributes;
2. adapts, via some rule-based formalism, the course structure (the narrative) to the
current learning scenario;
3. resolves the emerging inconsistencies to make sure that changes of state are epis-
temologically conservative (so that the resulting narrative is not subverted).
7 Future Directions
With the standardization of the building bricks of adaptive systems (such as LOM
[17], Learner model [16]– IEEE, LTSC for education, RDF [20], etc.) it becomes
more and more feasible to collaborate and share adaptive techniques, technologies but
also, system parts, developed adaptive hypermedia presentations, etc. Adaptive and
adaptable systems are necessary in education, where learners come with different
cultural and knowledge backgrounds, different learning styles, genders, ages, espe-
cially in the new context of life-long learning. Such systems are definitely necessary
in commerce (and are having a tremendous success in e-commerce, even with ex-
tremely simple adaptation techniques – such as the well-known Amazon.com tech-
nique of suggesting “buyers who bought this book also bought…”). But they can
have, as we have shown also some other, surprising applications, such as adaptive
literature and adaptive art.
It is extremely important to find the right balance between system complexity and
authoring efficiency, as shown in Fig. 1. Extending adaptive systems with extra adap-
tation rules can be beneficial, especially if these rules can express situations that were
not possible (or difficult) to express with the given set of tools/ rules. Moreover, this
enhancement makes sense if it does not impose on the checking mechanisms, by
increasing dramatically the types of tests an adaptive hypermedia author has to per-
form in order to verify his/her output.
As a direction we predict to be rewarding for such checking mechanisms we see
the replacement of large state trees with all possible situations to be reached from the
existing rule base (or, generally speaking, adaptation procedure) with a visual, dy-
namical representation of the processes involved. For example, the effect of a new
rule on the rest can be shown on the static (and much smaller) unit link graph, as the
propagation of some colored fluid through the graph, etc.
8 Conclusion
The paper has started with a criticism on the widespread practice to distinguish adap-
tation in hypermedia between an adaptive navigation support and an adaptive presen-
tation. This criticism is based on the claim that, in this way, authors of adaptive
courseware have to artificially separate links from concepts but still to coordinate and
tune them in order to provide an adaptation that is conceptually valid and that con-
tributes to a significant knowledge acquisition.
We have suggested a better way to look at the whole authoring problem in adap-
tive hypermedia. This approach consists in the combination of the concept mapping
paradigm to construct the course narrative and of several new adaptation rules. We
have highlighted a few new rules that should be integrated into an adaptive authoring
shell or toolkit.
We have shown how these two formalisms may be integrated in an example version
of AHA!. Moreover, we have augmented the present rule behavior performed by
AHA! to allow it more adaptivity. We claim that this approach is another step towards
flexible generic-purpose adaptive hypermedia.
References
1. AHA, Adaptive Hypermedia for All; http://aha.win.tue.nl/

2. Bernstein, M.: Patterns of Hypertext. Ninth ACM Conference on Hypertext (1998) 21-29
3. Brusilovsky, P.: Methods and Techniques of Adaptive Hypermedia. User Modeling and
User-Adapted Interaction 6 (1996) 87-129
4. Calvi, L.: ‘Lector in rebus’: the Role of the Reader and the Characteristics of Hyperread-
ing. Tenth ACM Conference on Hypertext ’99. ACM Press (1999) 101-109
5. Calvi, L., De Bra, P.: Using Dynamic Hypertext to Create Multi-Purpose Textbooks. In T.
Müldner and T.C. Reeves (eds.), ED-MEDIA’97, AACE Press (1997), 130-135
6. Cristea, A., Okamoto, T.: MyEnglishTeacher – A WWW System for Academic English
Teaching ICCE 2000, Taipei, Taiwan (2000)
7. De Bra, P., Calvi, L.: AHA! An open Adaptive Hypermedia Architecture. The New Re-
view of Hypermedia and Multimedia (1998) 115-139
8. De Bra, P., Ruiter, J.P.: AHA! Adaptive Hypermedia for All. WebNet (2001) 262-268
9. Eco, U.: Six Walks in the Fictional Woods. Harvard University Press (1994)
10. Goldman, A.: Telerobotic Knowledge: A Reliabilist Approach. In K. Goldberg (ed.). The
Robot in the Garden. Telerobotics and Telepistemology in the Age of the Internet. MIT
Press (2000) 126-142
11. Kendall, R.: A Life Set For Two. Eastgate Systems, Watertown, MA (1996)
12. Kendall, R., Réty, JH.: Toward an Organic Hypertext. Hypertext ’00. (2000) 161-170
13. Idinopulos, M.: Telepistemology, Mediation, and the Design of Transparent Interfaces. In
K. Goldberg (ed.). The Robot in the Garden. Telerobotics and Telepistemology in the Age
of the Internet. MIT Press (2000) 312-329
14. Goldstein, I.: The genetic graph: a representation for the evolution of procedural knowl-
edge. In D. Sleeman and J.S. Brown (eds, Intelligent Tutoring Systems. Academic Press,
1982.
15. Joyce, M.: Nonce Upon Some Times: Rereading Hypertext Fictions. Modern Fiction
Studies 43(3) (1997) 579-597
16. Learner Model. IEEE Learning Technology Standards Committee (LTSC);
http://ltsc.ieee.org/wg2/index.html
17. LOM (Learner Object Metadata). IEEE Learning Technology Standards Committee
(LTSC); http://ltsc.ieee.org/wg12/index.html
18. Novak, J. D.: Clarify with concept maps. The Science Teacher, 58(7), (1991) 45-49
19. Perrin, M.: What Cognitive Science Tells us About the Use of New Technologies. Invited
Talk given at the Conference on Languages for Specific Purposes and Academic Pur-
poses: Integrating Theory into Practice. Dublin, 6-8 March (1998)
20. RDF (Resource description framework) standard: http://www.w3.org/RDF/
21. Semantic Web: http://www.w3.org/2001/sw/
22. SMIL, W3C standard, http://www.w3.org/AudioVideo/
23. W3C Recommendation: Authoring Tool Accessibility Guidelines 1.0 (3 February 2000);
http://www.w3.org/TR/ATAG10/
24. Wu, H., De Bra, P.: Suficient Conditions for Well-behaved Adaptive Hypermedia Sys-
tems. Web Intelligence. Lecture Notes in AI. Vol. 2198. Springer. (2001) 148-152
A Methodology for Developing
Adaptive Educational-Game Environments
Rosa M. Carro1,2, Ana M. Breda1, Gladys Castillo1, and Antonio L. Bajuelos1
1 Departamento de Matematica, Universidad de Aveiro,

3810, Aveiro, Portugal
{rosac,ambreda,leslie,gladys}@mat.ua.pt
2 E. T. S. Informática, Universidad Autónoma de Madrid, Ctra. Colmenar Viejo km. 15,
28049, Madrid, España

[email protected]
Abstract. In this paper we present a methodology for describing adaptive edu-

cational-game environments and a model that supports the environment design
process. These environments combine the advantages of educational games
with those derived from the adaptation. The proposed methodology allows the
specification of educational methods that can be used for the game environment
generation. The educational goals, the activities that the users can perform,
their organization and sequencing, along with the games to be played and the
game stories are selected or dynamically generated taking into account the
user’s features and behaviors.
1 Motivation
Educational computer-based games are those games that promote the growth of peo-
ple’s reasoning and the acquisition of skills and knowledge in a pleasant way [1].
Their background is related to pieces of knowledge that the users have to put in
practice in order to reach the goals proposed in the games. From the first studies
about the use of games in education [2] until now, they have proved that can consti-
tute a very good resource of motivation [3] for the users to test the knowledge they
own, improve it by practicing, and learn what they do not know while enjoying. Par-
ticularly, the use of multimedia resources, stories that present real or figured goals
attractively, and agents that accompany the user during the game execution (motivat-
ing them to go on playing, providing feedback and so on) increases the learning
achievements [4].
From a high-level point of view, there exist basically two different kinds of educa-
tional game environments: those composed by a fixed sequence of sceneries which
the user has to interact with [5], and those who allow the user to select the game (s)he
wants to play among a set of games [6]. In both cases, the whole game environment is
developed ad-hoc. Adaptation exists in the sense that each concrete game’s behavior
depends on user actions, but the decisions about the next scenery or the available

games at every moment are fixed during the game development phase, being the
same for every user at runtime.
There exist some factors that affect the effectiveness of educational games. These
are related to personal user’s features, preferences and behaviors [7]. As it is well
known, not all the users have the same preferences or styles while interacting with
games and solving problems. From the game developer’s point of view, there are
several educational methods and strategies that can be applied during game-
environment creation [8]. They are related to the kind of tasks proposed to the users,
the sequencing among them and so on. Each developer may want to apply different
methods depending on the features of the users the game is intended for. Moreover,
(s)he can consider as necessary the use of different kind of multimedia games and
storylines for each type of user.
In order to broaden the number of potential users and to improve the effectiveness
of the games for each of them, we propose the development of adaptive educational
game environments. In these environments the cognition activities that users have to
perform, the difficulty of the problems behind the game, the sceneries presented and
the organization of these elements, among others, can be dynamically selected or
generated for each particular user depending on his/her personal features and behav-
iors. In such a way the advantages of using computer-based games in education [9]
along with those that come from the adaptation [10, 11] can be exploited together.
In Sect. 2 of this paper we present a methodology that has been created for sup-
porting the design of adaptive educational game environments along with the specifi-
cation of the educational methods to be used in them. In Sect. 3 we propose a model
that supports this methodology, and Sect. 4 shows the adaptation process carried out
during the dynamic game environment generation. Finally, Section 5 contains the
conclusions and future work.
2 The Proposed Methodology for the Environment Design Process
In order to assist the process of creating educational adaptive games we have devel-
oped a methodology that establishes a set of steps for the game environment design
process. In the following subsections this methodology is presented, as well as an
example about its application for developing an adaptive game environment.
2.1 Steps to Follow
The methodology establishes the following steps:

1. Identifying the types of users who the environment is directed to, by fixing the
personal features that will be taken into account during the adaptation process and
their possible values.
2. Specifying the game goals from an educational point of view, namely, the user’s
knowledge or skills needed to play the games or those that can be ac-
quired/improved while playing them.
92 R.M. Carro et al.
3. Creating or providing the computer-based games, indicating, for each of them, the
learning goals involved and the type of users the game is intended for, in case they
are specifically oriented to certain kinds of users. Games can be taken from exist-
ing game repositories, or generated and stored in one of them.
4. Establishing the educational method for each type of user, by:
4.1. Determining the activities that will be proposed to the users. These activities
can be of two different kinds: educational activities or relaxing activities. The
former are those proposed with the purpose of motivating the acquisition of
certain knowledge, that one described by the educational goals; the latter can
be non-related to education, they all have the same goal (‘relaxing’) and can
be included so that the users can relax after performing difficult tasks. For
every activity the developer must specify its type, goals and the number of
games that will be presented to the users to play with while performing the ac-
tivity. Optionally, the name of the concrete game(s) can be specified. Other-
wise, games will be selected from the ones in the game repository, selecting
the games whose goals match with those indicated in the activity.
4.2. Describing how the environment structure will be generated for each user at
runtime. This requires the specification of i) the organization of activities for
each type of user in activity groups, where several activities can be gathered ii)
the sequencing mode for these activities execution and iii) the prerequisites
that can be established among them.
5. Describing stories, where the game goals, activity feedbacks, agents and other
multimedia elements can be included. These components are associated to the
game structure components (activities and activity groups) and constitute alto-
gether a story that can be independent from the concrete games presented to the
users. In this case, the users can consider games as ways of achieving the story
goals. Different stories can be created for distinct types of users, being included in
a game environment by:
5.1. Creating or providing starting sceneries that are presented to the users before
they carry out an activity or access a group of activities. They present chal-
lenges and goals in an attractive way so as to catch the user's attention and
make them get involved in the environment. The goals presented can be either
related to the activity goals or fictitious, being part of a fantastic story.
5.2. Developing or supplying menu sceneries that will be used for game menu
generation. They can contain objects or agents that stimulates the users to se-
lect one activity among any set of available ones. Menu sceneries can be used
for story construction by presenting fictitious goals that will be achieved if the
user perform the activities, for example, or by motivating the users to go on
playing for reaching the goal (whichever the goal is).
5.3. Creating or supplying feedback sceneries for providing the users with com-
ments about the activity execution, giving them a positive reinforce both when
the results are good, to reward them, and also when they are bad, to motivate
them to learn and go on playing. This comments should be inserted into the
story explaining in an attractive way why the story goals have (not yet) been
achieved.
5.4. Classifying the games according to their subject (i.e., animals, sports, etc.) so
that games related to the same subject can be selected during the environment
generation.
2.2 Ecotoons2: An Adaptive Educational Game Environment
The educational environment Ecotoons2 has been created starting from Ecotoons, an
educational game (developed inside the frame of the Geometrix project) whose main
goal is to promote the children mathematical reasoning development in an attractive
and pleasant way. The game was conceived for children from 5 to 9 years old. Sev-
eral experiences with Ecotoons, involving about 120 children from three different
schools located in Aveiro, have proved that it is a well-designed game with respect to
the interface usability [12], the motivational elements, and so on. In Ecotoons the
stages are presented sequentially to all the users in the same way.
Ecotoons2 is an adaptive educational game environment that incorporates some of
the games and resources of Ecotoons. This environment is intended for users from 5
to 18. By now, there have been established three different ranges: children from 5 to
9, those from 9 to 11, and users from 11 to 14.
Its main goal is to help and motivate users to construct knowledge about mathe-
matical concepts and improve their mathematical reasoning. The educational goals
involved in the game are related to mathematical concepts and operations such as
counting, adding, subtracting, manipulating fractions, and so on. Ninety independent
computer-based games that support the performance of these activities are being
developed. Some of them are distinct games with the same goal, specifically devel-
oped for certain kind of users (i.e., younger users can add by counting and provide the
solution by selecting among several numbers, while older ones should be able to
solve problems by adding numbers and writing the result). All the games developed
until now for the youngest are related to environment conservation while those pro-
vided for teenagers are related to sports and animals, among others. Each game has its
own feedback messages that are shown to the users while they are playing it.
With respect to the educational method, in many cases educators have decided that
mixing activities with different goals or letting the users to choose the order of activi-
ties is better than establishing a fixed sequence. Otherwise, users could get tired or
bored while performing the same task for a long time, or feeling certain lack of free-
dom while interacting with the game. In other cases, the establishment of a fixed
sequence of activities has been considered convenient.
There have been specified thirty activities, which have been organized in groups of
five, resulting in six activity groups. These are grouped into three activity groups that
compose the main one. The corresponding relations among them have been estab-
lished in order to describe the game structure.
Stories are included by means of different starting and feedback sceneries that
have been attached to each activity group. We have used the story of Ecotoons for
sceneries generation. In the main activity group the starting scenery contains an agent
that states the main goal of the whole game: “Saving the planet Platoon by avoiding
that the Poluxes destroy it. If your help is positive, Poluxes will be transformed into
Platoons. Otherwise, the planet will be lost forever”. In every starting scenery, an
agent presents the fictitious goal (what Poluxe can be converted, in the case of the
three activity groups, or the planet area to be protected, for the six activity groups)
and the way of achieving it. Feedback sceneries contain animations showing the Po-
luxe transformation or the area saved, along with the Platoon explanation of what is
happening, depending on the score obtained while performing each activity. The
menu sceneries are related to Platoon landscapes and agents.
3 The Adaptive Game Environment Specification Model
In order to support the previously described methodology we have developed a model

for describing adaptive educational environments. These environments are generated
on the fly. The main components of the model are related to the types of users, educa-
tional goals, computer-based games, educational methods and stories.
- The types of users the game is intended for are described by a series of attributes.
The possible values for each attribute can be described by specifying the ranges
of values or the set of possible discrete values (see Table 1).
Table 1. Establishing some user features and their possible values
Attribute Values
Age [5, 9) [9, 11) [11, 14]
Language Portuguese/Spanish/English
Preferred-media Sounds/Texts/Images/Animations
- An educational goal is specified by its description (a word or a sequence of

words). Example: ‘adding numbers’.
- A computer-based game is included by specifying its identifier, description,
goals, location, icon (for menu generation), difficulty (‘H’ for high, ‘M’ for me-
dium, ‘L’ for low and ‘P’ to indicate that the game can be generated from a tem-
plate with parameters whose values can be set), features (list of descriptive game
attributes such as the age of the users it is intended for, the predominant media,
the language used in texts, sounds, etc.), and, optionally, subject (i.e., ‘sports’,
‘animals’, ‘races’, and so on). An example of a template-based game that is ori-
ented to portuguese children from 5 to 9 can be seen in Table 2. When generating
the game, the number of fishes that appear in each fishing rod is randomly se-
lected from the range of possible values.
In order to describe the educational method of an adaptive game environment the
following components are used:
- Activity: it is the basic unit of the game structure and represents a task to be per-
formed. It is described by its name, type (‘educational’/ ‘relaxing’), goals, and,
optionally, identifiers of starting sceneries, menu sceneries, and/or feedback
sceneries (the way of describing sceneries is explained below). The number of
games that should be played while performing the activity can also be specified
(otherwise one game will be played). Table 3 shows an example of an activity
with no sceneries associated. For the activity performance, one game whose goal
is ‘add’ will be selected for each particular player.
Table 2. Game example: adding fishes
Attribute Value
Identifier Adding-fishes
Description The user has to count the number of fishes that
appear in three fishing rods, write the three num-
bers, and write the total number of fishes
Goals Add
Location /games
Icon Fishing.gif
Difficulty P
Features [5-9), Portuguese
Subject Ecology
Table 3. Activity example: adding numbers
Attribute Value
Name Adding-numbers
Type Educational
Goals Add
- Activity Group (AG): it groups several activities or activity groups. It is described

by its name. It may have associated starting sceneries, menu sceneries, and/or
feedback sceneries that will be used for the story generation. This sceneries can
be different depending on the type of user, so it is possible to associate several
sceneries including the condition for each of them to be selected (which is related
to the user’s features). One of the AGs constitutes the main AG of the game envi-
ronment. An example of an AG with associated sceneries (the same for every
kind of user) is shown in Table 4.
Table 4. Activity Group example
Attribute Value
Name Oceanus-Place
Starting scenery Saving-Oceanus
Menu scenery Ocean
Feedback-OK scenery Feed-OK-Oceanus
Feedback-Wrong scenery Feed-Wr-Oceanus
- Decomposition Rule (DR): it describes which activities or activity groups are part
of a given activity group, and the order they should be performed at runtime, if
any. It is possible to define several DRs for the same AG, showing different ways
of decomposing it and/or different orders in which activities must be carried out.
In this case, the DRs must include activation conditions that have to be satisfied
for the DR activation. These conditions can be related to the user’s features
and/or behaviors while interacting with the environment. In such a way it is pos-
sible to specify different educational methods by providing each user with differ-
ent activities for the same AG. It is also possible to combine the same activities in
different ways for every type of user. Moreover, the use of conditions allows the
decomposition of the same AG in different ways depending on the results ob-
tained by the user while performing other activities. In order to describe a DR it
is necessary to indicate the AG identifier, the AGs and/or activities identifiers, the
sequencing mode among them (according to the sequencing modes proposed in
[13], ‘AND’ means that the activities must be performed in the order they appear
in the DR and ‘ANY’ means that they can be performed in any order) and, op-
tionally, the activation conditions, related to user’s features and/or behaviors
(other activities execution). Optionally, it is possible to associate a weight to each
subactivity, indicating the contribution of the score obtained by the user while
performing the subactivity to the calculation of the score obtained in the activity
group (by default the score will be calculated as the media of subactivity scores).
Table 5 shows an example of two strategies for the same AG: while the order of
activities for the youngest children will be fixed, children from 9 to 11 will be
able to select, a priori, the activity they want to perform at every moment among
those that appear in the DR. In addition, older children will have available the
‘Subtracting-numbers’ activity, which will not appear in the environment gener-
ated for younger ones.
Table 5. Decomposition Rule examples
Attribut Value
Activity group Oceanus-Place
Subactivities Identifying-objects, Adding-numbers, Associating-objects,
Ordering-numbers
Sequencing AND
Activation condition [5, 9)
Attribute Value
Activity group Oceanus-Place
Subactivities Identifying-objects, Adding-numbers, Associating-objects,
Ordering-numbers, Subtracting-numbers
Sequencing ANY
- Prerequisite rule (PR): it describes the dependence relation that can exist among
activities. By means of a PR it is possible to state that some activities should not
be performed before doing other ones previously. A PR is composed by the ac-
tivity whose availability depends on the performance of other activities, the pre-
requisite activities, those that have to be performed before that one, and the acti-
vation condition, that indicates for which type of users the rule will be active, that
is, in which cases the dependence relation will be taken into account. Optionally,
it is possible to associate a weight to each prerequisite activity, indicating the
minimum score that must be obtained during the activity execution for consider-
ing the prerequisite condition satisfied (by default it will be 5 in a scale from 0 to
10). Table 6 shows an example of a PR defined for children from 9 to 11 that es-
tablishes that they will not be able to perform the ‘Subtracting-numbers’ activity
unless they have performed the ‘Adding-numbers’ one.
Table 6. Prerequisite Rule example
Attribute Value
Activity Subtracting-numbers
Prerequisites Adding-numbers
As for the story generation, sceneries are included in the model. These sceneries
can be used for presenting the story, stating the fictitious or real goals at every mo-
ment, generating menus, providing feedback, etc.
- A scenery is described by its name, type (‘starting’/‘menu’/‘feedbackOK’/ ‘feed-
backWrong’), description, location and the kind of users it is created for, that is
directly related to the scenery features. If it is intended for every kind of users
this attribute will be empty.
4 The Adaptive Environment Generation
Once the description of an adaptive educational game environment has been pro-
vided, it is possible to generate the components that will be presented to each particu-
lar user while interacting with it. The adaptation mechanism operates in two stages: at
the beginning of the environment generation and at runtime.
4.1 First Stage: Structure Generation and Story Attachment
The environment structure is generated starting from the main activity group, consid-
ering the decomposition rules and user features in order to determine the activity
groups or activities in which the main one is decomposed for this particular user. This
process is repeated recursively until the game structure is generated. For every node
of the structure (activities and groups of activities) it is checked whether it has any
scenery associated. If it is the case, the scenery is attached to the activity. In case
there are several sceneries available for the same node, the most suitable one for the
user is selected.
In this stage the basic environment structure along with the story in front of it have
been generated. This does not mean that all the activities are available. The activity
availability will be decided for every user at runtime.
4.2 Second Stage: Selection of Available Activities and Games
While a user is interacting with the environment, the available activities at every
moment are selected on the fly. This selection is done by taking into account the
user’s features and behaviors, along with the decomposition and prerequisite rules.
For each set of available activities, a menu is generated so that the user can select the
activity to be carried out. For each activity execution, the most suitable game support-
ing it is selected, considering the available games whose goals match the activity
goals, their features and the user’s characteristics. Once the user has selected an activ-
ity or the system has decided which one will be the next, the computer-based game
will be presented/generated: if the game can be constructed starting from a template,
it is dynamically generated; otherwise, the game is directly presented to the user.
The availability of activities and the difficulty of the games presented will depend
on the user’s actions and behaviors while interacting with the game environment.
In this paper we have presented a methodology that establishes the suggested steps for
describing adaptive educational game environments. These environments combine
the advantages of educational games [9] with those derived from adaptation [10][11].
We have also presented a model that supports the adaptive environment design.
The environments described following this approach are dynamically generated
taking into account the particular user’s features and behaviors for the personalization
of: i) the activities that are part of the environment, ii) their organization in the envi-
ronment structure, ii) the order in which activities should be performed, if any, or the
free activity selection, iii) the computer-based games that will be presented for activ-
ity performance, including their subject and difficulty, and iv) the displayed storyline.
The use of activities, activity groups and rules for the environment description allows the
specification of educational methods. In such kind of environments, the users can acquire skills
and improve their own knowledge in a personalized way. The educators can include the most
appropriate educational methods, activities and games for each kind of user.
The separation among the activities to be performed and the games that support them makes
it possible to select the most suitable games for every user while performing the activity. We
have created a game repository with games labeled with their educational goals and difficulty.
This makes possible the game reuse in different educational environments. Furthermore, exist-
ing games stored in game repositories can also be tagged and used in the environments gener-
ated.
By providing different sceneries for distinct types of users, it is possible to create
several stories for the same activities, widening the possibility of attracting users'
interest in playing the game. Sceneries can compose a fantastic story that can be used
in different environments independently of the activities and the games behind them.
Currently we are experimenting with the use of the proposed methodology and
model for different game environment generation. We are also considering the adap-
tation of the kind of game selected at every moment (educational vs. relaxing games),
provided that some user features such as the age, the difficulty of the last activity
performed and/or the score obtained by the user while performing it may affect the
type of activity that should be proposed next. In the future we plan to create a model
for the internal game components and behavior representation.
References
1. Klawe, M., Phillips, E.: A classroom study: Electronic Games Engage Children as Re-
searchers. Proceedings of CSCL’95 Conference. Bloomington, Indiana (1995) 209-213
2. Gordon, A.: Games for Growth. Sc.Research Associates, Inc., Palo Alto, California (1970)
3. Malone, T. W.: What makes things fun to learn? A study of intrinsically motivating com-
puter games. Cognitive and Instructional Sciences Series, CIS-7, SSL-80-11, Palo Alto
Research Center, Palo Alto (1980)
4. Klawe, M.: Computer Games, Education And Interfaces: The E-GEMS Project. Invited
Presentation at Graphics Interface 1999, Online Papers (1999). Available at
http://www.graphicsinterface.org/proceedings/1999/20/
5. Breda, A.M., Bajuelos A.L., Castillo G, Lopes, M.: Computational Math Games Versus
Learning. Proceedings of the International Conference on New Technologies in Science
Education, Aveiro, Portugal (2001)
6. Hungry Frog Java Arcade Games. At http://www.hungryfrog.com/java/javamath.htm
7. McGrenere, J.: Design: Educational Electronic Multi-Player Games. A Literature Review.
Thesis from the Department of Computer Science, Univ. British Columbia, USA (1996)
8. Gonzalez, C.S., Moreno, L., Aguilar, R.M., Estévez, J.I.: Towards the Efficient Commu-
nication of Knowledge in an Adaptive Multimedia Interface. Proceedings de Interactive
Learning Environments for Children, Athens, Greece (2000). Available at
http://ui4all.ics.forth.gr/i3SD2000/proceedings.html
9. Papert, S.: The Children’s Machine, BasicBooks, New York, NY (1993).
10. Brusilovsky P., Kobsa A., Vassileva J. (eds.) Adaptive Hypertext and Hypermedia.
Dordrecht: Kluwer Academic Publishers (1998) 1-43
11. Carro, R.M., Pulido, E., Rodríguez, P.: Dynamic Generation of Adaptive Internet-Based
Courses. Journal of Network and Computer Applications. Academic Press. Available
online at http://www.idealibrary.com. Vol. 22 (1999) 249-257
12. Reynolds, A., Martin, J.V.: Designing an educational computer game: Guidelines that
work. Educational Technology, January (1988) 45-47
13. Carro, R.M., Pulido, E., Rodríguez, P.: TANGOW: a Model for Internet Based Learning.
International Journal of Continuing Engineering Education and Life-Long Learning,
UNESCO. Special Issue on “Internet based learning and the future of education” (2001).
At http://www.inderscience.com/ejournal/c/ijceell/ijceell2001/ijceell2001v11n12.html
Multi-model, Metadata Driven Approach to Adaptive
Hypermedia Services for Personalized eLearning
Owen Conlan1, Vincent Wade1,2, Catherine Bruen2, and Mark Gargan1

1 Knowledge and Data Engineering Group, Trinity College, Dublin
{oconlan, vwade, garganm}@cs.tcd.ie
http://www.cs.tcd.ie/research_groups/kdeg
2 Centre for Learning Technology, Trinity College, Dublin
[email protected]
http://www.tcd.ie/CLT
Abstract. One of the major obstacles in developing quality eLearning content is

the substantial development costs involved and development time required [12].
Educational providers, such as those in the university sector and corporate
learning, are under increasing pressure to enhance the pedagogical quality and
technical richness of their course offerings while at the same time achieving
improved return on investment. One means of enhancing the educational impact
of eLearning courses, while still optimizing the return on investment, is to
facilitate the personalization and repurposing of learning objects across multiple
related courses. However, eLearning courses typically differ strongly in ethos,
learning goals and pedagogical approach whilst learners, even within the same
course, may have different personal learning goals, motivations, prior
knowledge and learning style preferences. This paper proposes an innovative
multi-model approach to the dynamic composition and delivery of personalized
learning utilizing reusable learning objects. The paper describes an adaptive
metadata driven engine that composes, at runtime, tailored educational
experiences across a single content base. This paper presents the theoretical
models, design and implementation of the adaptive hypermedia educational
service. This service is currently being successfully used for the delivery of
undergraduate degree courses in Trinity College, Dublin as well as being used
as part of a major EU research trial.
1 Introduction
In the past Intelligent Tutoring Systems (ITS) traditionally have embedded experts’
knowledge in the structure of its content and applied appropriate design models.
However, such systems have continually been criticized for believing that this is
sufficient for effective learning to occur [13]. In reality, these early systems
constrained the learner and limited the opportunities for the learner to investigate
topics the ITS deemed to be of little relevance.
Later ITSs used knowledge about the domain, the learner, and about teaching
strategies to support flexible individualized learning and tutoring [4]. One of the goals
of these ITSs was to adaptively deliver content. The majority of such ITSs merge the

Metadata Driven Approach to Adaptive Hypermedia Services for Personalized Learning 101
content, narrative and learner modeling into a single engine, giving a system that
adapts effectively yet is very difficult to repurpose.
Adaptive Hypermedia is a newer research domain [3]. Adaptive Hypermedia
Systems (AHS) apply different forms of learner models to adapt the content and the
links of hypermedia pages to the user [4]. While there tends to be a clearer separation
of the learner model and content model in AHSs (as opposed to the integrated
approach of ITSs) the narrative or pedagogical model in usually either embedded in
the content or into the adaptive engine itself. This means that applying new or
different pedagogical models, e.g. case based learning, simulations, etc., to the
content model is more difficult and involves a re-authoring of the content model. This
results in learning content that is difficult to reuse or an engine that is domain
specific.
This paper proposes an approach that has a clear separation of content, learner
and narrative models, and a generic adaptive engine that employs a multi-tiered AI
model to achieve effective adaptation to the learner’s requirements. The approach is
to have very little semantics actually embedded in the adaptive engine itself.
Therefore the pedagogic semantics that govern the narrative sequence of the learning
experience are contained in a separate model. The adaptive engine reconciles the three
models to compose, at runtime, the personalized course. Such dynamic construction
of the learning experience is controlled by each learner via appropriate pedagogic
instruments, e.g. indirect access to their learner model via an instrument. This
approach enables multiple narrative models to be constructed to fulfill different
learning goals, while these goals may be achieved from a common repository of
content.
Section 2 presents the issues which impact on the development and
representation of content within the Adaptive Hypermedia Service. Section 3
discusses the aspects of the learner which can be modeled to enable adaptation to the
learner’s preferences. Section 4 introduces the narrative model that facilitates the
separation of content and structure and enables the course author to define how the
service adapts to the learner’s knowledge. Section 5 presents the Personalized
Learning Service, an implementation of the multi-model, metadata driven approach to
constructing Adaptive Hypermedia Services. Finally, Section 6 discusses the
conclusions drawn from this research.
2 Content Issues
The main goal of the multi-model approach is to separate the learning content from
the adaptive linking logic or narrative. This separation improves the possibilities of
reusing a piece of learning content as the learning object (LO) is no longer specific to
a given implementation or narrative model. A second goal (with respect to content) of
this approach is to allow course designers to easily discover learning content in the
content repository by providing appropriate descriptive metadata.
Metadata may describe both technical and pedagogical aspects of the LO. This
information is not only useful to a course designer in selecting appropriate learning
content, but can be used by an adaptive engine to select appropriate content where
there may be many candidate LOs available to fulfill a learning or technical
requirement.
102 O. Conlan et al.
2.1 Content Model for the Adaptive Service – The Content Model
The use of standards based metadata to represent the content model encourages the
reuse of the learning objects outside of the adaptive hypermedia service. Content may
also be imported into the services content repository from external repositories if that
content has similar metadata associated with it. The content model utilized in the
adaptive hypermedia service is based on IMS Learning Resource Metadata [9]. The
IMS Metadata specification was chosen as the basis for the content model schema as
it is based on the IEEE LOM specification and an XML binding is available.
Adaptivity is not, however, directly addressed by the IMS Metadata Specification
and in order for the adaptive engine to choose between several candidate pieces of
content (Sect. 2.3) it may be necessary for it to have further information about the
learning objects. Within the EASEL [6] IST project (through which this research has
been part funded) an extension to the IMS Metadata schema was developed. The
extension consisted of the addition of a sub-section called Adaptivity to the Education
section of the schema. This sub-section caters for user definable adaptivity types
allowing the metadata creator to develop complex relationships and dependencies
within the metadata description of the service [5].
From the perspective of a single reusable learning object, the adaptivity types
might include competencies.taught, competencies.required and learningstyle.
Included in each adaptivity type is the ability to reference, using a URI, an external
resource that enables the metadata author to describe that type and any vocabularies
or requirements associated with it.
2.2 Content Lifespan and Granularity
The separation of content from narrative eases the reuse of the learning objects and
potentially increases their lifespan. For example, if many learning objects covering
aspects of the Java programming language were developed three years ago then, if
viewed as a complete course today, may be out of date with respect to the current
version of the language. However, if viewed individually some of the learning objects
(LO) may be reusable, e.g. those describing the control loops. If the narrative was
embedded in the LOs then it may be more difficult to reuse them within another
adaptive course.
The potential reuse of LOs is related to the granularity, or the scope, of the
learning object. The smaller the granularity of the content the greater potential exists
for the LOs reuse. One possible disadvantage of this approach is that if the fine
grained LO, which the author terms pagelets, are poorly sequenced then they may
appear to be inconsistent or incoherent. It is the function of the narrative author to
ensure that the customized courses produced from the narrative contain pagelet
sequences that maintain learning concept coherency and have a logical flow (Sect. 4).
2.3 Candidate Content Groups
The mechanism employed by the narrative to refer to content is to use an indirection,

whereby the narrative doesn’t refer to individual pieces of content (LOs) directly, but
to candidate content groups. Each candidate content group contains learning objects
that fulfill the same content requirement. The LOs in a candidate content group may
differ technically (e.g. bandwidth requirements), in instructional approach or on any
other axes on which the adaptive service may be adaptive. The decision as to which
LO to deliver can be made at runtime based on some information about the learner
(Sect. 3). The requirement for different candidates can be determined by an
educational instructional designer, although the task of generating the content for the
candidates is generally a collaborative process between the domain expert and the
instructional designer.
3 Modeling the Learner
A learner/user model contains explicitly modelled assumptions that represent the

characteristics of the student which are pertinent to the system. The system can
consult the user model to adapt the performance of the system to each student’s
characteristics. User modelling allows the system to personalize the interaction
between the student and the content. To achieve effective learning this personalization
should put the content in a context that the student can understand and to which they
can relate.
3.1 Prior Knowledge and Learning Objectives
The learner model should be capable of storing the prior knowledge and learning
objectives of the learner to facilitate the personalized delivery of content based on the
learner’s experience and goals. This raises a number of questions –
• What vocabulary should be used to describe the prior knowledge and objectives?
• What level of detail does this vocabulary need to describe?
As the narrative model is constructed by an expert(s) in the knowledge domain it is up
to them to use whatever vocabulary they feel best describes the knowledge domain for
which they are building the narratives. As they are responsible for the mechanism(s)
that are used to populate the learner model, e.g. a knowledge pre-test, all they need to
ensure is that the vocabulary is consistent between the learner model and the
narrative.
The granularity to which the vocabulary exists and the scope of the pagelets
determine the level at which the engine can adapt to prior knowledge and learning
objectives. For example, if a course author decides there are ten learning objectives in
a course then the finest grain that the adaptive engine can personalize a course is at
the scale of one of those objectives. This is true even if the pagelets are finer grained,
as each learning objective may require several pagelets to fulfill it. On the other hand
if there was a learning objective associated with each pagelet then the engine could
personalize the course on a pagelet by pagelet basis. There is a balance between the
vocabulary granularity and pagelet granularity that determines the level of content
personalization that is achievable by the adaptive engine.
3.2 Pedagogical Considerations
Learning style is a term used to describe the attitudes and behaviors that determine an
learner’s preferred way of learning. Learning style preferences have implications for
all types of learning, whether the learning is dedicated to the acquisition of knowledge
through formal structured activities, e.g. lectures, case studies and books or through
experiential learning, i.e. learning through experience [8]. For the online paradigm, as
in traditional classroom situations, there is no consensus on how best to model the
learner’s preferred approaches to learning. Therefore, the approach taken in this
design of the Adaptive Hypermedia Service (AHS) was to enable the instructional
designers to impact the rendering of the personalized course at two levels –
• The structure/layout in which the content is placed.
• The type (or format) of content displayed.
To this end the AHS enables many candidate narratives, supporting different
pedagogical approaches to structuring the content, to be associated with a single
course. This association and an appropriate selection mechanism enables the AHS to
deliver a personalized course that, while dealing with the same subject matter, can be
structured in a way that best engages the learner’s preferred learning styles. The
subject matter, however, is not referred to directly in the narratives, but rather through
the mechanism of candidate content groups (Sect. 2.3). At runtime the AHS can
determine which candidate is most suitable (Sect. 5.4). The pedagogical approach
used in the design and selection of the content can be defined by the instructional
designer.
One such approach is the VARK (Visual, Auditory, Read/Write and Kinesthetic)
[15] model that can influence the design of individual content candidates to
emphasize one of these aspects – visual, auditory, read/write and kinesthetic. Using an
appropriate instrument the learner model may be populated with these values. The
AHS can reconcile the content model and the learner model to determine the
appropriate candidate at runtime.
3.3 Describing the Learner – What to Model?
The learner model employs a similar mechanism to the content model enabling an
extensible metadata framework where information pertaining to how the adaptive
engine creates a personalized course may be placed. There is an Adaptivity sub-
section in the metadata model that enables the definition of new adaptivity types. For
example, these types may include competencies.learned, competencies.required and
learningstyle. As this is an extensible framework the domain expert, who describes
the learning content requirements of a course, and the instructional designer, who
describes the pedagogy requirements of the course, can define new adaptivity types to
facilitate other forms of adaptivity they may wish to implement in the narrative.
The learner model also includes learner information such as forename, surname
and a unique identifier (within the adaptive hypermedia service). When the Adaptive
Hypermedia Service is integrating with a Learning Management System (LMS) it is
usual that some of this information is retrieved from the LMS (Sect. 5.6).
4 The Narrative Model
The narrative model for a course describes the rules, developed by domain and
pedagogical experts, which govern the range and scope of personalized courses that
the adaptive engine can produce for learners. The narrative enables the course
author(s) to separate the rules which govern how the personalized course will be
generated from the content that will be included in that course.
4.1 Encapsulation of Domain Expertise
The narrative is a representation of the expert’s knowledge of a domain. Narratives

can be used to generate courses that differ in ethos, learning goals, pedagogical
approach and learner prior experience from a common content repository. The
vocabulary used to describe the learning concepts embodied in the course is that of
the domain expert. As the narrative does not refer directly to individual pagelets, but
rather to candidate content groups using this vocabulary, the domain expert can create
the narrative without being constrained by pedagogical or technical delivery issues at
the content level. The author can simply refer to the Candidate Content Group in the
narrative and allow the adaptive engine determine which candidate from the group is
most suitable for delivery.
Using a similar mechanism to the candidate content groups it is possible to have
several candidate narratives for a single course. The candidates have the same ethos,
learning goals and require the same learner prior knowledge, but differ in pedagogical
approach. Where these kinds of pedagogical issues, i.e. those which impact the course
structure, are implemented the process of developing the narratives is often a
collaborative process between instructional designers and the domain expert. The
Personalized Learning Service is capable of selecting a narrative based on learner
model values (e.g. learning style) or on external factors (e.g. is a revision course
required).
4.2 Coherent Personalized Course Structures
The primary goal of the narrative is to produce courses that are structured coherently
and fulfill the learning goals for the course. It is, therefore, the domain expert’s task to
ensure that each learning goal has sufficient appropriate pagelets to fulfill that goal
and that those pagelets are sequenced in such a manner that engages the learner.
From this perspective the domain expert must consider how the exclusion or
inclusion of pieces or sequences of content will impact on the intelligibility of
neighboring content and on the personalized course as a whole. To this end it is often
useful to determine, before designing a narrative, what is the granularity of content
personalization that is to be achieved, i.e. personalization on the section, page or
paragraph level. It is also useful to determine what content, if any, is considered core
material and should always be present in all personalized courses. With these two
factors set, the expert has a framework in which to consider the impact of the
inclusion or exclusion of content based on the learners’ expertise and preferences.
4.3 Reconciling the Learner Model and the Content Model
As the vocabularies used to describe the knowledge domain and the learner’s prior
knowledge and learning objectives are determined by the expert (Sect. 3.1), that
expert must ensure that those vocabularies are one common vocabulary or that there is
a translation mapping available between the two vocabularies. The adaptive
hypermedia service does not place any restrictions on what constitutes the vocabulary,
only that the narrative model and any elements of the learner model that the expert
wishes to reference share that vocabulary or that there exists an appropriate mapping
between the narrative and learner model vocabularies. This enables the narrative to
reconcile learner prior knowledge and/or learning objectives with candidate content
groups.
This reconciliation enables the narrative to select appropriate candidate content
groups based on the learner’s prior knowledge and learning objectives. The narrative
is not directly concerned with learning style issues; only with the learning concepts
the final course should include to meet the learner’s learning objectives. Learning
style considerations are catered for by the appropriate selection of narrative (Sect. 5.3)
and content (Sect. 5.4).
5 Personalized Learning Service
The Personalized Learning Service (PLS) has been developed as a service to deliver
personalized educational courses based on the multi-model, metadata driven approach
outlined in this paper. The PLS is currently being used to deliver personalized online
courses in SQL (Structured Query Language) to final year undergraduate students in
two degree programmes in Trinity College, Dublin and a trial of eighty students, over
a period of two months, was been carried out to assess the students reaction to using
an online adaptive hypermedia service. It is also being used within the EASEL [6]
IST project to demonstrate the discovery and integration of Adaptive Hypermedia
Services with traditional (static) online learning content.
5.1 Architecture
The architecture of the PLS has the three models – learner, narrative and content – as
the basis of its design. The PLS utilizes three metadata repositories (Learner, Content
and Narrative) and two information repositories (Content and Narrative). There are
also two repositories that contain information about candidates – one dealing with
Candidate Content Groups and one concerning Candidate Narrative Groups.
At the core of the PLS is the Adaptive Engine (AE). The AE uses the Java Expert
System Shell (JESS) [10] with customized functions as the basis of its rules engine.
The rules engines role is to produce a model for a personalized course based on a
narrative and the learner model. The XML-based [16] personalized course model
encapsulates the structure of the learner’s course and contains the candidate content
groups that fulfill the learner’s learning requirements in accordance with the narrative.
Learner
Input Learner Learner
Modeler Metadata
Reposito
ry
Adaptive
Content
Metadata
Engine Reposito
ry
Personaliz Rules Candid
ed Course Engine ate Candidat
Model
Selecto e Content
Groups
Personaliz
ed Course
Content Narrativ Candidat
e Metadata e Narrative
Reposito Groups
ry
Narrative Content
Repository Repository
Fig. 1. Adaptive Hypermedia Service Architecture
The AE also utilizes a candidate selector for choosing the appropriate narrative
by reconciling information in the learner model with the candidate narrative groups.
The candidate selector is also used to choose the appropriate piece of content to
deliver from a candidate content group when the personalized course content is being
generated from the personalized course model (Sect. 5.4).
The AE has a learner modeler component that enables input from the course or
pre-tests to be translated into changes in the learner’s information. This component is
used to populate the learner’s model when the learner initially enters the Adaptive
Hypermedia Service. It can also be used at runtime to modify the learner’s model –
these modifications may either be initiated by the learner or by the engine itself and
can be initiated directly from the JSP.
5.2 Building the Learner Model
Before the PLS can create a personalized course for a learner it must have some
appropriate information about that learner. This information is obtained by asking the
student to complete an online instrument, typically a prior knowledge questionnaire,
that determines both their prior knowledge of the domain and any pertinent learning
style information. This instrument uses the learner modeler component of the AE to
modify the learner model.
The design of any instruments that determine learner information is the
responsibility of the domain expert and the instructional designer. The learner is asked
to interact with the instrument on their first visit to the PLS. The learner can access
the instrument at any stage during their learning and modify their answers. This
process gives the learner an indirect mechanism to change their learner model and
rebuild their personalized course accordingly.
5.3 Selecting the Narrative and Creating the Personalized Course
The first step in creating a personalized course model is to select an appropriate

narrative. For each course there may be several narratives available to achieve the
same learning objectives. These narratives differ in the pedagogical approaches used
to structuring the content they implement. The adaptive engine calls the candidate
selector to choose the best candidate narrative from the candidate narrative group for
the course. The selection is based on the metadata in the Narrative Metadata
Repository and on the learner’s metadata. For example, if the learner’s
preferredcoursestructure is abstract the candidate selector will select the narrative that
matches this preference most closely.
Once the narrative, which is represented by the JESS [10] language, is selected
the rules engine is invoked to interpret the rule set. This rules engine has access to the
learner model to determine which candidate content groups should be added. It can
compare elements of the learner model with rules in the narrative. In the PLS the rules
engine is primarily concerned with reconciling the learner’s prior knowledge and
learning objectives with appropriate candidate content groups.
The rules engine generates an XML representation of the personalized course
model that is described in terms of organizational structures such as sections, modules
or units (as determined by the narrative designer) and candidate content groups.
5.4 Selecting Content and Delivering the Course
Once the Personalized Course Model has been generated it needs to be translated into
a structure and format that the learner can access easily. The appropriate candidates
(chosen from the candidate content groups) need to be included in the delivered
course as well. These steps are performed by passing the XML representation of the
Personalized Course Model through as XSL [17] transformer. The transformer
produces many Java Server Pages (JSP) from the XML representation. The JSPs give
HTML form to the Personalized Course Model. During the transformation the
candidate selector is called to choose the content that best fits the learner’s VARK
[15] preferences. The candidate selector chooses this content by comparing the
prospective candidates metadata with the learner model and uses server-side includes
to include the content in the JSP files. The course is delivered as HTML using the
Jakarta Tomcat [14] JSP engine.
5.5 Initial Learner Trial Results
After completing the personalized SQL course, produced by the PLS, and a related
database design project, the eighty final year undergraduate students who took part in
the trial were asked to complete an evaluation questionnaire.
The evaluation of these results is currently being performed, but the initial
findings have shown that –
• 30% of the students had no prior experience of online learning, while only 6%
said they had much online learning experience.
• Over 80% said the personalized course generated represented the answers they
gave in the online pre-test instrument.
• 60% of the students believed the online instrument gave them sufficient control
over the content contained in the personalized course generated.
• 87% of the students were happy with how the content was structured in the
personalized courses.
These results appear to show the students’ satisfaction with the personalized courses
generated by the PLS, although some of the comments on the evaluation
questionnaires indicated that some students desired a finer level of content control
than that offered via the online instrument.
Also observed was a behavior that was not originally anticipated – some students
used the instrument to regenerate a personalized course for each study period. They
interacted with the instrument is such a way that the personalized course produced
contained only the content they wished to study for that period. This gave the students
greater control over their learning, echoing some of the fundamental concepts of
constructivism.
5.6 PLS as a Remote Third Party Service
In order to be called a service the Adaptive Hypermedia Service (AHS) must facilitate
ease of integration with heterogeneous Learning Environments (LE) in such a way
that learners are able to seamlessly launch and use the AHSs adaptive content from
within their preferred LE, e.g. WebCT, Blackboard, etc. This process, as far as the
learner’s interaction with the LE is concerned, should appear no different from regular
(LE native) content. It may be the case that the learner is studying content from both
native and adaptive sources to achieve a learning objective. Within the EASEL [6]
IST project the Personalized Learning Service has been successfully integrated with
Fretwell-Downing Educations [7] Learning Environment using the Content
Interworking API initially defined by the AICC [2]. The API version and data model
utilized is a JavaScript implementation of the ADL SCORM [1].
The information exchanged between the LE and the PLS is learner identification,
and performance and assessment information. The PLS requests the learner’s
identification from the LE when the service is initially launched and returns
assessment information when the learner has completed their personalized course.
The PLS can be a Third Party Service residing on a separate server from the
learning environment. The advantage of this approach it that there is a clear
separation of responsibilities – the LE provides the learning support facilities and the
PLS provides personalized content. If the learning support features of a Learning
Environment are not required the PLS can be launched as a stand-alone service or
integrated with other learning content.
6 Conclusions
This paper has presented an approach for developing Adaptive Hypermedia Services
based on separation of the narrative, content and learner into individual models. This
approach encourages the reuse of content, as the content does not embed the logic
used to produce personalized courses. The pedagogical approach and course structure
is instead embedded in the narrative model. This approach enables a single course to
support many pedagogical approaches to structuring learning content. Also proposed
is a mechanism to enable course authors to reference learning objectives, rather than
individual pieces of content, from the narrative, thus facilitating the selection of
learning content at runtime. This content is selected in accordance with the learner’s
content presentation preferences.
The paper also presents the Personalized Learning Service, an implementation of
the multi-model, metadata approach. The PLS is currently being used to successfully
generate and deliver personalized courses in SQL to eighty final year undergraduate
degree students in Trinity College, Dublin. Presented are some initial evaluation
findings from this trial.
Finally the paper discussed the PLS as a service, highlighting how the service
approach enables personalized learning content to be integrated with existing learning
content.
References
1. “ADL Sharable Content Object Reference Model”, Version 1.2 , http://www.adlnet.org/.

2. “AICC CMI Guidelines for Interoperability”, Revision 3.0.1, Release 24 November 1999.
3. Brusilovsky, P.: Methods and techniques of adaptive hypermedia. In P. Brusilovsky and J.
Vassileva (eds.), Spec. Iss. on Adaptive Hypertext and Hypermedia, User Modeling and
User-Adapted Interaction 6 (2-3), 87-129.
4. P. Brusilovsky. Adaptive educational systems on the world-wideweb: A review of
available technologies. In Proceedings of Workshop "WWW-Based Tutoring" at 4th
International Conference on Intelligent Tutoring Systems (ITS'98), San Antonio, TX,
1998.
5. Conlan, O., Hockemeyer, C., Lefrere, P., Wade, V., & Albert, D. (2001). Extending
educational metadata schemas to describe adaptive learning resources. In Hugh In
Proceedings of the twelfth ACM Conference on Hypertext and Hypermedia (Hypertext
2001), pp. 161-162, New York: Association of Computing Machinery (ACM), 2001.
6. Educators Access to Services in the Electronic Landscape (EASEL). EC IST project
10051, http://www.fdgroup.com/easel/.
7. Fretwell-Downing Education, http://www.fdlearning.com.
rd
8. Honey, P. & Mumford, A.: The Manual of Learning Styles, 3 Edition, 1992, ISBN 0
9508444 7 0.
9. IMS Learning Resource Metadata, Version 1.2, http://www.imsproject.com/metadata/.
10. Java Expert System Shell, Version 6.0, http://herzberg.ca.sandia.gov/jess/.
11. JavaServer Pages Technology, Sun Microsystems, http://java.sun.com/products/jsp/.
12. Marchionini, G.: The costs of educational technology: A framework for assessing change.
In H. Maurer (Ed.), Proceedings of Ed-Media 95, World conference of educational
multimedia and hypermedia, Graz, Austria.
13. Stauffer, K.: “Student Modelling & Web-Based Learning Systems”, 1996.
14. Jakarta Tomcat, The Apache Software Foundation, http://jakarta.apache.org/tomcat/.
15. Fleming, N.D. (1995), I'm different; not dumb. Modes of presentation (VARK) in the
tertiary classroom, in Zelmer, A., (Ed.) Research and Development in Higher Education,
Proceedings of the 1995 Annual Conference of the Higher Education and Research
Development Society of Australasia (HERDSA), HERDSA, Volume 18, pp. 308 – 313.
16. Extensible Markup Language, XML, http://www.w3.org/XML/.
17. Extensible Stylesheet Language, XSL, http://www.w3.org/Style/XSL/.
Adaptation and Personalization on Board Cars:
A Framework and Its Application to Tourist Services∗
+
L. Console, S. Gioria, I. Lombardi , V. Surano, and I. Torre
Department of Computer Sciences – University of Torino

Corso Svizzera 185 - 10149 Torino (Italy)
{luca.console, ila.torre}@di.unito.it
Abstract. In this paper we analyse the goals and problems that should be taken
into account when designing adaptive/personalized services that must run on-
board vehicles. This is, in fact, a very interesting and promising area of
application where adaptation and personalization can provide unique
advantages. We then introduce a framework and a multi-agent architecture for
on-board services supporting different forms of user and context modelling and
different forms of adaptation and personalization. Finally, to support our claims
and framework, we discuss a specific prototype system for on-board tourist
services.
1 Introduction
The convergence between the different information technologies has reached by now
the automotive world. Manufacturers and suppliers of electronic systems have been
reproducing, on cars, the whole of appliances and services of both home and office.
The main problem they are facing now, is how such different devices - mobile
phones, TV, computers, PDAs, new generation phones with internet access, CD and
DVD players, GPS, information and entertainment systems -, can be integrated in
dashboards and made compatible with the environment of a car and of a user who is
driving. Indeed, the solutions implemented or proposed so far are only partially
satisfactory and this integration is the subject of several research projects.
The project described in this paper aims at bringing a contribution to these
research issues. The main claim of our work is that adaptation and personalization
techniques can bring interesting contributions to research in this area. Roughly
speaking, in fact, these methodologies and technologies can contribute to making
information services on-board a car more useful and more compatible with the
context of a vehicle and with the driving task since they can be successful in reaching
the goal of providing “the right service, at the right time and in the right way, given
the driver’s ability/preferences and the current contextual conditions”. The aim of our
project is to support this claim. In more detail, we started analyzing the peculiar
problems arising in this context; as a result we defined a framework for adaptive and
∗ The work described in this paper has been carried on in co-operation with Centro Ricerche
Fiat and Magneti Marelli Electronic Systems.
+ Current affiliation: Magneti Marelli Electronic Division

Adaptation and Personalization on Board Cars 113
personalized services on-board cars; finally we experimented the framework

implementing a prototype for providing personalized tourist information (information
about hotels, restaurants, places of interests).
Although adaptation and personalization techniques have been applied to many
contexts and domains, including mobile systems (see [3, 4, 9,] for reviews and [8, 6]
for applications on mobile devices, dealing also with tourist information), the specific
case of automotive systems has not received the attention it deserves. Some
applications have been developed (e.g., [7, 10]), but there is no systematic study of
the problems that have to be dealt with in this special case.
The paper is organized as follows: Sect. 2 analyses the problems and goals for
on-board adaptation/personalization and introduces the framework we defined; Sect. 3
discusses how the framework has been used for implementing a tourist information
service called MASTROCARONTE; Sect. 4 concludes the paper with a few words about
implementation issues.
2 On-board Adaptation/Personalization: Goals and Framework
The starting point of our project was the analysis of the goals that must be taken into
account when designing adaptive/personalized services (systems) on-board cars. This
led us to the definition of a set of requirements for the design of our framework:
− Adaptation and personalization must focus on making a service effective and
usable on a car where the driver has no time to navigate a space of alternatives,
browsing or evaluating different options. This means that the system should be
able to provide only the right service (information), at the right time and using
the right communication system.
− The system should be very easy and intuitive to use (no effort should be required
to comprehend at a glance what the system is communicating and the options that
are open to the user) so that its use must not distract the user from her main task,
driving. This means also that the system should not be intrusive, especially when
it is not appropriate (e.g., driving conditions requiring maximum attention).
− The hardware and software limitations on-board must be taken into account. This
is particularly critical as regards the interface since many I/O devices (such as
keyboards or pointing devices) cannot be available.
− The services should be location based; this implies a connection with systems
such as the GPS and the car navigation system.
− The system should be able to operate in different modes: on demand answering a
user request or activating autonomously (if this is compatible with the situation).
− The system should adapt to at least two different aspects: (1) the user – adapting
to her features (dimensions), some of which are general and domain independent
(e.g., her capabilities or propensity to spend), while others depend on the specific
service (e.g., her tourist interests) and (2) the context of interaction. The latter
includes dimensions such as the car and driving conditions, the weather, the time
of the day, the presence of other passengers (see next item) and is a peculiarity of
the application on-board a car.
− The system should take into account that the same user may have different
preferences in different situations or contexts (e.g., when travelling alone or with
114 L. Console et al.
the family, when travelling for business or for pleasure). This means that the
system should be able to deal with multiple views of a user (on the user model).
− The system should support different forms of adaptation and personalization,
each of them requires to take into account both the user’s profile and the context:
!" Content adaptation: selection of the specific service to be suggested,
!" Detail adaptation: the amount of information to be presented (possibly on
demand),
!" Interface adaptation: the interface and channels for presenting the
information,
!" Behavior adaptation: the behavior of the system, especially as regards the
autonomous activation and the autonomous suggestions.
− The user model should follow the user on multiple cars: an individual should
receive personalized services on any car equipped with the system and the system
should support multiple users of the same car. This means that the user model
must be located on a portable support (e.g., a smart card) that can be inserted into
any car on which the adaptive system is installed.
− The user cannot be involved too actively in the user modelling process; in
particular, we cannot expect that this process can be performed in an interactive
and co-operative way. The system must be able to get and exploit at best indirect
(and weak) knowledge about the user’s behavior.
− Multiple services should be supported on the same car and should be integrated
into a unique interface (or unique device). This can be achieved if the services are
based on a client/server model, where the client on the car connects to a server
via the telephone line. This connection, however, can be critical since it may be
unavailable (e.g., in a tunnel or in a remote area) and is quite expensive and slow.
As a consequence, the number of connections and the amount of information to
be transferred should be limited as much as possible. On the other hand we
cannot expect to have powerful computing and storage resources on-board a car.
Architecture of the system. Starting from these goals, we designed a framework and
architecture for on-board adaptive/personalised systems (Fig. 1). In the following we
shall analyse the framework (using the tourist services application when needed to be
more concrete). The architecture is designed to support multiple services; a specific
server has to be set for each one of them. The architecture is constituted by a set of
agents, distributed in part on the car and in part on the server.
The interaction between the user and the system is mediated by an interface agent
running on the car.
The user model is stored on a smart-card that the user can insert into the system
on any car. This supports mobility of the user. The model is initialised on the server
using stereotypical information. The user model is then updated periodically by taking
into account the user’s behavior; this can be done either on the car or again on the
server. This process is performed by an agent which takes into account statistics on
the user behavior collected with the help of the interface agent. Performing the update
on-board has the advantage that these data need not be transferred via the GSM line;
however, in case the user has multiple cars, the statistics have to be stored on the
smart card. On the other hand, performing the update on the server may allow the
system to take into account data about the user collected on multiple cars and the
behavior of other users (to implement co-operative approaches to user modelling).
User Answers to
model questionnaire Server
card
Interface Agent for

agent User model Stereotypes
User model initialization
Context
On-board
Service Service
Managm.
Analysis of Manager DB
agent
Feedback and
User model GSM Analysis of feedbacks and
Replicated communication User model revision
revision
Service
DB on CD
Fig. 1. An architecture for on-board adaptive/personalized services
The user model must contain at least two types of information

− Domain specific information, depending on the application. In the case of tourist
services we maintain information about the user interests (e.g., for topics such as
art, architecture, museums, natural parks, …), preferences (e.g., about type of hotel
or restaurant) and life-style (propensity to consume and spend).
− Domain independent information, concerning, e.g., the user capabilities, cognitive
characteristics, preferences as regards the interface and the modes of interaction
and the level of autonomy of the system.
The context model contains information about the context of interaction, that is:
− Information about the position of the car, obtained by the GPS installed on the car.
− Information about the driving conditions (e.g., speed, duration and direction of the
travel, the fact that the car is inside a metropolitan area or not, estimated level of
traffic, presence of passengers). These pieces of information can be obtained from
trip computers and sensors on the car (possibly after some computation).
− Other useful information such as the date and time of the day or weather conditions
(these can be obtained from sensors available on the car).
The service data base (DB) are located on the provider site. Whenever it is possible, a
replication is distributed on a CD/DVD to be used on the car. In MASTROCARONTE
the DB contains information about hotels, restaurants and places of interest (like in a
tourist guide), augmented with information useful for personalization.
Three agents are in charge of adaptation/personalization: the interface agent and
the on-board manager are on the car, and the service manager is on the provider’s site.
The on-board manager is responsible of deciding when a service has to be provided
(after a user’s request or activating autonomously). It also decides which portions of
the user and context model are relevant for personalization. It then sends a request to
the service manager agent on the server.
The service manager, given information about the user and context model, selects
the appropriate personalized service (information) and sends it back to the car. In
MASTROCARONTE it selects from the tourist database the hotels or restaurants or
places of interest (depending on the type of request) that are most appropriate, given
the user preferences and the context (position, time of the day).
The on-board manager receives these pieces of information and evaluates them
according to the complete user model. In MASTROCARONTE it ranks the list of items
received from the server given the complete user model.
If the service DB is replicated on the car, a protocol for reducing the amount of
information to be transferred can be adopted: (i) the server only sends the keywords
of the items to be presented (ii) the on-board manager can check if the information on
the CD is up-to-date (the keyword is replaced at each update, so that if the keyword is
available in the DB on the CD, then the information on the CD is up-to-date), (iii) it
asks to transfer data from the server only if needed. This is the protocol adopted in
MASTROCARONTE which uses a replicated tourist DB on a DVD on the car.
The interface agent is in charge of selecting the right medium and form of
presentation, given the user and context model. For example, it may decide to use a
voice synthesizer to provide information in safety critical situations (e.g., high traffic
or speed) or to use the video display in other conditions (e.g., low speed, non
metropolitan area, limited traffic). Moreover, it decides the amount of information
and detail to be provided (detail adaptation).
This framework can be used to implement different types of services. In the
following section we discuss in more detail MASTROCARONTE, which is the first
prototype we implemented based on the framework.
3 MASTROCARONTE, an Application for Tourist Services
In this section we discuss MASTROCARONTE, a specific instantiation in the domain of

tourist services of the framework discussed in the previous section. In the following
we analyse in more detail the various agents and modules in Fig. 1.
User and Context Model

The analysis of the dimensions of the user to be modelled has been carried on using
the methodology proposed in [11, 12]). The methodology suggests that an application
is analysed by performing a decomposition of the adaptation goals and of the features
of the application domain and context. The resulting matrix provides guidelines for
choosing which aspects of the user are most useful (for achieving a given goal, in a
given application domain and context). We performed this analysis on several
potential applications on-board cars and this allowed us to single out the distinction
between a portion of the user model which is common to (almost) all the applications
(because it is mainly related to the goals of adapting the interaction with the user on a
car) and a portion which is domain dependent.
The former includes dimensions such as the user cognitive characteristics
(amount of information that she can read at a glance, level of attention that she can
pay to the on-board information system, …), her life style and propensity to spend
and consume and preferences (e.g., on the interface, on the medium to be used in the
interaction, on the level of initiative that the system may take..).
The latter includes, in the tourist service application, aspects such as the user’s
tourist interests (useful for selecting the most appropriate places of interests), her
preferences concerning food or location or type and level of comforts of
hotels/restaurants, and the propensity to spend under some specific conditions, such as
kind of travel, (useful for selecting the most appropriate hotels or restaurants).
The approach we adopted for user modelling is a modular one, where different
dimensions of the user are dealt with in a separate way (following a methodology
already adopted in [1, 2]). This has several advantages since user modelling
knowledge can be re-used in multiple applications for which the dimension is
relevant. In MASTROCARONTE we considered three dimensions: Cognitive
characteristics, Interests, Life style.
We defined stereotypes for these dimensions, starting from psychographic studies
about the Italian population (and specifically about their cultural interests) published
by Eurisko [5]. We thus have three groups of stereotypes:
− Interests. Starting from classificatory data such as the age, gender, type and field of
work, hobbies and recent travels, these stereotypes make a first prediction on the
user's interests in art (ancient or modern), technology, amusements, nature, ….
− Cognitive characteristics. This group of stereotypes uses classificatory data such as
the user's age, education level, job and makes a prediction on the user's receptivity,
a parameter used to determine the amount and detail of the information that can be
presented to her.
− Life style. These stereotypes classify users according to their psychographic
features, which include socio-demographic data and priorities. They make
predictions about the user propensity to spend and consume and about the preferred
types of tourist facilities (e.g., hotels and restaurants).
The user model is generated by matching the data provided by the user with the
stereotypes and then generating and merging the corresponding predictions. In
particular, each prediction regards a feature in the user model and corresponds to a
probability distribution for the linguistic values of the feature.
For the sake of brevity we omit the details of this classification process (which is
a fairly standard one based on probabilities and Bayesian classification, see [1]), and
we only report an example of a fragment of a user model:
interest for art(modern) very-high: 0.4, high: 0.3, medium: 0.2, low: 0.1, null: 0
interest for architecture very-high: 0.2, high: 0.3, medium: 0.3, low: 0.1, null: 0.1
interest for shopping very-high: 0.1, high: 0.1, medium: 0.1, low: 0.4, null: 0.3
….
receptivity: very-high: 0.5, high: 0.4, medium: 0.1, low: 0, null: 0
propensity-to-spend: very-high: 0.1, high: 0.2, medium: 0.4, low: 0.3, null: 0
For example, the first line specifies that the probability that the user’s interest for
modern art is very high is 0.4, etc.
The user model is refined (or even revised) after tracking the user behavior. In
particular, the system records the actions performed by the user whenever she
receives a suggestion and in this way tries to infer the user’s actual
interests/preferences. We noticed that we cannot ask the user to provide direct
feedback; nevertheless there are many events that can provide indirect (even if
uncertain) feedback. For example, strong feedback can be obtained whenever the user
makes a phone call or asks for further information about a place (this can be done
using facilities integrated into our interface, see Fig. 2). Weaker feedback can be
obtained from the GPS: from the position of the car (and in particular from that of the
parking) the system can guess which are the places that the user actually visited.
Very roughly, the revision is performed in three steps:

− Feedback (direct and indirect ones) about the user’s behavior are collected across
time and stored on the memory card containing the user model, with different
degrees of reliability associated with them.
− Statistics about the user’s behavior are computed periodically. In particular, these
statistics regard the type of places the user visited and their characteristics. For
example, for places of interest we compute statistics about the type of places that
have been visited (e.g., related to modern art or to architecture or to shopping; this
information can be extracted from the tourist database, see next paragraph).
− A set of rules is activated periodically to compute, from the statistics above, new
probability distributions for the features in the user model. For example, if the user
visited all the time places which are highly relevant for modern art, we may get the
following new probability distribution for the user interest in modern art:
interest for art(modern) very-high: 1, high: 0, medium: 0, low: 0, null: 0
This new distribution is then combined with the one in the current user model. The
two contributions are weighed according to the following schema:
• in the early phases of the use of the system (when the current user model mainly
comes from stereotypes), the distribution computed using feedback about the
user’s behavior and the revision rules has a stronger impact;
• this impact is progressively reduced and, after some time, the distribution in the
current user model becomes the most relevant one, so that the rules can only
perform a slight revision of the user model.
Database of Tourist Information

This database is a standard tourist database, augmented with additional information
useful for performing a personalized selection of the items.
For hotels and restaurants, it maintains information about the category, style,
level of price, type of amenities, look of the place, etc.
For places of interests it keeps track of how relevant the place is with respect to a
set of categories (which are those in terms of which the interests of a user are
classified in the user model, see above): art (ancient and modern), nature, architecture,
technology, amusements, shopping, etc. The relevance is defined on a scale from 0
(not relevant) to 4 (highly relevant). For example, the “Museum of Cinema” of Torino
(a museum concerning the history and techniques of Cinema), has the following
attributes:
relevance for: art (modern) = 4, technology = 3, amusements = 4
For each item we also keep track of its absolute location (in terms of GPS
coordinates). Finally, each item is identified by a unique key, which is changed
whenever a piece of information about the item is updated. The database is stored on
the server and replicated on a DVD to be used on the car. Through the identifying key
it is possible to discover, on the car, if the information on the DVD is up to date.
On-board Management Agent

This agent decides if and when the system must activate autonomously to provide
recommendations to the user. This is decided given information in the user model
(user preferences about the level of the system’s autonomy and statistics about the
situations in which the user likes to receive advice).
Whenever the system is activated, the agent is responsible of generating the

query to be sent to the service manager. The query concerns one of the services
(hotels, restaurants or places of interests) and includes information about the location
(GPS coordinates) of the car and about the user model. The latter is the list of the
interests/preferences for which the cumulative probabilities of the levels “medium”,
“high” and “very high” is over a threshold (0.66). Thus, the typical structure of a
query is the following (using am abstract SQL):
Select identificationKey
From (restaurants OR hotels or placesOfInterest) {depending on the query}
Where Location closeTo currentGPSposition AND
Attrubutes includeAtLeastOneOf UserPreferences
The query is sent to the server to the information retrieval agent.
When it receives the answer from the server, the query management agent ranks
the items in the reply. This is done by taking into account the preferences in the user
model and the information about each item contained in the tourist database. Let us
consider the case of places of interest. The ranking is obtained by computing a degree
of match between each item (and in particular its degree of relevance with respect to
the categories of interests, e.g., art(modern), architecture, …) and the preferences in
the user model (probability distributions regarding the same categories).
Service Manager Agent

The service manager agent is activated by a query sent from the car and is responsible
of retrieving the items that are most appropriate, given the information about the car
location and user’s preferences carried by the query. It queries the database and
returns the list of identifiers of the retrieved items. The selection takes into account:
− The location of the car: the items must be not farer than X km from the location of
the car (the actual value of X depends on the location – metropolitan or not – and
on the time of the day), possibly in the direction of the trip. In case the criterion is
too restrictive (less than 10 items are retrieved), it is weakened and the distance
doubled (this, however is applied only if the query was started by the user).
− Contextual information about the date and time so that all the items which are not
open or accessible are discarded.
− The preferences of the user: only the items which have a relevance of at least 2 for
at least one of the preference attributes associated with the query are retrieved.
Interface Agent
The interface agent is responsible of collecting requests from the user (who simply
has three buttons for asking about hotels, restaurants or places of interests) and of
presenting the list of suggested items as ranked by the on-board management agent.
The adaptation of the presentation depends on three main aspects:
− The context in which the interaction is taking place, that is the driving conditions
(e.g., speed, traffic, metropolitan area vs. motorway, weather, time of the day).
− The user’s cognitive characteristics. These include long term information stored in
the user model (i.e., the user receptivity which estimates the amount of information
that the user can read at a glance) and short term information concerning the user
tiredness (estimated given the time of the day and the number of hours driven).
− The user preferences concerning the style and layout.
The user interface agent evaluates these conditions every time it receives from the
query management agent a list of ranked items to be presented to the user. The
decision to be made is the choice of the presentation format, which corresponds to
choosing how many items to list in a page and how many pieces of information and
services to associate with the selection of each item. In order to simplify the process,
we decided to define a fixed set of presentation formats so that the interface agent
must select the most appropriate one among these formats. We use XML for the
exchanges of information and each format is defined by a style-sheet; thus, the
generation of the presentation simply amounts to the application of the selected style
to the ranked list of items. In the current prototype we defined five style-sheets.
Examples of presentation are showed in Fig. 2. All the screenshots refer to the
same driver but those in the first row refer to driving conditions that are less critical
than those in the second row. The screenshots on the left correspond to the list of
items presented to the driver (ten per page in the first case, four per page with larger
fonts in the second). The screenshots on the right refers to the details about one item
(a restaurant in this case) with a personalized ranking (“Voto” on a scale from 1 to 5)
and the possibility of activating other services (e.g., “Chiama” for making a phone
call or “Strada” for getting the route from the navigation system). Notice that in the
second row the number of available services is reduced.
Fig. 2. Examples of interaction

4 Summary and Implementation Issues
In the paper we introduced a framework that addresses the goals and problems to be
taken into account when designing adaptive/personalized services on board vehicles
and we discussed an instance of the framework in the area of tourist services.
The framework has been implemented as a multi-agent system, where the agents
running on the car are embedded in the VxWorks real time operating system used by
Magneti Marelli on its car navigation systems. In particular, we implemented the
agents using CLIPS (a rule-based system developed by NASA), embedding CLIPS-
based agent in VxWorks. All exchanges of information are based on XML schemas.
The prototype of MASTROCARONTE runs on a PC emulation of the CONNECT car
navigation system by Magneti Marelli, installed on cars of the Fiat group (the black
buttons in Fig. 2 correspond to the physical buttons in the dashboard of the car).
References
1. L. Ardissono, L. Console, I. Torre: An Adaptive System for the Personalized Access to

news, in AI Communications, Vol 14, no. 3, 2001, pp. 129-147.
2. L. Ardissono, A. Goy, G. Petrone, M. Segnan, P. Torasso: Tailoring the recommendation of
tourist information to heterogeneous user groups, in Springer LNCS 2266, pp. 280-295.
3. P. Brusilowsky: Methods and Techniques of adaptive hypermedia. In User Modelling and
User Adapted Interaction 6 (2-3): 1996, pp 87-129.
4. P. Brusilowsky: Adaptive Hypermedia. In User Modelling and User Adapted Interaction 11
(1-2): 2001, pp 87-110.
5. G. Calvi: Indagine Sociale Italiana, rapporto Eurisko, Franco Angeli, 1986
6. B.N. De Carolis, F. de Rosis: in Proc. Worksop on AI in Mobile Systems, Seattle 2001.
7. C-N Fiechter, S. Rogers: Learning subjective functions with large margins. In AAAI Spring
Symposium on Adaptive User Interfaces, Stanford March 2000, AAAI , pp. 40-47.
8. J. Fink, A. Kobsa, A. Nill : Adaptable and adaptive information for all users, including
disabled and elderly people, in New review of Hypermedia and Multimedia, Vol. 4, 1998,
pp. 163-188.
9. A. Kobsa, J. Koenemann, W. Pohl: Personalized hypermedia presentation techniques for
improving online customer relationships, in The Knowledge Engineering Review, 2001.
10. C.A Thompson, M. Göker: Learning to Suggest: The adaptive place advisor. In AAAI
Spring Symposium on Adaptive User Interfaces, Stanford March 2000, AAAI, pp. 130-135.
11. I. Torre: A modular approach for user modelling. In Adaptive Hypermedia and Adaptive
Web-Based Systems 2000 (AH 2000), Springer LNCS 1892, pp. 414-420.
12. I Torre: Goals, tasks and Application domains as the guidelines for defining a framework
for User modelling”, in Springer LNAI 2109, pp. 260-264.
Adaptive Authoring
of Adaptive Educational Hypermedia
Alexandra Cristea and Lora Aroyo

Eindhoven, The Netherlands
{a.i.cristea, l.m.aroyo}@tue.nl
http://wwwis.win.tue.nl/~{alex, laroyo}
Abstract. In this paper we propose a set of development guidelines for an

adaptive authoring environment of adaptive educational hypermedia. This set
consists of relevant and necessary functionalities and architectural features of
authoring systems (AS) for adaptive teaching and/or learning environments
(LE). We extracted the core functionalities by analysing MyEnglishTeacher
(MyET) and AIMS - two independently designed and built AS for adaptable/
adaptive LE. The extended core led us to a concept-based layered approach
with respect to concept- and attribute-level adaptation, as well as lesson and
presentation adaptation. We believe that providing adaptive authoring support
for adaptive hypermedia will have a strong impact on the authors’ motivation
and efficiency in performing their tasks and consequently will increase the
popularity of adaptive hypermedia.
1 Introduction
To keep up with the high demands in educational software and LE, it is already
widely accepted by the educational community that the future of such environments
lies in adaptability and adaptivity [11,13,18,19]. These high demands reflect also on
the authoring environments, general purpose or application oriented. At present, AS
for adaptive LE are almost non-existent, mainly because of the field novelty and the
growing demands– making any authoring system for LE outdated before it is used.
However, with the ripening of the field, standardization of adaptive and adaptable
techniques and methods [10,15,16,22] is starting to preoccupy the research commu-
nity, and AS become a serious need. Still, it is not reasonable to expect teachers to
skilfully apply adaptation to their new on-line courses. Transition from linear, book-
like courseware is not easy. Teachers are confronted with the heavy task of designing
contents alternatives, adaptation techniques and ultimately, the whole user-interaction
mechanism. It is clear that for adaptive courseware the authoring tool has to offer
tuneable complexity and automatically perform many of the authoring tasks.
In this paper, we build upon our previous research by shortly reviewing two AS for
adaptable/ adaptive LE, MyET and AIMS, which we designed and built independ-
ently. From our experience with these two systems we are extracting a set of func-

Adaptive Authoring of Adaptive Educational Hypermedia 123
tionalities that are relevant and necessary for AS for adaptive teaching and/or LE.
Given the complexity of the authoring task such environments [14], we independently
came to the conclusion that the authoring tool has to be also adaptive to the teacher.
Therefore, we extend our previous research to an adaptive authoring framework
based on the above-mentioned necessary functionalities.
2 Two Methods for Authoring Support of Adaptive Courseware
In this section we are presenting the two independently designed, developed and
tested systems for courseware authoring, with special focus on adaptive courseware:
(a) MyET [9], developed at the University of Electro-Communications in Japan, and
(b) AIMS [2], developed at Twente University in The Netherlands. In both MyET and
AIMS concept mapping paradigm [4,7] is used as a main structure to organise the
subject domain terminology and to link to course items (lessons, exercises, docu-
ments). We outline the common features shared by these systems with respect to
course content organization, maintenance and presentation, with the purpose of re-
usability and student adaptation. Our evaluation framework is based on the: (a) gen-
eral course structure; (b) text presentation and structure; (c) lesson composition; (d)
authoring views and (e) student adaptation facilitation. Due to lack of space, we have
skipped the analysis of tests and exercises creation, structuring and linking.
2.1 General Course Structure
The information exchange from tutor to system means input of lessons, texts, links,
but also asking for help in editing, etc. The data from the tutor is stored in a structured
way, as shown in Figure 1. Input can be audio, video, graphic or text.
Teacher-name COURSE: <name> <description>

AUDIO FILMS PICTURES LESSON INSTRUCTOR: <name>
113-1 115-1 119-1 123-1 190-1 196-1 65-1 DOMAIN: <name>
114-1 118-1 119-2 123-2 190-2 197-1 67-1 TOPIC-1: <name> <description>
EXERCISES conclusion keywords MAIN-TEXT: <link>
title TEXTS explanation pattern TASK-1-1: <name> <keywords> ! TEXT(s)
TEXT1 TEXT2
exercise1 keywords pattern TASK-1-N: <name> <keywords> ! TEXT(s)
(a) (b)
Fig. 1. Data structure in MyET (a) ; Course structure in AIMS Course Editor (b)
a. In MyET, the first three inputs have also a text version in the lesson and with
automatically generated index. Each lesson object is subdivided into exercises,
conclusion, keywords, title, explanation, pattern to learn and text objects. These
are further subdivided into exercise, keywords, pattern, title and actual main text.
b. A related structure appears in Fig. 1b. The teacher in AIMS creates the course
structure as part of an information model, defining the subject domain (course
124 A. Cristea and L. Aroyo
terminology/concepts), library (collection of texts), course (lessons and exercises)

and user profile objects and expressing all the links between them and their com-
ponents [3]. Each course lesson is divided into exercises, which are directly re-
lated to subject domain concepts and this way also to related documents (texts).
This domain-specific structuring and concept-based linking are provided [17].
2.2 Text Presentation and Structure
The smallest block / object in the course structure is a TEXT.

In MyET, each text also has (next to text body), some obligatory attributes: a short
title, keywords, explanation, patterns to learn, conclusion and exercises. This way,
not only titles and keywords but also explanation and conclusion files are used for
search and retrieval. Moreover, the corresponding text to any video/audio recording
allows any non-text resource to be retrieved via a text-based search.
The authors in AIMS are offered the choice to select keywords from the domain
terminology they have created as a concept map. These keywords also make the link
directly to the lessons (course structure of course topics and exercises). The AIMS
library editor provides the authors with presentation and instructional formats to map
each course related material in such a way that it is both task- and use-oriented [2].
2.3 Lesson Composition
One or more texts (with multimedia or not) build a LESSON object. In MyET, each
lesson also has (beside of texts, etc.) the following attributes: title, keywords, explana-
tion, conclusion, combined exercises (generated automatically or not). This structure
is very similar to the text object structure.
A text or lesson is generically called ‘SUBJECT’ in MyET or ‘TOPIC’ consisting
of course ‘TASKS’ in AIMS.
2.4 Student Adaptation Issues
In the MyET environment, adaptation to the students’ needs meant interpreting the
concept maps and the links created by the teacher. A global agent would copy the
map to create a global student model that would serve as a guide for all students.
Moreover, a private agent would make its own copy and alter it with respect to the
student’s needs and based on the interaction with the student. The system agents work
based on the embedded rule/knowledge systems. They act as learning objects, which
can adaptively change their representation of the subject space [10].
The adaptation to the students’ needs in AIMS is performed with a team of col-
laborating agents with respect to (a) user-oriented presentation of the information,
based on the general work progress of the student on the course, and (b) students’
search activities. This is realised by refining the student search query according to the
current course task context, by adjusting the search result presentation with relevance
to the course task and by providing students with alternative view of the results and
thus alternative ways to build a conceptual structure of the course knowledge [2].
2.5 Courseware Views for Authors
Once a map has about 50 subjects representing a number of concepts with their links,
it is impossible to get a sensible display in one map [1]. To manage the complexity
we designed extra courseware views showing different sub-sections: “views” of
whole graph, with reduced information (“bird’s eye views”); one concept and its
“star”-links (all concepts currently linked to it) (“fisheye views” [9,11]); non-linked
concepts: “floating”-concepts [10]; one concept and its “star”-linked documents
(texts) [3]; all link types currently in use [2]; all concepts ordered alphabetically by
name, by description and category; the place in the concept map of every newly de-
fined concept.
3 Our Concept Mapping Layered Approach to Adaptive Authoring
From the analysis and extraction of common proprieties of the two AS, a more intui-
tive division seems therefore to be to separate the course material into concepts
[2,10], as derived from the concept-mapping paradigm [7]. A low level concept
should represent an atomic piece of content/ information that has an independent
semantics. This atomic unit can be labelled with one concept. Collections of concepts
can, of course, build composite concepts, generating a concept hierarchy. Concepts
can be related to each other at any level of the hierarchy. The creation of these build-
ing bricks is the role of the course designer [3,9]. The division of the content into
concepts only gives us the primitive building blocks of the courseware. Putting these
building blocks together with different sequences generates different presentations or
lessons. This can be done by a course designer or by a teacher. In a more advanced
environment it can be automatically generated by a system [10].
At this level, we would only speak of adaptive navigation support. That is because
adaptive presentation is (normally) at a lower level than the concept level and binds
actually parts (fractions) of concepts with each other. Clearly, it makes no sense to
just transform parts of concepts into sub-concepts, as it is possible that they make no
sense but in the context of the atomic concept (so have no independent semantics
attached).
"#Example: Consider, for instance, the case of the introduction to some text. This
is a construct that appears very often and that can be dropped in later versions of
the browsing (or used together with other introductory fragments in an introduc-
tory chapter). However, this construct usually has no independent meaning.
The solution is quite obvious [9]: the concept can be sub-divided into its attributes.
These can be anything from a concept name to alternative contents or fragments.
By dividing/mapping the course content into a concept hierarchy, and the concepts
into a set of attributes, the adaptation has only to deal with concept-level adaptation
and attribute adaptation. The advantage is that it can all be performed (and viewed)
from a high level and does not need separate consideration of different conditions
written within the text, which are more difficult to re-use by other authors. Basically,
the adaptation becomes only a matter of combining concept attributes into pages
(pieces of information that can be show at a time), chapters and subchapters. This
way, the adaptation is only at a content level (equivalent to adaptive presentation),
while the navigation is only dependent on the presentation format. We will return to
these issues in Section 3.1 (e.g., short pages will mean that the “next” button within
the same lesson appears more often, but the content of the page is shorter [20]).
3.1 Lesson Map
A lesson map is, in the simplest case, the lesson sequence that tells the student how
the lesson should proceed (according to the teaching style, learning style or learning
goals). In a more general case (Figure 2), the lesson map is a directed (not necessarily
a-cyclic) graph with at least one beginning (START) and at least one ending state
(GOAL). Circles here represent either whole concepts, or concept attributes.
Fig. 2. Alternatives ‘start’ and ‘stop’ positions for the lesson map
The START and STOP positions do not have to be unique. The user model (UM)
based adaptation engine determines the selection between alternatives.
In Figure 3 we present a sketch of the concept mapping layered approach to lesson
adaptation. A lesson X is composed of some concept fragments (attributes) that are
grouped into chapters (here, 2). At presentation, there is another separation into pages
(here, 6 pages). The latter is presentation means dependent (e.g., laptop screen
browser, hand-held device, etc.).
It is interesting to note that, although the fragments themselves have no semantics
(and therefore could not have been subdivided into concepts), they can be addressed
via the concepts they belong to.
3.2 Authoring Adaptive Courseware with Concepts
Next we present two types of concept editing processes. This way we show that the
transition from classical course editing - to pure concept editing style is an easy one.
Fig. 3. Lesson adaptation: lesson X (straight line square), composed of, e.g., 3 concepts A, B
and C (cylinders), with attributes (darkened cylinder); presentation order shown by directed
connections between attributes; lesson has 2 chapters, that contain (parts of) concepts (dotted
line squares). In chapters information is presented (in browser window, e.g.) in pages (point-
dotted line squares). “Next” buttons at page level are navigation support for presentation only.
3.2.1 Concept–Based Authoring in Traditional Order

In order to make the transition from traditional editing to concept editing easier a
concept-based authoring tool should be able to allow, in extreme cases, pure tradi-
tional editing. In this section we present a five-step procedure for concept editing in
traditional order, illustrated in Figure 4.
1. First, the author writes sequentially the text of the course/lesson together with the
respective multimedia. The instructor can stop the process here or at any of the
next steps and let the system automatically perform the rest of the steps and show
the results again to the teacher for approval, or to another author for reusing.
2. The content is divided and organised into a concept structure. First, the concept
hierarchy of concepts, sub-concepts and atomic concepts is created, where the
atomic concepts are the smallest semantically indecomposable building blocks
within this structure. Further, the main attributes of the concepts are filled in, such
as concept name and content and possibly any other related attributes [9].
3. Adaptive features are added by writing a separate rule-base [22] or setting impor-
tance coefficients and weights [10] and generating at least one lesson map (se-
quence). In some cases it may be necessary here to return to STEP 2 in order to
refine the division granulation or to STEP 1 to add more concepts/ material.
4. The author can define the different items to display depending on certain condi-
tions (only mentioned in the separate rule-base, started at STEP 3: lesson maps).
"#At first glance, we have here alternatives (such as alternative texts in differ-
ent languages for the same concept/ lesson/ etc.) or conditionals (such as addi-
tional information that is only presented if “conditions are right” – classical
adaptive presentation). However, from the adaptive engine point of view, this
distinction cannot be meaningful, as conditionals can be seen as alternatives,
where, if the condition is not satisfied, nothing is displayed (instead of an al-
ternative). Here we don’t discuss the cases where empty information pages re-
sult due to, e.g., bad authoring. This is a matter of presentation means adapta-
tion and not user-modelling adaptation.
After new attributes have been generated the author can return to STEP 3 to refine
the rules. If necessary, s/he can jump to STEP 2 to refine granulation of division or
to STEP 1 to add more concepts/ material.
5. The author creates the lesson maps following the procedure from Section 3.1.
Step 1: write text + multimedia of lecture with an editor
Text + MM
Step 2: divide content into concepts
Concept 1,
Concept 2,
Concept k
Sub-step 1: define concept hierarchy (concepts and Sub-step 2: define (at least) main attributes for ALL concepts
add
sub-concepts) (such as concept name, concept content- maybe more)
more
content father concept
name
sub-concepts
concept
[...] content
atomic concepts
Step 3: add adaptive features
refine refine the rules

granularity
Step 4: define secondary attributes
Step 5: Form new lessons with the concepts (sub-

concepts and concept attributes)
Fig. 4. Concept-based editing: traditional order
The procedure above allows the author to be as precise and detailed as s/he wants
but at the same time, it allows him/her to do as little authoring as possible. This is
realized by the clear division between the authoring stages. An author/ teacher can be
just content creator – or the creator of non-adaptive hypermedia (STEP 1). Other
authors nevertheless can reuse and refine the created content, by performing the steps
starting from STEP 2. In this way, we have accomplished several goals:
"#Simple authoring, yet
"#Complex results; moreover
"#Collaborative authoring.
The latter is made possible by the semantic structure, based on a flexible, growing
common ontology based on well-annotated concepts (as concept attributes can also
take over the role of concept annotations).
Moreover, the transition from STEP 1 to STEP 2 is not unique and loops to refine
the granulation of the concepts are possible – so, subdivisions of higher level into
lower level concepts – as long as the units obtained still have an independent seman-
tics. This structuring is in the sense of the Semantic Web, while the division of con-
tents into concepts, with their attributes and links is similar to the RDF W3C standard
recommendations (of separating resources, literals and properties).
3.2.2 Non-traditional Authoring: Direct Concept Editing

The previously presented method is only to make the bridge between traditional (clas-
sical) editing and concept-based editing. In a purely concept oriented editor, an au-
thor would follow the following steps:
STEP 1: write concepts + concept hierarchy
STEP 2: define concept attributes (define main and extra attributes)
STEP 3: fill concept attributes (write contents)
STEP 4: add content related adaptive features regarding UM (des. alternatives, conditions)
STEP 5: define format (presentation means-related; define chapters)
STEP 6: add adaptive features regarding presentation means (define pages).
3.3 Resulting Layers, Author Adaptation and Automatic Processing per Layer
Concluding from this procedural explanation, it is clear that the courseware resulting
will have a layered structure as illustrated in Figure 5.
Fig.5. Authoring adaptation layers
1) The first layer is the conceptual layer, with

a) a first sub-layer of atomic concepts, which cannot be changed anymore.
Adaptation and automatization at this level can mean search of related (possi-
bly identical) concepts, in order to warn the author that the new concept s/he
wants to define possibly already exists. Also, this is the moment to automati-
cally establish connections between concepts (computed by system via heuris-
tics [9] and approved by the author).
b) a second sub-layer of composite concepts, which are sets of concepts of atomic

granulation or larger, and which have a hierarchical structure. Course design-
ers and/or system can change this hierarchy automatically. In this layer, con-
nections as relatedness connections [10] also appear.
The system can perform concept – sub-concept adaptation by helping in sug-
gesting relevant higher concept classes, etc.
2) The second layer contains lessons (the hierarchy of chapters, sub-chapters and
the directed graph called lesson – Section 3.1). This represents the way and the
order in which the concepts should be taught. Note, that the lesson layer deals
with concept attributes and not necessarily directly with concepts. This allows
having both adaptive navigation support and adaptive presentation within the
same layer and guarded by the same mechanism. The system can search for al-
ternative existing orders and make the author attentive to other possibilities or in-
quire in a dialog with the author for the possibility that the order is not compul-
sory.
3) The third layer contains
a) a UM-based sub-layer, an adaptation engine that specifies what material
should be presented and when (under which conditions or equivalent sets of
conditions – Section 3.4).
From this point, user adaptation means student adaptation.
b) a sub-layer, formed by the presentation means based adaptation. This adapta-
tion part is concerned with the formatting so that the information appears
nicely in the page, with questions such as the ideal page length – so where
chapters should be cut to form pages, how and where multimedia presentations
should appear, colours, fonts, etc. These matters should, in the simplest author-
ing version, be generated fully automatically in order to simplify the task of
authors. Of course, such issues should also be designable by authors, if neces-
sary [20] Here, rhetorical structure adaptation should be overlayed.
4 Conclusions
Providing adaptive authoring for adaptive hypermedia is quite a crucial task as the
authoring process involves, beside of what was mentioned in this paper, also a num-
ber of other complicated tasks, such link-checking (e.g., issues of termination and
confluence [22]) which become almost impossible for a human to keep track of.
Therefore, next to the adaptation support with respect to the content organization and
presentation, an important issue is also the provision of support tools to analyse and
monitor the information input by the author. The editing environment must provide
support for a number of editorial tasks, such as information search and retrieval, in-
formation visualisation, selecting, restructuring, annotating information with meta-
data, generation of adaptive user feedback and user preferences information. In order
to make the course related content and knowledge to be most efficiently maintainable
it is of a vital need for the authors to be provided with facilities to view the content
from different perspectives and to perform various analyses and statistics on it.
Note that the few authoring systems which allow auhoring of adaptive hypermedia
are quite restrictive. The well-known InterBook shares the concept-based approach,
but relies only on the simple overlay model, and is based on a strict prerequisite struc-
ture [6]. A more advanced system, Tangow [8], has a top-down approach (whereas
here we describe a bottom-up one), and requires a predefined set of concept (here,
task) attributes – whereas here we allow more flexibility.
We obtained from the analyses and tests of two separate, independent systems
(MyET and AIMS) a framework of a concept-based, layered architecture and guide-
lines for adaptive hypermedia, in concordance with the stratified hypermedia struc-
ture for information disclosure [5]. This way we set the basis towards standardization-
based authoring [15, 16, 21].
References
1. Ackerman, F., Eden, C., Cropper, S.: Cognitive Mapping – a user guide,
http://www.banxia.com/depapsum.html (1993)
2. Aroyo, L., Dicheva, D.A.: Concept-based approach to support learning in a Web-based
support environment, In: Moore, J. (eds.) Proc. of AIED’01, Leipzig: IOS Press (2001)
1-12
3. Aroyo, L., Dicheva, D., Velev, I.: Conceptual visualisation in a task-based information
support system, In: Bourdeau, J., Heller, R. (eds.): Proc. of EdMedia’00 Conf. (2000)
125-130
4. Beyerbach, B.: Developing a technical vocabulary on teacher planning: preserves teachers’
concept maps, Teaching and Teacher Education, Vol. 4 (4) (1988) 339-347
5. Bruza, P. D., Van der Weide, T. P.: Stratified Hypermedia Structures for Information
Disclosure. The Computer Journal, Vol. 35 (3) (1992) 208-220
6. Brusilovsky, P., Schwarz, E., Weber, G.: A Tool for Developing Adaptive Electronic Text-
books on WWW, WebNet’96
7. Buzan, T., Buzan B.: The Mind Map Book: How to Use Radiant Thinking to Maximize
Your Brain's Untapped Potential. New York: Plume (1996)
8. Carro, R. M., Pulido, E., Rodriguez, P.: Designing Adaptive Web-based Courses with
TANGOW, ICCE’99, v. 2, 697-704
9. Cristea, A., Okamoto, T.: MyEnglishTeacher–A WWW System for Academic English
Teaching. ICCE’00, Taiwan (2000)
10. Cristea, A. I., Okamoto, T.: Object-oriented Collaborative Course Authoring Environment
supported by Concept Mapping in MyEnglishTeacher, Edu. Tech. & Society 4(2) (2001)
11. De Bra, P., Brusilovsky, P., Houben, G.-J.: Adaptive Hypermedia: From Systems to
Framework. ACM Computing Surveys, http://wwwis.win.tue.nl/~debra/public.html (1999)
12. Dicheva, D., Aroyo, L.: An approach to intelligent information handling in Web-based
learning environments. In: Arabnia, H. R. (ed.), Proc. of ICAI’00 Conf. (2000) 327-333
13. Fink, J., Kobsa, A., Schreck, J.: Personalized Hypermedia Information Provision through
Adaptive and Adaptable System Features. Mullery, A. et al (eds.)Springer (1997) 456-467
14. Höök, K., Karlgren, J., Waern, A., Dahlbäck, N., Jansson, C-G., Karlgren, K., Lemaire, B.:
A Glass Box Approach to Adaptive Hypermedia, Journal of UMUAI, 6 (1996) 157-184
15. IEEE P1484.6 Course Sequencing Working Group: http://ltsc.ieee.org/wg6/index.html
16. IEEE P1484.2 Learner Model Working Group: http://ltsc.ieee.org/wg2/index.html]
17. Murray, T. et al.: MetaLinks - A Framework and Authoring Tool for Adaptive Hyperme-
dia. S. Lajoie and M. Vivet (eds.): AIED‘99, (1999) 744-746
18. Okamoto, T., Cristea, A., Kayama, M.: Future integrated LE with Multimedia. JCAI, Ad-
vanced information technologies for learning in the APR. Vol. 17 (1) (2001) 4-12
19. Opperman, R., Rashev, R., Kinshuk: Adaptability and Adaptivity in Learning Systems,
Knowledge Transfer (Volume II) (Ed. A. Behrooz), 1997, pAce, London, UK, pp173-179
20. SMIL, W3C standard, http://www.w3.org/AudioVideo/
21. W3C Recommendation: Authoring Tool Accessibility Guidelines 1.0 (3 February 2000):
http://www.w3.org/TR/ATAG10/
22. Wu, H., De Bra, P.: Sufficient Conditions for Well-behaved Adaptive Hypermedia Sys-
tems. Proc. of WI Conf. Lecture Notes in AI. Vol. 2198. Springer-Verlag. (2001) 148-152
Hypermedia Presentation Adaptation
on the Semantic Web
Flavius Frasincar and Geert-Jan Houben

PO Box 513, NL-5600 MB Eindhoven, The Netherlands
{flaviusf, houben}@win.tue.nl
Abstract. Web Information Systems (WIS) present up-to-date infor-

mation on the Web based on data coming from heterogeneous sources.
In previous work the Hera methodology was developed to support the
design of a WIS. In this paper we target the design of an intelligent
WIS. For this reason the Hera methodology is extended with two kinds
of hypermedia presentation adaptation: adaptability based on a profile
storing device capabilities and user preferences, and adaptivity based on
a user model storing the user browsing history. While adaptability is
considered to be static, i.e. the presentation is fixed before the brows-
ing starts, adaptivity is dynamic, i.e. the presentation changes while the
user is browsing it. The models used in Hera and their adaptation aspects
are specified in RDF(S), a flexible Web metadata language designed to
support the Semantic Web.
1 Introduction
The Web is the most rapidly growing information source. As huge amounts of
data are today stored in the “deep web” (searchable databases), there is an
increasing need to automate the presentation of this data. Designed originally
for human consumption, the Web is nowadays augmented to target machines.
In the Semantic Web [1] era, Web data will evolve from machine readable to
machine understandable, i.e. it will have associated semantics described by its
metadata.
The Web can be accessed through a number of different devices (PC, Laptop,
WebTV, PDA, WAP phone, WAP watch etc.) each having its own capabilities
(display size, memory size, network speed etc.). At the same time, the user
preferences (desired layout, navigation patterns etc.) and browsing history can
be taken into account during the presentation generation.
Web Information Systems (WIS) [2] offer Web presentations of data typi-
cally coming from heterogeneous sources (relational databases, object-oriented
databases, XML repositories, WWW etc.). In order to generate an appropriate
hypermedia presentation (hyperdocument), the presentation needs to be tailored
to specific device capabilities and user preferences.
The Hera methodology [10,11] supports the design of a WIS. It distinguishes
three important design steps: conceptual design that produces the conceptual

134 F. Frasincar and G.-J. Houben
model of the integrated data, application design that focuses on the naviga-
tional or logical aspects of the hypermedia application, and presentation design
that gives an abstraction of the physical level of the application. The heart of
Hera is the Application Model, a model inspired by Relationship Management
Methodology (RMM) [12,13]. In previous work [9] we built a prototype using
the Hera methodology based on XML.
This paper extends Hera by considering the adaptation of the presenta-
tion with respect to devices capabilities and user preferences stored in a profile
(adaptability). Moreover, we target also the automatic generation of adaptive
presentations based on user browsing history stored in a user model (adaptiv-
ity). In our methodology the different models lead to a lot of metadata that
describe different aspects of the application. Semantic Web technology appears
to be a natural solution to represent this metadata. As there is not yet a W3C
recommendation for a semantic markup language (only a note on DAML+OIL
[5]) we base our future prototype on RDF(S) [3,15].
The rest of the paper is structured as follows. In Sect. 2 we introduce the Hera
methodology and discuss its individual design activities. The artifacts produced
by Hera activities are: Conceptual Model, Application Model, and Application
Model with Adaptation, presented in Sects. 3, 4, and 5, respectively. Section 5
distinguishes two kinds of adaptations: adaptability described in Subsect. 5.1
and adaptivity described in Subsect. 5.2. Section 6 concludes the paper.
2 Hera Methodology
The Hera methodology is a model-based Web Engineering [16] method for design-
ing WIS. Figure 1 depicts the four different activities of the proposed method:
Conceptual Design, Application Design, Adaptation Design, and Presentation
Design. The newly introduced activity Adaptation Design is further decomposed
in two sub-activities: Adaptability Design and Adaptivity Design.
Presentation Design
Adaptation Design
Adaptativity Design
Adaptability Design
Application Design
Conceptual Design
Fig. 1. Hera Methodology

Hypermedia Presentation Adaptation on the Semantic Web 135
Each activity has specific design concerns and produces a model which is an
enrichment of the model built by the previous activity. Hera models are repre-
sented in RDFS [3], the schema language for RDF [15]. There are several rea-
sons that motivated us to choose RDF(S). RDFS offers the subclass/subproperty
mechanisms useful for building taxonomies for classes/properties. As RDFS is
expressed in RDF, it has all the benefits of property-centric models like extensi-
bility and sharability. Extensibility enables the building of each model on top of
the previous one and sharability fosters re-use of the developed models. There
exist already Composite Capability/Preference Profiles (CC/PP) [14] vocabu-
laries (in RDF(S)) for modeling device capabilities and user preferences which
ease the burden of defining new ones.
Conceptual Design provides a common representation for the schema of the
retrieved data. A WIS gathers data from multiple sources each having its own
dialect. In order to be able to further process this data one needs to define a
uniform semantics for it. This semantics is captured in the Conceptual Model
(CM) as an application specific ontology. The basic elements in the CM are
concepts and concept relationships. Concepts have properties to describe their
features.
Application Design is concerned with the navigational aspects involved in the
hypermedia presentation of the retrieved data: the structure of the hyperdoc-
ument. It extends the CM with navigational views that build the Application
Model (AM). The basic elements in AM are slices and slice relationships. Slices
are units of presentation for data contained in one or more concepts from CM.
Adaptation Design adds adaptation features to the previously defined AM.
We distinguish two kinds of adaptation: adaptability and adaptivity. Both condi-
tion the appearance of slices and the visibility of slice relationships. Adaptability
does it based on information about device capabilities and user preferences prior
to browsing. Adaptivity uses information about the user browsing history stored
in a User Model (UM) during the browsing of the presentation. Adaptability is
considered to be static, i.e. the presentation is fixed before browsing starts, while
adaptivity is dynamic, i.e. the presentation changes while the user is browsing
it.
Presentation Design takes into consideration the physical aspects of the pre-
sentation. In the Presentation Model (PM) we define slice appearance in terms
of regions [10]. The basic elements in PM are regions and region relationships.
Regions are rectangular shaped areas that present some data from one or more
slices. Slice relationships are materialized by navigational, spatial, or temporal
region relationships which can be synchronized. PM is outside the scope of this
paper, nevertheless we acknowledge the need of extending the adaptation aspects
also to the PM (e.g. font colour, page layout etc.).
3 Conceptual Model
The Conceptual Model (CM) presents a uniform view of the domain semantics
for the input data sources. It is an application specific ontology that will consti-
tute the basis for the subsequent data transformations. The retrieved data is so-
called “instance data”, as it represents specific instances of the concepts defined
in the ontology. CM specifies in a hierarchical manner the concepts (classes) in
the domain and the relationships among them. These relationships are expressed
as class properties (slots). One can associate logic to this ontology that will en-
able reasoning about CM, e.g. the transitivity of some properties enables the
derivation of new properties which can be used at a later stage in our design
process.
For CM serialization we were inspired by the RDF(S) representation of on-
tologies built with Protege-2000 [17]. Protege-2000 is a graphical tool intended
for defining ontologies without having to know any ontology language. This
graphical representation can be saved in a Protege-extended RDF(S). Figure 2
presents an example of CM that corresponds to a subset of the Rijksmuseum
catalogue in Amsterdam.
String Integer
String name name year name String
exemplified_by created_by
Technique Artifact Creator
exemplifies creates
description biography
String String
painted_by
Painting Painter
Property paints
picture
subClassOf
subPropertyOf Image
Fig. 2. Conceptual Model
As RDF(S) has its own modeling limitations, we added two new properties
to characterize RDF(S) properties that represent relationships: the cardinality
(single or multiple) of a relationship and the inverse of a relationship. We also
defined a hierarchy of Media classes that have specific properties to be used in
the Adaptation Design phase. Example 1 shows an instance of the Image class:
the image dimensions (pix_x and pix_y) can be considered in adapting the
presentation to the display size.
Example 1. Media Type
<Image about="http://www.example.com/sunset.jpg"
pix_x=326
pix_y=230
...
</Image>
4 Application Model
The Application Model (AM) describes the navigational view over CM. The AM
is the most abstract form of the presentation. We define meaningful presentation
units called slices and relationships among them. The simple slices contain only
a concept attribute (concept property that points to a media item). Complex
slices are defined in a tree-structure manner having simple slices as leaves [12,13].
We distinguish two kinds of slice relationships: aggregation and reference [9].
Reference relationships are also called links as usually they are materialized to
hyperlinks (navigational relationships in the Presentation Model). Each slice
belongs to a certain concept from CM. While the reference relationships do
not leave the context of a certain concept, aggregation relationships can link
slices belonging to different concepts. The designer needs to carefully specify the
relationships from the CM, which make such a slice embedding possible. For
relationships having cardinality single-many the access structure Set (of slices)
is used.
Figure 3 describes a part of the AM for the Rijksmuseum example, i.e. two
complex slices Slice.technique.main and Slice.painting.main.
technique painting
picture
name
name
description year
painting painted_by
picture
exemplified_by painter
Set name
main main
Fig. 3. Application Model
Slice.technique.main has two attributes from the concept technique, i.e.

name and description and an attribute from painting, i.e. picture. Since the
relationship exemplified_by (inherited from CM) has cardinality “single-many”
one needs to add an access structure, e.g. Set. Note that the previous attributes
are slices on their own, i.e. simple slices. Each picture has a reference (link) to
the Slice.painting.main that describes the current painting. This description
includes the picture, name, year, and the name of the painter the painting was
painted_by. We chose a very simple AM example but rich enough to show, in
the next section, different kinds of presentation adaptation.
5 Adaptation in the Application Model

A WIS can be accessed through a multitude of devices and by different users.
Each device has its own capabilities (display size, memory size, network speed
etc.). Every user has specific preferences (desired layout, navigation patterns
etc.) and browsing history with respect to a particular presentation. An intel-
ligent WIS needs to take into account these constraints (abilities) coming from
both devices and users, and adapt the presentation accordingly.
The adaptation we consider in this paper is based on conditioning the ap-
pearance of slices in the AM (and derived, the visibility of slice relationships or
links). Figure 4 shows two examples of slice conditioning, one for adaptability
and one for adaptivity. Both examples will be discussed at the level of RDF(S)
in the following two subsections.
um:Painter > 10 prf:ImageCapable = Yes
technique painting
picture
name
name
description year
painting painted_by
picture
exemplified_by painter
Set name
main main
Fig. 4. Adaptation in the Application Model
Example 2 illustrates the definition of the condition for slice appearance in

RDFS. A link pointing to a slice that has the appearance condition unfulfilled
will be hidden [7].
Example 2. Slice Condition
<rdf:Property rdf:ID="condition">
<rdfs:domain rdf:resource="#Slice"/>
<rdfs:range rdf:resource="#Literal"/>
</rdf:Property>
We consider two kinds of presentation adaptation: adaptability and adaptivity.
Adaptability means conditioning the appearance of slices and the visibility of
slice relationships based on device capabilities and user preferences stored in
a Profile. Adaptivity takes into account the user browsing history stored in a
User Model (UM) to condition the appearance of slices and links during the
browsing of the presentation. While we consider adaptability to be static, i.e.
the presentation is fixed prior to browsing, we consider adaptivity to be dynamic,
i.e. the presentation changes while the user is browsing it.
Two existing techniques fit well in the Hera methodology to model the two
kinds of adaptation mentioned above. For adaptability the Composite Capabil-
ity/Preference Profile (CC/PP) [14] offers a framework to model profiles that
characterize device capabilities and user preferences. The AHA (Adaptive Hy-
permedia Architecture) system [6] adds adaptivity functionality to a hypermedia
presentation based on UM.
5.1 Adaptability
Adaptability is the adaptation that considers the device capabilities and user
preferences stored in a Profile. A Profile contains attribute-value pairs used by
the Hera system to determine the most appropriate presentation for the retrieved
data items. Example 3 shows how to build a Profile using two vocabularies, the
CC/PP [14] UAProf (User Agent Profile) vocabulary [18] developed by WAP
Forum to model device capabilities, and our own vocabulary for describing user
preferences.
Example 3. Device/User Profile
<Description rdf:about="Profile">
<ccpp:component>
<prf:HardwarePlatform>
<prf:ImageCapable>No</prf:ImageCapable>
<prf:ScreenSize>600x400</prf:ScreenSize>
...
</prf:HardwarePlatform>
</ccpp:component>
<ccpp:component>
<up:UserPreferences>
<up:Language>English</up:Language>
...
</up:UserPreferences>
</ccpp:component>
</Description>
A Profile has a number of components, each component grouping a number
of attributes. In the previous example we defined two CC/PP components. The
HardwarePlatform component has two attributes, ImageCapable that specifies
if the device is able to display images and ScreenSize that defines the dimen-
sions of the device display. The UserPreferences component has one attribute
Language, the language the user prefers. Example 4 illustrates an adaptability
condition that models the presence of Slice.painting.picture in AM based
on the ability of the device to display images.
Example 4. Adaptability Condition

<rdfs:Class rdf:ID="Slice.painting.picture"
slice.condition="prf:ImageCapable=Yes">
<rdfs:subClassOf rdf:resource="#Slice"/>
</rdfs:Class>
In an adaptability condition, one can use the media properties defined in
Example 1, e.g. one can test if the dimensions of a particular image fit the size
of the screen. Note that the adaptation based on user preferences can be treated
in the same way as the adaptation based on device capabilities.
5.2 Adaptivity
Adaptivity [4] is the dynamic adaptation that considers the user browsing his-
tory stored in a User Model (UM). AHAM (Adaptive Hypermedia Application
Model) [8], a Dexter-based reference model for adaptive hypermedia, defines in
the Storage Layer three models: the Domain Model, the Teaching Model later
on renamed Adaptation Model [19], and the User Model (UM). We use the
AHA system to add adaptivity functionality to our methodology. In Hera, the
Domain Model will have as atomic concepts slice instances, which stand for
both AHAM pages and fragments, and as composite concepts class instances
from CM and additional concepts introduced by the application author. The
UM contains attribute-value pairs: for each concept from DM a value (from 0
to 100) is associated. The Adaptation Model contains adaptation rules based
on the Event-Condition-Action paradigm (like in active databases) that can be
executed by the software from the AHA engine.
In Hera when the user visits a certain (page) slice, update rules get triggered
that in the end determine the appearance. A slice is desirable if its appearance
condition evaluates to true. The desirability of a slice included in another slice
implies its appearance in the presentation. Standard AHA functionality implies
that a link pointing to a desirable page is displayed in “good” colour (“blue” for
active) if it was not visited before, or “neutral” colour (“purple” for visited link)
if it was visited before, while “bad” colour (“black” for hidden link) is used if
the link points to an undesirable page.
The appearance condition in adaptivity for the slice Slice.painting.main,
illustrated in Example 5, models the desirability of the slice based on the interest
of the user for the painting’s painter.
Example 5. Adaptivity Condition
<rdfs:Class rdf:ID="Slice.painting.main"
slice.condition="um:Painter > 10">
<rdfs:subClassOf rdf:resource="#Slice"/>
</rdfs:Class>
Initially, all concepts from UM have their values set to 0. Example 6 shows
an update rule expressed in RDF.
Example 6. Update Rule
<um:Slice.painting.main>
<aha:updatelist>
<aha:SetOfConcepts>
<aha:item><um:Painting aha:update="+80"/></aha:item>
<aha:item><um:Painter aha:update="+40"/></aha:item>
</aha:SetOfConcepts>
</aha:updatelist>
</um:Slice.painting.main>
Suppose that when seeing the slice Slice.technique.main, the user is in-
terested in a particular painting description and clicks its link, which for the
moment is a link to an undesirable slice (condition not fulfilled), as are all the
Slice.painting.main instances. Before this slice Slice.painting.main is ac-
tually visited, the corresponding update rule is triggered. First, the value of
the particular instance of Slice.painting.main (i.e. associated to the chosen
painting) is updated to 35 (default update for a undesirable slice). Then, the
values of the Painting and Painter instances from DM (note that these concepts
were present also in the CM) corresponding to this particular slice instance have
their values updated by 80% and respectively 40% of the slice update increment.
As opposed to AHA which specifies update rules for instances, in Hera we
specify update rules based on classes (rules are defined at schema level, because
the particular instances are not known beforehand). Nevertheless, the Hera up-
date rules will become instance update rules (as in AHA) at runtime.
When after visiting the slice describing a particular painting, the user would
go back to Slice.technique.main, the user would observe that all the paintings
related to the painter of the previously chosen painting are now active links
pointing to desirable slices (the condition is fulfilled since the Painter value was
updated; 40% × 35 = 14 > 10).
6 Conclusion and Further Work
The Hera methodology has been extended with adaptation models (Profile and
User Model) that condition the slice appearance in the Application Model. Dur-
ing the different activities that compose our methodology there are a lot of on-
tologies involved: Conceptual Model, Domain Model, Profile etc. RDF(S) proves
to be a flexible language for modeling, but as we saw, it has also its own short-
comings (we added RDF(S) extensions to cope with them). The Semantic Web
promises a richer Web where data will have semantics associated to it. This se-
mantics can be fully exploited in an adapted automatic generation of hyperme-
dia presentations. In the future we plan to include these new emerging Semantic
Web standards like DAML+OIL in the modeling of the different ontologies in-
volved in our processes. A prototype based on RDF(S) that experiments with
the proposed Hera methodology is under development.
References
1. Berners-Lee, T.: Weaving the Web. Orion Business, London (1999)
2. Bieber, M., Isakowitz, T., Vitali, F.: Web Information Systems. Communications
of the ACM, 41(7) (1998), 78–80
3. Brickley, D., Guha, R.V.: Resource Description Framework (RDF) Schema Speci-
fication 1.0. W3C (2000), http://www.w3.org/TR/rdf-schema
4. Brusilovsky, P.: Adaptive Hypermedia. User Modeling and User-Adapted Interac-
tion 11(1/2), Kluwer Academic Publishers (2001), 87–110
5. Connolly, D., van Harmelen, F., Horrocks, J., McGuinness, D.L., Patel-Schneider,
P.F., Stein, L.A.: DAML+OIL (March 2001) Reference Description. W3C (2001),
http://www.w3.org/TR/daml+oil-reference
6. De Bra, P., Aerts, A., Houben, G.J., Wu, H.: Making General-Purpose Adaptive
Hypermedia Work. In Proc. WebNet 2000 World Conference on the WWW and
Internet, AACE (2000), 117–123
7. De Bra, P., Brusilovsky, P., Houben, G.J.: Adaptive Hypermedia: From Systems
to Frameworks. ACM Computing Surveys, 31(4es) (1999)
8. De Bra, P., Houben, G.J., Wu, H.: AHAM: A Dexter-based Reference Model for
Adaptive Hypermedia. In Proc. The 10th ACM Conference on Hypertext and
Hypermedia, ACM Press (1999), 147–156
9. Frasincar, F., Houben, G.J.: XML-Based Automatic Web Presentation Genera-
tion. In Proc. WebNet 2001 World Conference on the WWW and Internet, AACE
(2001), 372–377
10. Frasincar, F., Houben G.J., Vdovjak, R.: An RMM-Based Methodology for Hy-
permedia Presentation Design. In Proc. Advances in Databases and Information
Systems, LNCS 2151, Springer (2001), 323–337
11. Houben, G.J.: Hera: Automatically Generating Hypermedia Front Ends for Ad
Hoc Data from Heterogeneous and Legacy Information Systems. In Proc. Third
International Workshop on Engineering Federated Information Systems, Aka and
IOS Press (2000), 81–88
12. Isakowitz, T., Kamis, A., Koufaris, M.: Extending RMM: Russian Dolls and Hyper-
text. In Proc. 30th Hawaii International Conference on System Sciences, Computer
Society Press (1997)
13. Isakowitz, T., Stohr, E., Balasubramanian, P.: RMM: A Methodology for Struc-
tured Hypermedia Design. Communications of the ACM, 38(8) (1995), 34–44
14. Klyne, G., Reynolds, F., Woodrow, C., Ohto, H.: Composite Capabil-
ity/Preference Profiles (CC/PP): Structure and Vocabularies. W3C (2001),
http://www.w3.org/TR/CCPP-struct-vocab
15. Lassila, O., Swick, R.R.: Resource Description Framework (RDF) Model and Syn-
tax Specification. W3C (1999), http://www.w3.org/TR/REC-rdf-syntax
16. Murugesan, S., Deshpande, Y., Hansen, S., Ginige, A.: Web Engineering: A New
Discipline for Web-Based System Development. In Proc. First ICSE Workshop on
Web Engineering, ACM Press (1999), 1–9
17. Noy, N.F., Sintek, M., Decker, S., Crubezy, M., Fergerson, R.W., Musen, M.A.:
Creating Semantic Web Contents with Protege-2000. IEEE Intelligent Systems,
16(2) (2001), 60–71
18. Wireless Application Group: User Agent Profile Specification. WAP Forum (1999),
http://www.wapforum.org/what/technical/SPEC-UAProf-19991110.pdf
19. Wu, H., De Bra, P., Aerts, A., Houben, G.J.: Adaptation Control in Adaptive
Hypermedia Systems. In Proc. International Conference on Adaptive Hypermedia
and Adaptive Web-based Systems, LNCS 1892, Springer (2000), 250–259
User Data Management and Usage Model
Acquisition in an Adaptive Educational
Collaborative Environment
Elena Gaudioso and Jesus G. Boticario
Dpto. de Inteligencia Artificial

Universidad Nacional de Educacion a Distancia
Senda del Rey, 9; Madrid, Spain
{elena,jgb}@dia.uned.es
Abstract. In this paper we describe the basis of aLF (active learning

framework ) an environment for web-based distance learning on the web
and how it can be extended with a user modelling subsystem aimed at
providing an adaptive response which can meet users’ knowledge state,
preferences and goals. We concentrate on user data collection and its
further processing to create a user model. Furthermore, as regards the
user modelling task, we claim that a combination of different machine
learning algorithms should perform better than individual methods. To
support our approach we will provide some preliminary experimental
results.
1 Introduction
Distance education takes place when a lecturer and student(s) are separated by
physical distance. The proliferation of accesses to the Internet has made the
World Wide Web (WWW) an ideal environment for lecturer-student commu-
nication with no restrictions of time and space. When a web site is designed
to support certain courses it is called an educational web site. Educational web
sites reduce student isolation by offering a number of communication channels,
usually newsgroups, e-mail, mailing lists and chats. They may also provide the
necessary course material that can be downloaded by the students, offer HTML
pages that guide students through the course contents or provide references to
other information sources, as well as many other educational possibilities.
Often the content provided by an educational web site is fully static. The
HTML page navigation sequence, the set of references, and the newsgroup mes-
sages are the same for every student on the course. This homogeneous response
is provided to students with changing needs and whose level of experience on
the subject being taught and on the use of resources available on the website are
different. Furthermore, adaptation is essential in distance education because stu-
dents are isolated and have a wide variety of backgrounds and interests. Therefore
an environment that adapts to each individual student appears to be desirable.
In order to solve this problem, the web-based adaptive educational systems [3]

144 E. Gaudioso and J.G. Boticario
carry out an adaptation based on a user model, representing the user’s knowl-
edge state, preferences and goals. These are not entirely new kinds of systems,
they borrow heavily from adaptive hypermedia systems.
Adaptive hypermedia systems [4] apply different forms of user models to
adapt the contents and the hypermedia page links to the user. The main goal
of this adaptation is to provide the user with efficient access to the site by first
presenting the links that should be of interest. Adaptive hypermedia application
in educational systems is a very hot research area. According with [4], many
interesting adaptive educational hypermedia systems have been developed and
reported during the last eight years. An interest to provide distance education
over the Web has been a strong driving force behind these research efforts.
The introduction of the Web has impacted both the number and the type of
systems being developed. All the early system were essentially lab systems, built
to explore some new methods that used adaptivity in an education context. In
contrast, several more recent systems provide complete frameworks and even
authoring tools for developing web-based courses.
Traditionally adaptation decision in adaptive systems was based on taking
into accounts various characteristics of their users represented in the user model.
Currently the situation is different. A number of adaptive web based systems are
able to adapt to something else than user characteristics. Kobsa et al. [7] suggest
to distinguish adaption to user data, usage data, and environment data. User
data comprise the traditional adaptation target, various characteristics of the
users. Usage data comprise data about user interaction with the systems that
can not be resolved to user characteritics (but still can be used to make adapta-
tion decisions). Environment data comprise all aspects of the user environment
that are not related to the users themselves. To achieve a good adaptation it is
neccesary to collect correctly the data that will be used to construct the user
models.
To summarise, although distance learning benefits from the use of the educa-
tional web sites, it poses several challenges regarding the management of student
collaboration and the adaptation of the contents to student needs. Within this
context, an interesting and relatively new area is the adaptation to students
with regard to their activities. In this paper we will describe aLF (active learn-
ing framework), a system for web-based distance learning developed to support
the requirements for collaborative work between all the different students in a
course delivered in aLF.
We aim to describe how to provide advice to students in their collaborative
work with fellow students. For this we construct a usage model. We will describe
the user data collection and its further procesing to form a user model and we
also suggest that, as regards the user modelling task, a combination of differ-
ent algorithms should perform better than individual methods, providing some
experimental support.
Usage Model in an Adaptive Educational Collaborative System 145
Fig. 1. aLF course workgroup
2 An Environment for Web-Based Distance Learning

2.1 Overview
Our research is developed in a distance-learning university (UNED1 ) with over
180,000 students and a traditional distance-education model. In this context
there is a set of educational requirements and needs to be fulfilled that are
different for each student. Distance-education tools have been shown to be useful
in such situations and, particularly, web-based adaptive educational systems
seem to be one of the best solutions.
A platform called aLF (active learning framework ) has been developed in or-
der to support the requirements for collaborative work between all the different
communities involved. The platform builds upon the ArsDigita Community Sys-
tem 2 , a multiplatform and open source set of tools for constructing web-based
applications. aLF is composed of a Web server connected to a relational database
and a set of TCL (Tool Command Language) scripts allowing management of
the interaction with the data model. The database does not only store the user
personal information and the contents they send, but also serves to structure
potentially everything that is happening in the site.
Another key feature of aLF is that it offers different views based on user
navigation and actions through the site. All kinds of information on users can
thus be gathered with TCL scripting language. The scripts can interface with the
1
http://www.uned.es
2
http://www.arsdigita.com/. Related web sites built with ArsDigita
http://www.arsdigita.com/pages/customers
database so that this information can be easily stored and retrieved on demand.
TCL scripts can also contribute to maintaining a dynamic environment, since
they enable web pages to be dynamically constructed and this process can make
use of the database information if required. Users registered on aLF are grouped
into workgroups (see Fig. 1). The administration of each workgroup is done by
the person in charge who may not be the web site manager. The setting up of
these kind of workgroups is particularly useful for distance learning since it allows
different university departments, courses, study groups and research groups to
be managed.
aLF workgroups offer several services thus allowing a learning and collabo-
rative environment to be easily set up, such us forums, chat rooms or shared file
storage area. A complete description of the services that offers aLF can be found
at [1].
3 An aLF Course Experience
We will now describe an experience of a distance course being taught through

aLF. The goal of this course was precisely to teach the use of the Internet in
education. We set up a group in aLF for the students on the course; this group
had a bulletin board, news service, chat, file-storage area and a personal web-
pages storage area where the students could publish their own web pages.
Regarding the conceptual organization of the contents, it was based on the
proposal made by Roger Schank in the ASK system [10]. The course concepts
were organized as a conceptual network where each node was a course concept for
the student to learn and the arcs represented certain links between concepts such
as, prerrequisite, consequence, example or exercise. These links were presented by
several questions or options that the student could choose . The course contents
were represented by a set of HTML pages, each one corresponding to one concept
in the course conceptual network. This conceptual network, the goals of the
course, and other elements were represented by an XML page that the system
may interpret to construct the student knowledge model. For each student on
the course, the system stored a model of the contents that the student had
already learned. For this course the system assumed that a student had learned
a concept if he/she had successfully done a test or had simply visited the pages
related to the concept. The system guided the student through the contents of
the course according his/her conceptual model; in this paper we are not going
to describe it further since it is described elsewhere [2].
First we presented the students with a survey in order to evaluate how fa-
miliar they were with educational software or the Internet. In most cases the
students were in turn lecturers interested in the use of the Internet in education,
and on the whole, they had little experience in the use of computers and Inter-
net services. However, as the course proceeded we realised that we had a very
hetereogenous group of students.
Consequently, the course lecturers had to closely monitor the students on
the course. The lecturers solved (on demand) the difficulties that arose. They
proposed several activities (such as sending a message to the bulletin board) in

order to guide the students more closely.
Although there was also the possibility of contacting the lecturers by phone
or face-to-face, the main channels of communication were the course bulletin
boards and electronic mail. The lecturer also visited the chat room established
for the course to detect if the students had difficulties in the course.
We have also discovered that students do not collaborate among themselves
except when they use the chats and send personal e-mails. If this lack of col-
laboration could have been foreseen, lecturers could have suggested scenarios or
tasks where students were encouraged to interact.
These problems could have been avoided if the system had detected student
difficulties, such as problems in the use of the course services and student sim-
ilarities. Then lecturers could have been informed of these problems right from
the very beginning and they could have set up subgroups of students with simi-
lar characteristics, needs and interests. The system could also automatically do
this.
Thus our proposal for a personalized environment is based on a cooperative
tutorization scheme. In this scheme the lecturer, besides solving the doubts of
the students in a personalized way, is responsible for improving collaboration
between the group components.
As we have seen before the lecturers proposed several activities through the
forums of the course. So it is very important to monitor the student correctly to
control the user performance in the course, thus we have to acquire some kind
of usage model of the user to describe his/her user interaction with the forums
and with any other service in the web site. All this proccess will be describe in
the next section.
4 Adaptation and User Modelling in aLF

4.1 Usage Model in aLF
At first, the main problem within the web-based adaptive systems were the
user data collection. However, nowdays, more and more systems are capable to
monitor the user actions to record the complete user interactions [8,11].
It is important to point out that aLF manages all the services through the
database. Information about the messages sent to a forum or a bulletin board, the
tickets sent to the project management tool, the appointments in the calendar,
the annotations to a particular news in the board or the conversation in a chat,
are stored in the database, so a great amount of information on the interaction
of a user with the platform is available. Another key feature of aLF, is that all
this data is stored separately in tables of the database, and that differs from
traditional web logs in which we only usually store the urls accessed by the user.
We are now to describe the user features that we consider to construct the
user model in aLF. They can be divided in two fundamental groups, personal
data and usage data. In the first group we include typical user attributes such
us, user-id, first-names, last-name, screen-name, e-mail, url?, portait?,
Table 1. Usage features related to the user interaction with the system
Feature Description
n-sessions Number of sessions of the user
msg-to-other-forums Number of messages that the user may have sent to other
forums
msg-to-group-forum Number of messages that the user may have sent to
the forums of the group
email-alerts-out? Determines if the user has activated some alerts in forums
out of the course
email-alerts-aprfor? Determines if the user has activated some alerts in the
forum of the course
wp-presentation? Determines if the user has constructed a presentation
bm? Determines if the user has bookmarks in aLF
num-msgs-chats-aprfor Number of messages that the user have sent to the chat
of the course
num-msgs-chats-out Number of messages that the user have sent to chats out
of the course
num-entrances-chat-aprfor Number of entrances to the chat of the course
num-entrances-chat-out Number of entrances to chats out of the course
files-personal-storage-area Number of files in the user personal file storage area
files-group-storage-area Number of files in the file storage area of the group
groups-other? Number of groups that the user is registered
num-news-sent Number of news that the user has sent
num-static-pages-visited Number of pages that the user has visited
average-session-duration Duration of the user’s sessions duration
spam? Determines if the user has sent a email to the whole
group or to the teachers
bio?, and several attributes describing certain users preferences (obtained from
an initial survey) such us do-you-have-interest-in-teleeducation, do-you-
have-interest-in-collaborative-environments,. . . and users background
(obtained also from an initial survey) such us computer-use-skill-level,
internet-use-skill-level, knowledge-about-computers, use-of-internet
-resources.
In the second group we include all those attributes that allows us to deter-
mine certain characteristics in the user interaction with the system and they are
described in Table 1.
However, as we have seen before, the course activities are controlled through
the forums of the course, so it is neccessary a closer control in the user messages
to the forum (all these last features allow us to determine the usefullness of the
contributions of the user to the forum.). So we consider the following features:
– num-threads-user-started: Number of threads that the user has began.

– num-threads-started-and-not-replied: Number of threads that the user
has began and his partners have not replied.
– num-threads-user-in-and-tutor: Number of threads in which the user
and the tutor have participated.
– num-threads-user-finished: Number of threads that the user has finished

– num-replies-to-user-messages: Number of replies to user messages
– fellows-replayed-user-messages: List of users that have replayed a par-
ticular user messages
With all this information we can determine some user characteristics that
are not directly observable from the data gathered by aLF. This features will
be the result of some machine learning tasks and we will be described in next
section.
4.2 Learning Some Features in the User Model

As we have seen before, there are some features of the user model that are not
directly observable from the data in the database. These features are very useful
in order to give the user some recommendations about his collaboration within
the course.
Currently the recommendations of the system are very simple. We focus in
determining the level of activity of the user in the course and the possibility
that a particular user may fail the course. With these features, the system may
recommend a user to participate more in the course, or may alert the teacher
that a user is going to fail.
Each of these attributes constitutes a separate machine learning task. The
particular data used in learning the value of these two attributes is taken from
user interaction with the web site.
First the goal is to learn a set of user profiles in order to infer the level of
activity that the user is going to exhibit. We measure this activity taking into
account a certain period of time and from the data described in Table 1. A set of
three categories is defined corresponding to the low, medium and high activity
profiles.
Extracting and preparing data from the aLF database, we composed a train-
ing set of instances as input data to the learning algorithm. Each instance cor-
responds to a user and includes information about his behavior in his first one
or two interactions (some of the attributes have been described earlier). As re-
sult of this process, we obtained a dataset of 125 instances described each by
21 attributes and labeled with a class value indicating the level of activity in
the forum of the course (this training data were constructed taking account the
labeled data constructed by external tutors of the course).
However, in a web environment there are a great variety of interaction data.
Single classifiers would fail in some regions of the training data or would be ac-
curate in others (competence region). If we have certain classifiers with distint
competence regions, we can combine them in order to achieve a better perfor-
mance, such a combination is usually called ensemble of classifiers.
So for our experiments we consider three classifiers; a decision tree learner
(C4.5), a rule set learner (C4.5Rules) and a Bayesian learner (Naive Bayes).
To asses the individual performance of each of the algorithms, we performed a
10-fold cross validation test. This procedure divides the available data into 10
Table 2. Error rate from 10-fold cross validation for each learning algorithm with the
standard deviations
Algorithm Accuracy
C4.5 13.46 (6.02)
C4.5Rules 24 (5.58)
Naive Bayes 42.37 (8.24)
Best Classifier 7.34 (5.08)
subsets of instances and successively takes one different subset as the test set,
using the remaining instances in the training set. In this manner, the algorithm
is ran 10 times learning with a subset of the 90% of instances and tested with the
remaining -unseen- 10%. The average of accuracies can be taken as an estimate
of the accuracy of the system for the dataset.
Table 2 shows the final results for each of the algorithms together with an
estimate of the upper bound that we could obtain if we pick up the best algorithm
for each testing instance. To compute this accuracy, we simply recorded a correct
prediction for all the instances for which at least one of the algorithms made a
correct guess.
As we can observe in the Table 2, the individual performance of any of the
algorithms is not outstanding. However, the room for improvement by using a
combination of these algorithms is huge.
Taking into account that an ensemble of classifiers will perform better than a
single classifier, there are a lot of work done directed to improve the accuracy of
certain learning tasks [5]. In our case, the user model construction implies several
learning tasks with very heterogenous data, we need some kind of mechanism to
improve the learning accuracy of the user modelling subsystem and the flexibility
to achieve a good performance in possible new learning taks.
So, following [9] we are modifying the user model construction and we are
constructing an ensemble formed by several classifiers constructed by running
cross-validation with a particular training set and c4.5. For each classifier and
each testing set, we store the prediction made by the classifier for a particular
instance in the testing set, and the correct prediction for that instance. For
each classifiers we have a referee that predicts with the set described before
(using c4.5 also) if its corresponding classifier will make a correct prediction.
We combine the solution given by the different classifiers that is supposed to be
correct (acording to their referes’ prediction), by simply mayority voting. The
results of our preliminary experiments are shown in Table 3.
Table 3. Error rate of the c4.5 classifier and the corresponding ensembles
Algorithm Error rate

Ensemble with 5 classifiers 12.11 (8.34)
At the moment we have only focused in choosing the best response for a
particular learning task. To learn the best algorithm for each of the learning task
involved in the user model construction is in turn another learning task. One
related work is [6], who present an approach to combine multiple learning agents
that constructs a preference table which is updated according the performance of
each individual agent. Although the results are not very relevant at the moment
because we have not yet enough interaction data, we are currently working in
this direction, together with using simpler approaches in order to get further
insight on this problem.
5 Conclusion and Future Work
aLF, the platform described in this paper, is especially intended to facilitate the
development of workgroups on the Web. What differentiates aLF from all other
initiatives is that it provides an integrated framework for all the Internet services
through HTML pages, storing all the information about the interactions of the
user in a database which allows a processing of that data to be done.
Although aLF is not only intended to serve for educational purposes, it is
specially useful in the establishment of on-line courses since it provides many
integrated services such us forums or file storage areas. However, supporting
distance education with educational web sites, does not guarantee that the stu-
dents will find the required information or will engage in effective collaborative
learning behavior. So, it is needed some kind of adaptation to help the users in
the variety of tasks he/she is expected to solve such as self-assessment exercises,
collaborative tasks, or practical exercises.
In order to access all the relevant data arising from each user’s interaction,
aLF draws up a complete register among the members of a group of users.
Although we have only discussed data preparation at a cursory level, this is not
a trivial problem as pointed out in related areas such as web mining [8]. Storing
information about the user behavior on a database should serve better to model
his preferences than, for example, simply using the navigation traces in web logs.
However, this information may be also more difficult to manage and prepare for
the user modeling stage.
Currently, the recommendations of the system are simple (tell the student
to visit a forum, or alert the teacher that a user seems to fail the course) but
it can be expanded given the information the system is gathering about the
student’s collaborative activities, such as, level of expertise in the use of the
services, preferences and interests. A possible scenario could be, if the student
is having trouble with topic X (perhaps as indicated by failing the online com-
petency test on topic X), the system might suggest contacting a student, by
email, who has shown competency and interest in topic X (as evidenced by this
latter student’s contributions to forums about topic X)). Also, the system could
encourage students to participate in chats/etc. about topic X. Or simply encour-
age the student to participate in activities other than forums - e.g., chats - when
there is evidence that the student has not done so. This advice can be useful
in having the student be part of the learning community and all its attendant
benefits.
We have shown that all these features are better infered from the past user
interaction applying machine learning techniques. In addition the complexity of
the learning tasks in these contexts lead us to explore more ellaborated solutions
based upon combining different learning techniques.
References
1. Jesus G. Boticario, Elena Gaudioso, and Carlos Catalina. Towards personalised
learning communities on the web. In Kai Hakkarainen Pierre Dillenbourg, An-
neke Eurolings, editor, European Perspectives on Computer-Supported Collab-
orative Learning. Proceedings of the First European Conference on Computer-
Supported Collaborative Learning, pages 115–122. Maastricht McLuhan Institute,
2001.
2. Jesus G. Boticario, Elena Gaudioso, and Felix Hernandez. Adaptive navigation
support and adaptive collaboration support in webdl. In Proceedings of the In-
ternational Conference on Adaptive Hypermedia and Adaptive Web-based Systems,
number 1892 in Lecture Notes in Computer Science (LNCS), pages 51–61, Trento,
Italy, August 2000. Springer Verlag.
3. Peter Brusilovsky. Adaptive educational systems on the world-wide-web: A review
of available technologies. In Proceedings of Workshop WWW-Based Tutoring at
Fourth International Conference on ITS (ITS’98), San Antonio, TX, August 1998.
MIT Press.
4. Peter Brusilovsky. Adaptive hypermedia. User Modelling and User Adapted Inter-
action, 11(1):87–110, 2001.
5. Joao Gama. Combining classification algorithms. PhD thesis, Facultad de Ciencias,
Universidad de Porto, Portugal, 1999.
6. J. Ignacio Giraldez, Charles Elkan, and Daniel Borrajo. A distributed solution to
the pte problem. In Giuseppina C. Gini and Alan R. Katritzky, editors, Predictive
Toxicology of Chemicals: Experiences and Impact of AI tools, Papers from the 1999
AAAI Spring Symposium, TR SS-99-01, pages 82–85. AAAI Press, March 1999.
7. A. Kobsa, J. Kownemann, and W. Pohl. Personalized hypermedia presentation
techniques for improving online customer relationships. Technical Report 66, Ger-
man National Research Cetner for iNformation Technology, St. Augustin, Ger-
many, 1999.
8. B. Mobasher, R. Cooley, and J. Srivastava. Web mining: Information and pat-
tern discobery on the world wide web. In Proceedings of the Ninth International
Conference on Tools with Artificial Intelligence (ICTAI97), 1997.
9. J. Ortega. Exploiting multiple existing models and learning algorithms, 1995.
10. Roger C. Schank and Chip Cleary. Engines for education. Lawrence Erlbaum
Associates, Hillsdale, New Jersey, 1995.
11. Gerhard Weber and Marcus Specht. User modeling and adaptive navigation sup-
port in www-based tutoring systems. In Proceedings of the Sixth International
Conference on User Modeling, pages 289–300, Chia Laguna, Sardinia, Italy, June
1997.
Personalizing Assessment
in Adaptive Educational Hypermedia Systems
Evangelia Gouli, Kyparisia Papanikolaou, and Maria Grigoriadou
Department of Informatics and Telecommunications, University of Athens,

Panepistimiopolis, GR-15784 Athens, Greece
{lilag, spap, gregor}@di.uoa.gr
Abstract. In this paper, we present a comprehensive framework for assessment,

developed through the web-based module named PASS-Personalized ASSess-
ment, which can be integrated in an Adaptive Educational Hypermedia System
to provide personalized assessment. PASS estimates learner’s performance
through multiple assessment options - pre-test, self-assessment and summative
assessment - tailored to learner’s responses. The adaptive functionality of
PASS, which is mainly based on the adaptive testing and the adaptive questions
techniques, is described. The first results from the formative evaluation of
PASS are encouraging, concerning the total number of questions posed to esti-
mate learner’s knowledge level, which is usually less than the maximum
needed and the accuracy of the outcome results compared to the estimations of
the expert-tutor.
1 Introduction
Educational assessment is a process of drawing reasonable inferences about what

learners know on the basis of evidence derived from observation of what they say, do
or make in selected situations [8]. An important consideration affecting the design of
the assessment is the purpose on which it will be used. Two broad purposes served by
assessment are the assessment to assist learning usually referred as formative assess-
ment and the assessment of individual achievement usually referred as summative
assessment. Formative assessment is part of the developmental process of learning
and assesses the quality of learning while summative assessment assesses the quantity
and retention of learning following the completion of a unit of instruction [5, 8].
In a web-based distance learning setting, information derived from both assess-
ment approaches as well as the learner’s interaction with the system can be stored by
the learning environment for further exploitation: (i) by learners to observe their
personal learning progress and to decide how to further direct their learning process,
and (ii) by tutors to individually support learners and formulate judgments about the
quality and the effectiveness of the provided content.
Moreover, when the learning environment is an Adaptive Educational Hypermedia
System (AEHS) [1], assessment provides a way to estimate learners’ knowledge level

154 E. Gouli, K. Papanikolaou, and M. Grigoriadou
and skills through learners’ responses in tests, characteristics that are usually used as
sources of adaptation. Thus, the generation of assessment tests tailored to learners’
responses could mainly enhance the adaptation of a system in this educational context
and extend its adaptive capabilities. Several techniques have been recently used in
learning environments introducing adaptation to the assessment process such as adap-
tive testing [11] and adaptive questions [2].
In this paper, we propose a comprehensive framework for assessment, developed
through the web-based module named PASS-Personalized ASSessment. Multiple
assessment options tailored to the learner’s responses are provided: (i) pre-testing and
self-assessment that correspond to the formative assessment, and (ii) summative as-
sessment. The assessment process is based on the central idea of the adaptive testing
and the adaptive questions techniques. The PASS module can be integrated in any
AEHS with certain characteristics.
The paper is organized as follows. In Sect. 2, a description of the adaptive testing
and the adaptive questions techniques is provided. Then, in Sect. 3, the PASS module
is presented and in Sect. 4, the functionality of PASS module is described according
to the proposed framework of the assessment. In Sect. 5, experimental results from
the formative evaluation of PASS module are discussed and in Sect. 6, the paper ends
with concluded remarks and plans for further research.
2 Adaptation in Assessment
The enhancement of the assessment process with adaptive capabilities is meaningful

for at least two reasons: (i) the assessment process becomes dynamic and individual-
ized, as it is adapted to the learner’s performance, and (ii) the number of questions
required to estimate learner’s knowledge level is usually reduced, resulting in a less
tedious assessment process. The adaptive testing and the adaptive questions tech-
niques, which can be used to introduce adaptation in the assessment process, are pre-
sented below.
Adaptive Testing involves a computer-administered test in which, the selec-
tion/presentation of each question and the decision to stop the process are dynami-
cally adapted to the learner’s performance in the test [11], [4].
Computerized adaptive testing relies on Item Response Theory [11]. In the adap-
tive testing procedure, a learner answering questions correctly (or incorrectly) will be
gradually administered more difficult (or easier) questions. Questions are selected so
that their difficulty matches the learner’s estimated knowledge level. The questions
that provide most amount of “information” about the actual knowledge level of the
learner are usually those with difficulty similar to the learner's knowledge level, as it
is estimated by the system, and low guessing factor. The learner’s knowledge level
estimation depends on the number of questions answered correctly and on the diffi-
culty level of the answered questions. The assessment process will hopefully reach an
accurate estimation of the learner’s knowledge level after a number of questions
posed.
Personalizing Assessment in Adaptive Educational Hypermedia Systems 155
Several approaches exploit the idea of adaptive testing. Huang in [4] describes an
adaptive testing algorithm, CBAT-2, that generates content-balanced questions and
the SIETTE system [10] is a Web-based testing system with adaptive capabilities.
Adaptive questions technique defines a dynamic sequence of questions depending on
learner’s responses. Specific responses or learner’s performance to certain questions
trigger the next series of questions according to several predefined rules, in contrast to
the adaptive testing technique, where the triggered question is the one that provides
the most amount of “information” about the learner’s actual knowledge level.
The adaptive questions technique uses a pool of questions, which are highly struc-
tured. Questions are grouped in different classes according to criteria specified by the
tutor, such as specific learning outcomes. Furthermore, each class may contain a
number of subclasses with different characteristics, such as the difficulty level of their
questions. Classes/subclasses are triggered following a certain sequencing mainly
determined by the learner’s responses. Each time a class/subclass is triggered, all its
questions are posed.
Adaptive questions have been used mainly in computer-assisted surveys. Pitkow
and Recker in [9] showed that Web-based adaptive questionnaires can reduce the
number and complexity of questions posed to users. Furthermore, in CATES [2],
“adaptive questionnaires” are used to assess Web users’ attitudes.
3 PASS - A Personalized ASSessment Module
PASS is a web-based assessment module, aiming to estimate learner’s performance

and to assess specific learning outcomes, which are congruent with the learner’s
learning goal. To this end, it provides multiple assessment options. Tutors can use it
in order to define the assessment specifications and to have a detailed overview of the
learners’ performance and progress. Learners can use PASS for taking a pre-test, self-
assessment and summative assessment tests (multiple assessment options) tailored to
their responses. PASS can be integrated to any AEHS with a structured domain
knowledge and a learner model that keeps information about the navigational behav-
iour of the learner.
PASS module has been integrated and evaluated in a simulated environment of
INSPIRE [7]. INSPIRE is a web-based Adaptive Educational Hypermedia System,
which aims to facilitate distance learners during their study, providing them with
personalized support. Based on the learning goals that the learner selects, INSPIRE
generates lessons that correspond to specific learning outcomes, accommodating the
knowledge level of the learner and his/her learning style.
In order to follow the functionality of the different components comprising the
PASS module, we provide a brief description of the domain knowledge of INSPIRE.
This is represented in three hierarchical levels of knowledge abstraction [7]: learning
goals, concepts and educational material. The outcome concepts, which are the most
important concepts of a learning goal, are associated with a number of prerequisite
concepts. The educational material of an outcome concept is organized in three dif-
ferent levels of performance [6]: Remember level (associated with learner’s ability to
recall the provided content), Use level (associated with learner’s ability to apply the
provided content in specific problems) and Find level (associated with learner’s abil-
ity to propose and solve original problems). On each particular level of performance,
one or more educational material pages, comprising of multiple types of knowledge
modules, such as examples, theory presentations, exercises, activities, are provided.
A ssessm ent Base (A B )
Q uestion and
Test Editor
Q uestion, Feedback and (Q TE)
A ssessm ent Base
Expert-tutor
Tem p orary Learner

M odel
Perform ance Estim ation

com ponen t (PE)
Test A daptive
G eneration
com ponen t (TA G ) Learner’s respon ses
Presen tation
C om ponen t (PC )
Learn er
Fig. 1. Schematic of the PASS module
The schematic of the PASS module is illustrated in Fig 1. It is comprised of the

following components:
– Assessment Base (AB): a data storage, which contains a pool of questions with a
variety of parameters, assessment parameters and a Temporary Learner Model
(TLM). The TLM stores all the information regarding the learner’s interaction with
the PASS module and the temporary estimations concerning the knowledge level
of the learner.
– Question and Test Editor (QTE): a tool that allows tutors to define: (i) the ques-
tions, (ii) the feedback provided to the learner for each of the alternative answers
of the question, (iii) the question’s parameters such as the initial difficulty level,
the number of alternative answers, the level of performance (in this case: Remem-
ber, Use, Find) that the question assesses, (iv) the assessment parameters such as
the maximum number of questions in each particular test, the minimum number of
questions for each level of performance, the degree of confidence in the estimated
knowledge level, the termination criteria, and (v) the weight of each educational
material page and of each prerequisite concept denoting their importance for the
outcome concept. The QTE allows the tutor to concentrate on authoring questions
and relieves him/her of the strain of technological details.
– Test Adaptive Generation (TAG): selects the appropriate question to be posed to
the learner according to: (i) the questions’ parameters, (ii) the assessment parame-
ters, and (iii) the current estimation of the learner’s knowledge level recorded on
the TLM. It takes all the required information from the AB and following the algo-
rithm of the selected assessment option, it selects the “best” next question/class of
questions. Furthermore, the TAG uses multiple termination criteria in order to ter-
minate the assessment procedure (see Sect. 4 for more details).
– Performance Estimation (PE): estimates the learner’s performance level and up-
dates the TLM (see Sect.4 for more details).
– Presentation Component (PC): presents the question selected by TAG and the
appropriate feedback according to the learner’s response. Also, it provides a
graphical and a text-based representation of the learner’s learning progress.
4 PASS Module’s Functionality
PASS, as an assessment module, provides a variety of evidence to support educa-

tional decision making, offering three assessment options: (i) pre-testing, (ii) self-
assessment on the content that the learner has already studied, and (iii) summative
assessment on the overall content. These options, in the context of AEHSs, can pro-
vide essential information about the learner’s performance in different stages of the
learning process. The adequate exploitation of this information can mainly enhance
the dynamic adaptation of an AEHS towards: (i) personalizing the content that will be
delivered to the learner, and (ii) providing adaptive navigational support through the
content.
Pre-testing functionality. Pre-testing has been designed: (i) to provide a preview of
the subject matter, (ii) to investigate the learner’s prior knowledge in the provided
content, (iii) to diagnose learning difficulties in the prerequisite material, and (iv) to
initialise the learner’s temporary model.
To fulfil the above aims, the three following categories of questions have been
specified: (i) questions assessing learners’ prior knowledge on the subject matter by
providing a preview of it, (ii) open questions providing information about learners’
experience and prior knowledge on related subjects, and (iii) questions assessing
learners’ knowledge on the prerequisite material.
Specifically, the technique of adaptive questions is adopted for the selec-
tion/presentation of questions assessing learners’ knowledge on the prerequisite mate-
rd
rial (3 category of questions). This technique is adopted in order to enable the learn-
ers to refrain from frustration, which is usually caused when they are asked to answer
questions the level of which is above their knowledge. In the proposed approach, a
class of questions is specified for each prerequisite concept of an outcome concept.
For each particular class, a number of subclasses is defined, aiming to: (i) assess the
level of learners’ performance, and/or (ii) draw meaningful inferences through learn-
ers’ responses. The rules that trigger the next class/subclass of questions are based on:
(i) the percentage of learners’ correct responses, and/or (ii) qualitative criteria defined
to draw specific inferences about the gaps in the learners’ knowledge.
An overview of the way the pre-testing is provided through the PASS module is
described below (not yet fully implemented). Once the learner has decided to take a
pre-test, the TAG selects the first two categories of questions and the PC presents
them to the learner. Next, the TAG applies the adaptive questions procedure, for
questions assessing learners’ knowledge on the prerequisite concepts, taking into
account the learner’s responses. The TAG triggers the appropriate classes/subclasses
of questions following certain rules. All the questions of each triggered subclass are
posed. When the assessment procedure is completed for a class, the PE estimates the
learner’s knowledge level for the particular prerequisite concept with which the class
is associated. Then, a new class of questions is triggered. A qualitative model, which
classifies learner’s knowledge level to one of the four levels of proficiency {Insuffi-
cient (In), Rather Insufficient (RI), Rather Sufficient (RS), Sufficient (S)} is used.
Following the termination of the pre-testing, the PE estimates the prior knowledge
level of the learner on each outcome concept by taking into account: (i) the percent-
age of correct responses on questions presenting a preview of the subject matter, (ii)
the learner’s knowledge level estimation for the prerequisite concepts of the outcome,
and (iii) the weights of the prerequisite concepts denoting their importance for the
outcome with which they are associated. This estimation is used as an initial estima-
tion of the learner’s knowledge level in the adaptive testing procedure, which is per-
formed during the self-assessment and the summative assessment.
Self-assessment and summative assessment functionality. Self-assessment aims to
stimulate the learner to contemplate and reflect on the content that s/he has studied, to
assess the learner’s performance and to provide immediate and informative feedback.
Summative assessment intends to evaluate the learning outcomes of the instruction by
indicating an end mastery of the learning outcomes for each outcome concept. The
learner has always the option to select self-assessment or summative assessment.
Moreover, the system has the ability to propose the most appropriate one, according
to the navigational behaviour of the learner through the content i.e. in case the learner
has not visited all the provided material of an outcome concept, self-assessment is
proposed for the outcome.
The construction of the self-assessment and the summative assessment tests is dy-
namic, depending on the current knowledge level of the learner. Moreover, in self-
assessment, the navigational behaviour of the learner through the content is also con-
sidered. The assessment procedure takes into account the content that the learner has
visited as well as the time that s/he has spent studying the material. Thus, the AEHS
needs to keep such kind of information in the learner model. In summative assess-
ment, questions associated with the prerequisite concepts and questions relevant to
each outcome concept, are posed.
The technique of adaptive testing is adopted for self-assessment and summative as-
sessment. The assessment procedure in both these approaches works as follows:
st
– 1 step: The TAG looks for an initial estimation of the learner’s knowledge level in
the TLM. If no estimation exists, an average is assumed.
nd
– 2 step: Depending on the assessment option selected, i.e. self-assessment or
summative assessment, the maximum number of questions to be posed for each
performance level (Remember, Use, Find) is estimated. In the case of self-
assessment, the weight of each educational material page (see in Sect. 3, QTE) and
the navigational behaviour of the learner, are taken into account. In the case of
summative assessment, just the weight of each educational material page is consid-
ered.
rd
– 3 step: The TAG selects the candidate questions according to the assessment op-
tion, the navigational behaviour of the learner and the minimum and maximum
number of questions for each performance level. Questions that are associated with
important educational material pages (pages with higher weight) have a greater
chance of being chosen.
th
– 4 step: For each candidate question, the TAG calculates the Item Characteristic
Curve (ICC) and the Item Information Function (IIF) based on the current estima-
tion of the learner’s knowledge level, the difficulty level of the question and the
guessing factor [11] (the discriminatory power is omitted). The ICC represents the
probability that the learner with a certain knowledge level will be able to provide a
correct answer and the IFF is a representation of the amount of “information” pro-
vided by each question about the learner’s knowledge level. According to the
weight of each educational material page and the amount of “information” pro-
vided by each question, the TAG selects the “best” question. Usually, questions
with difficulty similar to the learner's knowledge level and low guessing factor
provide the most amount of “information”. In the proposed approach, the difficulty
level of each question is initially assigned by the tutor and as the question is used
in the assessment procedure, it is re-estimated according to the number of times
that it has been answered correctly or incorrectly [4]. The calibration of the diffi-
culty level of questions is necessary.
th
– 5 step: The PE, according to the learner’s response, estimates his/her knowledge
level. The knowledge level is adjusted by a quantity, which depends on the previ-
ous estimation of learner’s knowledge level and on all his/her previous responses.
The PE classifies learner’s knowledge level for each outcome concept of the se-
lected learning goal, to one of the four levels of proficiency: {In, RI, RS, S}.
− 6 step: The TAG terminates the assessment procedure when any of the following
th
predefined criteria is met: (i) when the number of questions posed exceeds the
maximum number of questions defined for the test (see Sect. 3, QTE), (ii) when all
the questions in the AB have already been posed, (iii) when the degree of confi-
dence in the estimated learner’s knowledge level is high, (iv) when the number of
questions posed exceeds the maximum number of questions estimated for each
nd
particular level of performance (see 2 step).
Steps 4 and 5 are repeated until the termination criteria are met. For a detailed de-
scription of the algorithm, which is followed by the above assessment procedure, see
[3]. At the end of the assessment procedure, information on the learner’s achievement
and charts showing the learner’s progress over the learning process are provided and
stored in the system for further exploitation by the tutor.
5 Experimental Results
The first stage of the formative evaluation of the PASS module aims to check the
validity of its estimations concerning the knowledge level of the learner and the effi-
ciency of the proposed procedures. The study, performed in this context, focused on
the module’s estimations provided on the self-assessment and summative assessment
options, which were compared to: (i) the diagnostic process of an expert-tutor, and
(ii) the simplified process of accounting the percentage of right answers, a method
adopted in many AEHSs. In order to check the validity of the estimations, we exam-
ined the accuracy of the outcome results, i.e. the percentage of learners who were
classified by PASS into the same level of proficiency as classified by the expert-tutor.
The efficiency of the assessment procedure used in the summative assessment option
(the same procedure is used in self-assessment) was examined through the total num-
ber of questions posed to the different learners.
The students’ data used in this study have been obtained from an experiment,
which was performed as a part of the evaluation of INSPIRE at the Dept. of Informat-
ics and Telecommunications, University of Athens. In this experiment, twenty under-
graduate students used INSPIRE for two hours in order to study the learning goal
“Describe the Role of Cache Memory and its basic operations”. Once the students had
studied the educational material of the outcome concept "Mapping Techniques", they
were asked to submit an assessment test on the particular outcome (summative as-
sessment). The professor of the “Computer Architecture” course who had the role of
the expert-tutor assessed these tests. The estimations of the expert-tutor were based on
the general impression given by the tests, the difficulty of the questions, the number
of correct answers on each performance level and the comparison between students’
answers.
Students’ answers on the test were also provided to PASS and the expert-tutor as-
signed the required parameters. The maximum number of questions posed in the test
was assigned to 15. The initial difficulty of the questions was assigned according to
the level of performance that the question assessed, e.g. for questions assessing the
Remember level of performance, the initial difficulty was 0.3. For question’s diffi-
culty calibration, data from the final exams of the course “Computer Architecture”
were used. The initial knowledge level of the students was assumed as {RI}. The
scale of the four levels of proficiency, i.e. {In, RI, RS, S} was experimentally set.
This scale provides results, which are closer to the estimations of the expert-tutor. The
study was also performed by classifying learner’s knowledge level to five and to six
levels of proficiency. We found that if four levels of proficiency are considered then
the accuracy of the estimations is higher compared to the other two approaches.
Additionally, we estimated the students’ knowledge level based on the percentage
of correct answers, according to heuristic rules i.e. if the percentage of correct an-
swers is between 0-25% or between 26-50% or between 51-75% or over 75% then the
proficiency level is estimated as {In} or {RI} or {RS} or {S} correspondingly.
Figure 2 shows the summative assessment results obtained from the classification
of 20 students into the four levels of proficiency using the three estimation ap-
proaches. Unfortunately, none of the students’ proficiency level has been character-
ized as {S}. The reader may notice that for 17 out of 20 students, the proficiency
level estimations, resulted from the PASS module, coincide with the expert-tutor’s
estimations. On the other hand, for only 4 students, the expert-tutor’s estimations
coincide with estimations resulted from the percentage of correct answers approach.
Although the sample is rather small to reach a safe conclusion, the experimental re-
sults imply that the student’s proficiency level estimations provided by PASS follow
the expert-tutor’s estimations with high accuracy.
Percentage estimation Expert's estimation PASS estimation
RS
Proficiency Level
RI
In
1st 2nd 3rd 4th 5th 6th 7th 8th 9th 10th 11th 12th 13th 14th 15th 16th 17th 18th 19th 20th
Learners
Fig. 2. The proficiency level of 20 students on the concept “Mapping Techniques” as it is

estimated by: (i) calculating the percentage of students’ correct responses, (ii) the expert-tutor,
and (iii) the PASS module.
Furthermore, another interesting point, which we considered, was the total number
of questions that were posed to each student, a parameter that mainly affects the test-
ing time. Figure 3 shows the total number of questions posed to students in order to
estimate their proficiency level. The reader may notice that 11/20 students answered
less than 15 questions. These results imply that less time is needed to take the assess-
ment test since fewer questions are required to achieve an acceptable accuracy. Fur-
thermore, Fig. 4 shows the relationship between the average number of questions
posed and the estimation of student’s level of proficiency. The initial students’ profi-
ciency level was assumed as {RI} and as we have mentioned above, questions were
selected in a way that their difficulty was similar to the student’s knowledge level.
The reader may notice, in Fig.4, that for students with a proficiency level estimated as
{RS}, the average number of questions posed was less, compared to those with a
Number of learners
10 14
Number of questions
9
8 12
7 10
6
posed
5 8
4
3 6
2
4
1
0 2
15 questions 10-14 6-9 0
questions questions In RI RS
Number of questions posed Levels of proficiency
Fig. 3. Number of questions posed by Fig. 4. Average number of questions

the PASS module posed for each level of proficiency
lower level of proficiency, i.e. {In} or {RI}. One possible explanation is that students
with lower levels of proficiency usually guess the answers, so the assessment proce-
dure needs to pose more questions for estimating accurately their knowledge level.
6 Conclusions and Further Research
In this paper, we presented PASS, a web-based and easy to integrate in an AEHS,

assessment module. PASS offers three assessment options tailored to learner’s re-
sponses: pre-testing, self-assessment and summative assessment. The technique of
adaptive questions has been adopted for pre-testing. The discriminative characteristics
of the self-assessment and summative assessment procedures adopted in PASS are:
the adoption of the adaptive testing technique, the consideration of the learner’s
navigational behaviour, the re-estimation of the difficulty level of each question each
time it is posed (initially assigned by the expert-tutor), and the consideration of the
importance of each educational material page.
The experimental results from the formative evaluation of PASS, even performed
on a limited test group, have been encouraging and indicate that the estimations on
learners’ proficiency are close to the expert-tutor estimations. We have also shown
that a reduction in the number of questions posed can be achieved, especially for
learners with higher levels of proficiency, resulting in a reduction of the testing time.
Our future plans, apart from the evaluation of the PASS module with a wider
group of learners, include the completion of the pre-testing implementation and the
adaptation of the feedback provided to the learners’ according to their performance.
References
1. Brusilovsky, P.: Adaptive and Intelligent Technologies for Web-based Education. In:
Rollinger, C., Peylo, C. (eds.): Special Issue on Intelligent Systems and Teleteaching.
Kunstliche Intelligenz, Vol. 4 (1999) 19-25
2. Chou, C.: Constructing a computer-assisted testing and evaluation system on the World
Wide Web–the CATES experience. IEEE Transactions on Education 43(3)(2000) 266-
272
3. Gouli, E., Kornilakis, H., Papanikolaou, K., Grigoriadou, M.: Adaptive Assessment im-
proving interaction in an Educational Hypermedia System. In: Avouris, N., Fakotakis, N.
(eds.): Advances in Human-Computer Interaction I. Proceedings of the PanHellenic Con-
ference with International Participation in Human-Computer Interaction (2001) 217-222
4. Huang, S.: A Content-Balanced Adaptive Testing Algorithm for Computer-Based Train-
ing Systems. In: Frasson, C., Gauthier, G., Lesgold, A. (eds.): Intelligent Tutoring Sys-
tems, Third International Conference, Springer (1996) 306-314
5. Kommers, P., Grabinger, S., Dunlap, J.: Hypermedia Learning Environments. Instruc-
tional Design and Integration. Lawrence Erlbaum Associates (1996)
6. Merril, D.: Component Display Theory. In: Reigeluth, C.M. (eds.): Instructional design
theories and models: An overview of their current status. Lawrence Erlbaum Associates,
Hillsdale NJ (1983) 279-333
7. Papanikolaou, K., Grigoriadou, M., Kornilakis, H., Magoulas, G.: INSPIRE: An INtelli-
gent System for Personalized Instruction in a Remote Environment. In: Reich, S., Tzaga-
rakis, M.M., De Bra, P. M.E. (eds.): Hypermedia: Openness, Structural Awareness and
Adaptivity. Lecture Notes in Computer Science Vol. 2266, pp. 215-225. Springer-Verlag,
Berlin (2002)
8. Pellegrino, J., Chudowsky, N., Glaser, R. (eds): Knowing what students know: The Sci-
ence and Design of Educational Assessment. National Academy of Sciences, National
Academy Press, Washington (2001)
9. Pitkow, J., Recker, M.: Using the Web as a Survey Tool: Results from the Second WWW
User Survey. Computer Networks ISDN Systems 27(6) (1995) 809-822
10. Rios, A., Millan, E., Trella, M., Perez-de-la-Cruz, J., Conejo, R. : Internet Based Evalua-
tion System. In: Lajoie S., and Vivet M. (eds.): Artificial Intelligence in Education, Open
Learning Environments, New Computational Technologies to Support learning Explora-
tion and Collaboration. Frontiers in AI and Applications, Vol 50. IOS Press (1999) 387-
394
11. Wainer, H., Dorans, N., Eignor, D., Flaugher, R., Green, B., Mislevy, R., Steinberg, L.,
nd
Thissen, D. (eds): Computerized adaptive testing: A primer. 2 edition. Lawrence Erl-
baum Associates, Hillsdate NJ (2000)
Visual Based Content Understanding
towards Web Adaptation
Xiao-Dong Gu1,*, Jinlin Chen2, Wei-Ying Ma , and Guo-Liang Chen1

3
1 Dept. of Computer Science, Univ. of Sci. & Tech. of China,

230027, Hefei, China
[email protected]
2 School of Information Science, Univ. of Pittsburgh,
Pittsburgh, PA 15260 USA

[email protected]
3 Microsoft Research Asia
Beijing 100080, China

[email protected]
Abstract. Web content structure is proposed to facilitate automatic web page

adaptation in this paper. By identifying the logic relationship of web content
based on layout information, web content structure effectively represents au-
thors’ presentation intention. An automatic top-down, tag-tree independent ap-
proach to detect web content structure is presented. It simulates how a user un-
derstands web layout structure based on his vision. Comparing to other content
analysis techniques, our approach is independent to physical realization and
works well even when the physical structure is far different from layout struc-
ture. Besides, our approach is an O(n)-time process which is much more effi-
2
cient comparing to other approaches with O(n )-time complexity. Furthermore,
our approach is tag tree independent, which means it can be applied to web
contents of arbitrary physical realization formats. Experiments show satisfac-
tory results.
1 Introduction
Web publishing is playing a more and more important role for information distribu-
tion today. When creating a new web page, the author first decides what to present,
i.e., the semantic content. He then further decides how to present the information.
Finally, a markup language is used to realize the presentation, which gives a physical
structure to the content. Considering the whole process, we have XML to represent
semantic structure and many markup languages such as HTML to represent physical
structure. However, we still lack an effective way to represent presentation structure,
which indicates authors’ intention towards the presentation logic of web content.
Furthermore, with the exponential growth of information and increasing diversity
in terms of devices and networks in today’s web, it becomes increasingly pressing to

Visual Based Content Understanding towards Web Adaptation 165
access desired contents accurately and conveniently. Various content adaptation tech-
nologies [1-4,13,14] have been developed for this purpose. One crucial issue for
content adaptation is the need to effectively represent and understand presentation
structure of web pages. Many web content analyzers have been proposed to extract
structural information from web pages either manually [5-8] or automatically [9-11].
However, a big problem with these approaches is that they try to extract structural
information from HTML tag tree directly, which often leads to unstable results be-
cause HTML tags were designed for both presentational and structural representation.
Besides, most of them are bottom-up approaches which are time consuming. In addi-
tion, these approaches are only suitable for HTML documents.
To solve these problems, we first propose the web content structure, which at-
tempts to represent author’s presentation intention by identifying the logic relation-
ship of web content based on visual layout information. An automatic top-down, tag-
tree independent approach to detecting web content structure is then presented. It
simulates how a user understands the web layout structure when he browses a page
such as objects’ size, position, color, background, etc. A projection-based algorithm is
applied to segment a web page into blocks. Blocks are further divided into sub-blocks
or merged if they are visually similar. In this way we avoid the breaking of logical
chunks. Comparing to other existing approaches, our approach is independent of
physical realization and performs even when the physical structure is different from
visual layout structure. Besides, our approach is an O(n)-time process which is much
2
more efficient comparing to the O(n )-time process in [11] (n is the number of basic
objects). In addition, our approach is tag tree independent, which can be applied to
contents with various physical realization formats.
This paper is organized as follows. In Sect. 2 we will first introduce the web con-
tent structure. It is followed by the automatic detection of web content structure in
Sect. 3. Experimental results and conclusions are given in Sects 4 and 5, respectively.
2 Web Content Structure
When designing a web page, the author first collects all the basic objects for the page.
A basic object is the smallest unit of a web page, which cannot be further divided to
perform some certain functions. He then groups related basic objects together to
achieve a major function. This group of objects is called a composite object. Compos-
ite objects can be further grouped into a more complex one. This is a recursive proc-
ess until all the objects in a web page are grouped together. During this process the
author also needs to consider how to visually separate the composite objects.
Based on the analysis above, to fully express the presentation design of a web
page, web content structure should represent the layout structure of a web page,
which is the result of dividing and subdividing the content of a web page into increas-
ingly smaller part, on the basis of presentation. Layout structure includes the structure
of composite objects and how they are separated from each other. Similar to the de-
scription of document representation in [12], the basic model of web content structure
is described as below:
166 X.-D. Gu et al.
A web page • can be represented by a triple

• =(•,•,•) . (1)
where •={•1, •2,…, •n} is a finite set of objects, including basic objects and composite
objects. Each composite object can be viewed as a sub-content structure. •={•1, •2,…,
•t} is a finite set of separators in visual representation, including horizontal and verti-
cal separators. • is the relationship of every two objects in O and can be expressed as:
•=•×•••∪{NULL} . (2)
Suppose •i , •j ∈O, •(•i, •j)•NULL indicates that •i and •j are exactly separated by the
separator •(•i, •j), or we can say that the two objects are adjacent to some extent. Oth-
erwise there exists other object between •i and •j.
Based on the definition above, authors can easily represent their presentation de-
sign in a formulas manner.
Fig. 1. An example of web content structure
Figure 1 gives an example of web content structure, in which

Ο = {Ο1 , Ο 2 , Ο 3 , Ο 4 , Ο 5 , Ο 6 }
Φ = {L1 , L2 , L3 , L4 , L5 }
 (Ο1 , Ο 2 )   L1 
   
 ( Ο 2 , Ο 3 )   L2 
 (Ο , Ο )   L 
δ = Ο×Ο → Φ :δ  3 4  =  3 
 ( Ο 4 , Ο 5 )   L4 
 (Ο , Ο )   L 
 5 6   5 
 else   NULL 
   
In the page L1-L5 are all horizontal separators, O1-O6 are all Composite Objects,
which can be further subdivided. For example, O1 can be divided into two sub-blocks
O11 and O12 by a vertical separator.
3 Automatic Web Content Structure Detection
Since web content structure reflects authors’ intention directly, it plays a crucial role
for content understanding. Therefore, detecting web content structure from existing
web pages is very important for content adaptation. Although it is desirable that addi-
tional information be added for the generation of web content structure during the
authoring stage, authors often decide not to do so. Thus, it is important that web con-
tent structure can be automatically detected.
In this section, we describe a top-down tag-tree independent approach to detecting
web content structure based on page visual presentation. Our proposed approach
simulates how a user understands the web layout structure when he browses a page.
Below is the detailed description to the algorithm.
3.1 System Structure
Web content structure detection is a reverse process to web authoring. We start from
physical structure of web content to find out the presentation scheme. The detection
of web content structure is to analyze and establish the quintuple of (1), i.e. •, • and •
defined in Sect. 2.
Figure 2 illustrates the framework of our approach. The detection process is the di-
vision and repeated subdivision of a web page into increasingly smaller parts (ob-
jects). The detected structure can be visually represented by a geometric tree as in
Physical Preprocessing
structure
Dividing
Acceptable
Web content judgment
structure Merging
Fig. 2. Web content structure detection

168 X.-D. Gu et al.
Fig. 1 (c). To construct web content structure, first basic objects are extracted from
the physical structure (tag tree). We then preprocess the basic objects to find out
decoration objects and group similar ones to reduce complexity. Then based on web
visual representation, the whole page is divided into blocks through projection. Adja-
cent blocks are merged if they are visually similar. This dividing and merging process
continues recursively until the layout structure of the whole page is constructed. Be-
low we will first introduce the two major components in our system: dividing and
merging. We then give a brief introduction to preprocessing module.
3.2 Dividing Blocks via Projection
Projection is applied to divide a block into smaller sub-blocks. Projection refers to the
mapping of a web page into a waveform whose value is the sums of the values of the
object weights along projection axis. All objects in a web page are contained in rec-
tangular blocks. Blanks are placed between these rectangles. Thus, the projection
profile is a waveform whose deep valleys correspond to the blank areas of the web
page. A deep valley with a width greater than an established threshold can be consid-
ered as separator between objects. The process of projection is performed recursively
until all objects are located. Figure 3 (a) gives an example of projection. Every object
is projected along a projection axis. The sub-lines in the projection axis without any
objects projecting into indicate the separators (see L1 and L2 in Fig. 3 (a)).
Fig. 3. Using projection to divide block
Figure 3(c) explains the details of the projection process. Let {O1,O2,…,On} be the
n objects in a block, and si , ei be the starting and ending point of Oi along a projection
axis. Our goal is to find out all separators • = {L1, L2, …, Lt} along the projection axis.
Let link[j] be the judge value of point j at the projection axis. Then the weight value
of Oi (1 by default) is added to link[si]. Similarly, the weight value of Oi is deducted
from link[ei]. Point j is considered a separator point if link[j] is zero. Below is the
detailed algorithm.
Step 1: Let C be the set of the starting and ending points of all the objects. C =
(∪i{si})∪ (∪i{ei}), and |C| = K;
Step 2: Sort C into ascending order;
Step 3: Let link[0] = 0, we then project all the objects onto the projection axis:
for (j = 1; j <= K; j++)
link[j] = link[j-1]+c-d;
where c is the number of objects with starting point at C[j], and d is the number of
objects with ending point at C[j]. This repetitive expression realizes the projection
process in current block. If there are totally c objects starting at point j, the sum of the
weights of these c objects should be added to link[j-1] to get link[j]. Correspondingly,
if there are totally d objects ending at point j, the sum of the weights of these d ob-
jects should be deducted from link[j-1] to get link[j].
Since default weight is equal to 1 for each object, the resulting values of the sum of
weights are c and d, respectively. More fine-grained weight can be assigned to com-
pute link[j] for more accurate result.
Sub-line C[j]+1…C[j +1]-1 is a separator if link[j] is 0. This means that there is no
object in the area between Sub-line C[j]+1…C[j +1]-1.
Our algorithm requires O(n) time because each step above costs O(n) time. Com-
2
paring to the O(n )-time required in [11], our approach significantly reduces the com-
putational complexity.
Using the algorithm above, we detect separators L1[0…t1-1] in X-axis and L2[0…t2-
1] in Y-axis. Then the division in current level is:
Ο = {Ο1 , Ο 2 ,..., Ο (t1 +1)( t2 +1)−1 }
Sub-blocks are indexed from left to right as shown in Fig. 3 (b);
Φ = L1 [0...t1 − 1] ∪ L2 [0...t 2 − 1]
δ (Oi (t1 +1) + j , Oi ( t1 +1)+ j +1 ) = L2 [ j ] (0 ≤ i ≤ t 2 ,0 ≤ j < t1 )
δ (Oi (t1 +1) + j , O(i +1)( t1 +1) + j ) = L1 [i ] (0 ≤ i < t 2 ,0 ≤ j ≤ t1 )
.
As can been seen in Fig.3(b), projection in Y-axis reveals two horizontal separators
L1 and L2 while in X-axis reveals one vertical separator L3. These three separators
divide current block into six sub-blocks: O1~O6. The seven bidirectional arrows repre-
sent the relationship between sub-blocks and separators. Thus •,•,• are all detected.
3.3 Merging Sub-blocks via Layout Similarity
Since the division method above is only related to the position of objects, the separa-
tors detected may break a holistic object. Therefore the merging of some adjacent
sub-blocks is necessary.
Simulating human’s decision on whether two objects are similar, we use visual
similarity to decide whether two objects are holistic and should be merged. For basic
objects, if two objects are of different media type, their similarity x=0. Otherwise, x is
related to the media type. Below is an example of basic text objects. In a similar way
we can compute the similarity for basic objects of other media types.
170 X.-D. Gu et al.
• Starting from x=1.0, we first compare key HTML attributes (like <H1> …<H6>,
<A>). If not equal, x = x*Modifier_key.
• Compare alignment and other common attributes. If not equal, x =
x*Modifier_Common
• Compare font size attribute. If not equal, x = x*Modifier_Size.
• Compare styles (bold, italic, underline…). If not equal, x = x*Modifier_Style.
• Compare font face. If not equal, x = x*Modifier_Face.
• Compare text length, we have the following modification method:
factor Adjust
 min(length1, length2) 
x = x *  
 max(length1, length2) 
To calculate the similarity of two composite objects, we use an approximate similarity
measurement to compare two element strings that enables weighted mismatches and
omissions (skips). Weight of skipping may differ from element to element, because
some of the objects (such as those with attributes like <H1>...<H6>) could be more
important than others, and thus skipping of them would be costly (a small weight) or
not allowed (zero weight).
Some additional rules are used to modify visual similarity between objects for our
application based on the definition of visual similarity above:
1) Distance is an important factor for human to decide visual similarity. Let d be
the distance between two objects. The larger d is, the less their visual similarity is. Let
x be the visual similarity obtained above, a monotonic degressive function
Dist_Modifer(d) is applied to embody the impact of distance to similarity:
x = x ⋅ Dist _ Modifer (d )
2) Color has great influence for human to decide visual similarity. Two objects
with the same distinguished color and background color (different from those of
surrounding objects) are considered as a whole. Thus, x=1 if adjacent objects have the
same color and background color. Otherwise x remains its original value.
Based on the visual similarity between two adjacent objects, we can then decide
whether to merge them or not.
3.4 Preprocessing
Preprocessing is necessary to prevent decoration objects from enveloping separators

and prevent unexpected separators (e.g., shadowed separator in Fig. 4 is unexpected if
the dotted rectangle is a tightly-united object) from breaking object integrity.
By preprocessing we can not only decrease the opportunity of detecting unex-
pected separators, but also reduce the problem scale thus speedup the whole process.
Fig. 4. Unexpected separator
To evaluate our proposed approach, we randomly selected 50 web pages from popu-
lar sites listed in http://www.yahoo.com. We run our web content structure detection
over these pages and the results are listed in Table 1.
Table 1. Experiment results with a test set of 50 pages
DETECTING NUMBER OF
RESULTS DOCUMENT
Correct 45
Acceptable 4
Failed 1
Totally the layout structures of 45 (90%) pages are correctly detected. Some ap-
parent chunks are missed for 4(8%) pages. In most of these cases, one logical chunk
is broken into two or more segments. These segments are usually well separated by a
big white space from surrounding segments. Human consider these segments as a
whole chunk mostly according to their semantic meaning. Therefore, this is not
strictly the fault of our algorithm. Our detector fails to provide proper analysis to the
page because the page is confused in visual representation that even human eyes
cannot partition it correctly.
One benefit with our approach is that it is independent of how the web page is real-
ized in HTML tag-tree. Figure 5 gives an example. The left side bar (indicated in
dotted rectangle and was actually detected as a navigation list), which should be one
logical chunk, is actually realized with two HTML tables, which leads to mis-
detection in tag-tree based approach [11]. But with our approach the side bar is de-
tected as a whole.
172 X.-D. Gu et al.
Fig. 5. Our approach successfully detects the side bar as a whole while tag-tree based
approaches fail in this case because the tags are spread across different part of HTML
file.
5 Conclusions
In this paper the web content structure was proposed for web authoring, adaptation
and information retrieval. By identifying the logic relationship of web content based
on visual layout information, web content structure can effectively represent authors’
presentation intention. An automatic top-down, tag-tree independent algorithm to
detect web content structure was presented. It simulates how a user understands the
web layout structure based on its visual representation. Comparing to other ap-
proaches, our method is independent of physical realization and works well even
when the physical structure is far different from visual structure. Besides, our ap-
proach is an O(n)-time process which is much more efficient than other approaches
2
with O(n )-time complexity. Experiments show satisfactory results.
References
1. Ma, W.Y., Bedner, I., Chang, G., Kuchinsky, A., and H.J.Zhang. A Framework for Adap-
tive Content Delivery in Heterogeneous Network Environments. in Proc. MMCN2000
(SPIE Vol.3969), San Jose, USA (2000) 86-100.
2. Chen, J.L., Yang, Y.D., and Zhang, H.J.: An Adaptive Web Content Delivery System, in
Proc. AH2000, Springer (2000) 284-288.
3. Smith, J.R., Mohan, R., and Li, C.S.: Scalable Multimedia Delivery for Pervasive Comput-
ing, in Proc. of the 7th ACM International Conference on Multimedia (1999) 131-140.
4. Bickmore, T.W., and Schilit, B.N.: Digestor: Device-independent access to the World Wide
Web , in Proc. WWW6 (1997) 655-663.
5. Hammer, J., Garcia-Monlina, H., Cho, J., Aranha, R., and A. Crespo: Extracting semistruc-
tured information from the web, in Proc. PODS/SIGMOD’97 (1997) 18-25.
6. Ashish, N., and Knoblock, C.: Wrapper generation for semi-structured Internet sources, in
Proc. PODS/SIGMOD’97 (1997) 8-15.
7. Simth, D., and Lopez, M.: Information extraction for semi-structured documents, in Proc.
PODS/SIGMOD’97 (1997) 117-121.
8. Nestorov, S., Abiteboul, S., and Motwani, R.: Inferring Structure in Semistructured Data,
in Proc. PODS/SIGMOD’97 (1997) 39-43.
9. Embley, D.W., Jiang, Y., and Ng, Y.K.: Record-Boundary Discovery in Web Documents,
in Proc. SIGMOD’99, Philadelphia PA (1999) 467-478.
10. Lim, S.J., and Ng, Y.K.: An Automated Approach for Retrieving Hierarchical Data from
HTML Table, in Proc. CIKM’99, Kansas City, MO (1999) 466-474.
11. Yang, Y.D., and Zhang, H.J.: HTML Page Analysis Based on Visual Cues, in Proc. of the
6th International Conference on Document Analysis and Recognition, Seattle, USA (2001)
12. Tang, Y.Y., Cheriet, M., Liu, J., Said, J.N., and Suen, C.Y.: Document Analysis and Rec-
ognition by Computers, Handbook of Pattern Recognition and Computer Vision, World
Scientific Publishing Company (1999)
13. Chen, J.L., Zhou, B.Y., Shi, J. Zhang, H.J., and Wu, Q.F.: Function-based Object Model
Towards Website Adaptation, Proc. of the 10th International World Wide Web Conference,
Hong Kong, China (2001) 587-596.
14. Yang, Y.D., Chen, J.L., and Zhang., H.J.: Adaptive Delivery of HTML Contents, in
WWW9 Poster Proceedings (2000) 24-25.
Knowledge Modeling
for Open Adaptive Hypermedia
Nicola Henze and Wolfgang Nejdl
University of Hannover
ITI- Rechnergestützte Wissensverarbeitung
Appelstr. 4
D-30167 Hannover
{henze,nejdl}@kbs.uni-hannover.de
http://www.kbs.uni-hannover.de/˜{henze,nejdl}
Abstract. This paper proposes a knowledge modeling approach for

adaptive, open corpus hypermedia systems. Our approach towards adap-
tive, open corpus hypermedia is based on interpreting standard metadata
of learning objects. For each corpus of documents integrated into the
open adaptive hypermedia system (OAHS) we are calculating subgraphs
of the ontology for estimating the user’s knowledge with respect to this
corpus. This enables the OAHS to understand the knowledge contained
in learning materials, to make estimations about an individual user’s
knowledge state and to learn the prerequisite knowledge required for
learning objects from given structures in the materials.
1 Introduction
Hypermedia systems have become more and more popular as tools for user driven
access to information. Adaptive hypermedia systems bring together ideas from
hypermedia systems and intelligent tutoring systems, and enable personalized
access to information.
Recently, many approaches like SCORM [15], LOM [7] or IMS [1] begin to
standardize information for learning. They introduce metadata descriptions for
so called learning objects which are entities containing some learning material.
This metadata is e.g. an identifier, the creator of this material, the language,
keywords describing its content, the interactivity level of this material, etc. In
LOM there are currently about 80 metadata attributes defined. With these stan-
dardized descriptions of learning materials it is possible to build large learning
repositories in which teachers and students will be able to store or retrieve valu-
able materials for learning, for preparing courses etc. Current approaches are
e.g. ARIADNE [3] or EDUTELLA [8].
The individualization of the access to large learning repositories is a great
challenge for adaptive hypermedia. Normally, adaptive hypermedia systems work
on a closed set of materials which are often described by proprietary metadata
and whose adaptational functionalities are tailored to the current application.

Knowledge Modeling for Open Adaptive Hypermedia 175
The shift towards adaptation for open learning repositories will require to inter-
pret the standardized metadata of the learning objects accordingly and to find
generalizations of adaptational functionalities which can handle the demands of
open learning repositories.
In this paper we present an approach towards open adaptive hypermedia sys-
tems (OAHS). Our approach is driven by the claim to learn as much as possible
from standard learning descriptions (such as keywords) and from structures in
the learning materials (like reading sequences). This is motivated by the nature
of open learning repositories where authors of learning objects should have a
minimum overhead as possible for adding their learning objects into the reposi-
tory. For a definition of adaptive hypermedia systems we follow [4]:
Definition 1 (Adaptive Hypermedia System [4]). “By adaptive hyperme-
dia systems we mean all hypertext and hypermedia systems which reflect some
features of the user in the user model and apply this model to adapt various
visible aspects of the system to the user.”
Definition 2 (Open Adaptive Hypermedia System). An open adaptive

hypermedia system (OAHS) is an adaptive hypermedia system which operates
on an open corpus of documents.
The definition of OAHS includes that the system must be able to adapt
hypermedia documents to individual needs of a user regardless of the origin of
the materials: The materials may belong to some set of course materials, may
be a sequence of pages from a tutorial, may refer to content on personal home
pages, or, generally spoken, may be learning objects from an an arbitrary corpus
of documents.
The approach choosen here stems from research in adaptive hypermedia. We
see a strong relation to the area of Open Hypermedia Systems (OHS) [9], too.
Research in OHS focusses on cross-application access, on distributed hyperme-
dia and group work by developing e.g. advanced hypermedia data models and
hypermedia protocols [2]. So far as the authors know there has not been any
attempt to systematically apply adaptation strategies from adaptive hyperme-
dia research on top of open hypermedia - however this might be a very fruitful
combination (see for example [6]).
2 Adaptation in Open Corpus Hypermedia

Hypermedia systems can be described as a set of nodes (containing multiple
forms of media such as text, video, audio, graphics, etc.) which are connected by
links [14]). Consequently there are generally two aspects of hypermedia systems
which can be adapted to a user: the content (content level adaptation), and the
links (link level adaptation). By adapting the content for a user, the document’s
content is tailored to the needs of the user, for example by hiding too special-
ized information or by inserting some additional explanations. By using link
level adaptation, the user’s possibilities to navigate the hypermedia documents
176 N. Henze and W. Nejdl
are personalized. Techniques for content level and link level adaptation are for
example described in [4].
In case of open corpus hypermedia we face a large collection of learning
objects. As many authors can add arbitrary learning objects to an open learning
repository it is highly probable that we have an oversupply of information. Thus
for link level adaptation in OAHS it will be a central issue to decide which is
the best learning object for a user. For content level adaptation in OAHS the
metadata of a specific learning object must deliver information to redesign the
presentation of the learning object itself. A user of an OAHS might be a teacher
or a student as well. To consider students as users is very common in adaptive
educational hypermedia. However it is an important issue to support teachers as
users of an OAHS in the process of both storing and retrieving qualified learning
objects. In this paper we will focus on personalized access to learning objects for
students. Characteristics for a “best learning object” are e.g.
– current situation
E.g. a student might want to prepare an exam, to learn or to refresh her
knowledge.
– current interest
E.g. a learning goal, a teaching goal.
– current knowledge state
This is the most used characteristic in adaptive educational hypermedia and
very important for student modeling.
– current learning experience
E.g. based on previous studies. curriculum, general education.
– learning style
E.g. problem-based, constructivist.
– presentation style
E.g. graphics, textual descriptions, audios, animations.
– preferred author / corpus
Individual preferences for a specific way of presentation or a teaching style.
Most of the above stated characteristics have successfully been used in adap-
tive hypermedia systems so far. The main new characteristic for OAHS is “pre-
ferred author / corpus” which we think will play an important role during the
access to open learning repositories. For example, a student favors a specific
teacher and her way to explain things. Or a student wants to make sure that he
masters all topics of the class he is currently enrolled in.
3 Knowledge Modeling for OAHS

The knowledge of a user is the main source of information for adaptation in
educational hypermedia. Based on the user’s actual knowledge state, selection of
information takes place, examples are proposed, reading sequences are generated
and next learning steps are calculated for and presented to the user.
A knowledge model for OAHS has to fulfill the following requirements: It
must be
1. universally valid
Throughout the different corpora of the learning repository
2. expandable
E.g. learning objects on new topics may be added to one of the corpora
3. expressive
The knowledge contained in one single corpus must be describable as a valid
model on its own.
Ontologies have been developed in Artificial Intelligence to facilitate knowl-
edge sharing and reuse and therefor meet the first requirement. There are many
definitions of ontologies in literature. We follow Gruber [10]
Definition 3 (Ontology [10]). “An ontology is a formal, explicit specification
of a shared conceptualization.”
By generating the ontology as a shared conceptualization we had to find a
way to support different views on the categorization of concepts which we want
to illustrate by an example from the domain of Java Programming: “implements”
is used within a class declaration for implementing an interface. Therefor “im-
plements” can be expected to occur in the ontology nearby “interfaces” or “class
declaration”. But one can take the point of view that “implements” is a Java
keyword thus a person looking up Java keywords in the ontology can expect to
find “implements” there, too. All these different viewpoints on the classification
of “implements” are valid. In addition, as we will see in Sect. 4.1, the neigh-
borhood of an entity in the knowledge model is a rich resource for improving
the rating of relevance of learning objects during the retrieval process. Thus we
allow multiple occurrences of concepts in the ontology-based knowledge model.
Ontologies are able to meet the second requirement for the knowledge model
as well. As the example of the Java domain shows, new concepts are likely to
appear (there have been many new developments in the past, for naming one:
the migration from “AWT” to “Swing” in the Java Programming Language).
However, applying an ontology for knowledge modeling in OAHS includes that
the adaptation strategies and algorithms that we define must be capable to un-
derstand and to deal with new concepts which have been added to the ontology.
In order to model the knowledge of each individual corpus there are two
different strategies possible: We can apply a separate ontology for each corpus
and construct the overall ontology by ontology mapping [13]. The other strategy
is to use one common ontology and express the knowledge described in each of
the corpora as subgraphs of the ontology by labeling the nodes accordingly. We
decided for the latter approach as it allows us to easily compare the knowledge
of a user in relation to each of the corpora.
3.1 Assessment of Knowledge

If we employ an ontology-based knowledge model we have to rethink the way a
user can acquire knowledge in an OAHS. In OAHS U can learn by using learning
materials from different courses.
Example 1. User U can read about object-oriented programming in the course

materials of her class and then starts writing her first Java program with aid of
the Sun Java Tutorial [5].
Thus the knowledge of U must be estimated by rating her/his knowledge in
a broader context, not solely in terms of the curriculum of one course or one set
of learning materials.
To assess U ’s knowledge about a concept a with respect to some corpus
C we determine a minimal topic-subgraph of the ontology. This topic-subgraph
contains at least the topic a and all its child nodes. If neither one of the child
nodes nor a itself are contained in the corpus we add recursively parent nodes of
a to the subgraph until one parent node is contained in the corpus. Now we can
assess the knowledge of U with respect to some corpus C in the following way:
A concept which is contained in C can either be learned (value 1) or not learned
(value 0). The knowledge of a concept not contained in C is estimated by the
ratio of the number of learned concepts in the topic-subgraph which belong to
C by the number of all concepts in the topic-subgraph which belong to C .
Example 2. In Fig. 1 we interpret the knowledge of U in relation to corpus C as
follows
– Fig. 1A: U ’s knowledge of topic “a” in respect to corpus C : a|C = 1
U ’s knowledge of topic “b” in respect to corpus C : b|C = 1
U ’s knowledge of topic “c” in respect to corpus C : c|C = 1
– Fig. 2A: a|C = 1, b|C = 0.5, c|C = 0
– Fig. 3A: a|C = 1, b|C = 0.5, c|C = 0, d|C = 0.5
– Fig. 4A: a|C = 1, b|C = 0.33, c|C = 0, d|C = 0
– Fig. 5A: a|C = 1, b|C = 1, c|C = 1
– Fig. 6A: a|C = 0, b|C = 1, c|C = 0.5
Representation in one ontology
1A 2A a 3A 4A 5A 6A
a a a a
a
b c b c b c d b c b c
b c d
Representation in separate ontologies
1B 2B 3B 4B 5B 6B
a a a
a a
c c d b b
c
not read & not contained in corpus C not read & contained in corpus C
read & not contained in corpus C read & contained in corpus C
Fig. 1. Assessment of knowledge in an ontology-based knowledge model

3.2 Used Metadata of Learning Objects
The concepts in the knowledge model define a controlled vocabulary for describ-
ing the knowledge of the application domain and can be used for metadata
annotation of the learning objects (LOM, category General, data element 1.6
Keywords) (see Fig. 2). The learning objects themselves are solely described
and identified by a set of keywords describing their content.
Fig. 2. The role of the Controlled Vocabulary
We have implemented an early prototype of an OAHS which proves the ap-

plicability of our approach. This prototype is an adaptive hypermedia system
on Java Programming, the Java Hyperbook 1 . The Java Hyperbook is an adap-
tive system which uses course materials from an undergraduate course on Java
Programming held at University of Hannover and guides the student through
the course by showing next reasonable learning steps, by selecting projects, gen-
erating and proposing reading sequences, annotating the educational state of
information, and by selecting useful information, based on a user’s actual goal
and knowledge [11].
The adaptation component of this prototype was based on a separate knowl-
edge model which described the learning dependencies in our Java course. The
documents of the hypermedia system were annotated by a short metadata de-
1
http://www.kbs.uni-hannover.de/hyperbook
scription which contained a set of keywords describing the content of each doc-
ument. To prove the openness of the Java Hyperbook, we added the content of
the Sun Java Tutorial [5], a freely available online tutorial, to the Java Hyper-
book. The Java Hyperbook was capable to adapt to both corpora [12]. However
the coding of learning dependencies in the knowledge model has shown to be a
drawback by the simultaneous integration of different corpora: Each collection of
learning materials might follow their own learning / teaching strategy and there-
for might define different learning dependencies. Our approach was functioning
very well for our Java course and the Sun Tutorial but is not generalizable.
Thus we claim that for OAHS we need an independent knowledge model like the
ontology-based knowledge model proposed above.
4 Prerequisite Estimations and Enhanced Retrieval

In this section we will discuss two applications of the above presented ontology-
based knowledge modeling approach for improving the retrieval of learning ob-
jects and for making estimations on prerequisites required for learning objects.
We therefor need the following definition:
Definition 4 (Content Map). Let CV be the set of topics contained in our
ontology-based knowledge model (the controlled vocabulary). Let O be some set
of learning objects. Then
I : O → P(CV)\{∅}
is the content map, which gives for each learning object ∈ O the set of keywords
of this resource, e.g. the values of the metadata attribute 1.6 General, keywords
from LOM (P is the power set).
4.1 Enhanced Information Retrieval

An ontology defines a context for each of its topics x: topics “nearby” x describe
details or more general aspects of x. This observation can be used for enhancing
the retrieval of learning objects. Each query for learning objects can be checked
in the following way: if all topics of the query are handled in the learning object
(which means that they are contained in the learning objects metadata) we are
done. Otherwise we rate how “near” the topics covered by this object are to he
query by determining the minimal distance from each keyword from the learning
object to the set of query terms in the graph of the ontology.
More formally: Let O be some learning object, and {q1 , . . . , qn } the query
set. An x, y up-down path in a directed graph G is the edge disjoint union of a
continuously directed x, z path and a continuously directed z, y path for some
z. The up-down distance dG (x, y) is the minimum number of edges an x, y up-
down path may have, and +∞ if there is no such path. dG is reflexive and
symmetric, but, in general, does not satisfy the triangle inequality. For each
keyword k ∈ I(O) we define the following weighting function w(k):
w(k) := min {dG (k, q1 ), . . . , dG (k, qn )}
Root: Java Ontology
Programming Elements Programming Environment Programming Strategies

- Advanced Programming Concepts - JDK - Functional
-... -... -...
- Basic Programming Concepts - Java Programming Language - Imperative
- Classes and Objects - Features of Java -...
- Classes -... - Logical
- Class Body - keywords -...
- Class Declaration - implements
- Class Modifier - new
-... -...
- extends -...
- implements -...
-...
-...
- Inheritance
-...
- Interfaces
- implements
-...
-...
- Control Structures
-...
- Methods
- Class Method
- Instance method
- Method Body
-...
- Method Declaration
-...
-...
-...
Fig. 3. Part of the ontology for the Java domain knowledge
Example 3. Let qs be a search query, qs ={“implements”}, and let O1 , O2 be two

learning objects, I(O1 ) ={“keywords”,”interfaces”,”implements”}, and I(O2 ) =
{“class declaration”, “class body”, “implements” }. Then

k∈I(O1 ) w(k)
w(O1 ) :=
|I(O1 )|
= w(“keywords”) + w(“interfaces”) + w(“implements”)
1+1+0
= = 0, 6̄
3

k∈I(O2 ) w(k)
w(O2 ) :=
|I(O2 )|
= w(“class declaration”) + w(“method declaration”) +
w(“implements”)
1+5+0
= =2
3
As w(O1 ) < w(O2 ) we expect that object O1 will fit better to the users
information need than O2 . In a normal keyword-based query we can not further
distinguish between O1 and O2 .
4.2 Knowledge Estimation Based on Reading Sequences

We claim that structured materials deliver useful information with great po-
tential for adaptation. Here we will consider the structure “reading sequences”
which is a predefined sequence of learning objects defined by some author.
A reading sequence links together some closed set of learning objects. For
each element in the sequence we calculate the knowledge that a (hypothetic)
user has gained before starting with this element:
Let O1 , . . . , On be a sequence of learning objects in some corpus C and CV
be the set of topics contained in our ontology-based knowledge model. ∀i ∈
{1, . . . , n}: Mark all topics k ∈ CV with: ∃j, j < i, and k ∈ I(Oj ) as “read”
(marking all topics contained on the previous steps of the sequence). Whenever
we want to retrieve a learning object out of some sequence we compare the
required knowledge (as calculated above) with the user’s actual knowledge state
(as proposed in Sect. 3.1). This allows the estimation whether the user lacks
some knowledge to successfully work with this learning object, whether the user
has the required prerequisite knowledge to work with this object or whether
the user has already learned the content of the learning object. This enables
the adaptive presentation of learning objects covering prerequisite knowledge,
or the presentation of hints indicating that the content of some learning object
is assumed to be known to the user.
In case that we have one or more reading sequences which cover all docu-
ments of a corpus, we can make the following knowledge estimation: Let again
O1 , . . . , On be a sequence of learning objects in some corpus C and CV be
the set of topics contained in our ontology-based knowledge model. Then ∀i ∈
{1, . . . , n}:
1. Mark all topics q ∈ CV with: ∃j, j < i, and q ∈ I(Oj ) as “read” (marking all
topics contained on the previous steps of the sequence)
2. Mark all topics q ∈ CV with ∃j, j ≤ n, and q ∈ I(Oj ) as “contained” (marking
all topics contained in the overall sequence).
3. ∀k ∈ CV : Estimate the knowledge of a (hypothetic) user who has followed
the sequence by
# “read” topics in the subgraph of k
knowledge(k) =
# “contained” topics in the k-subgraph
Here the k-subgraph is the subgraph of the ontology containing the node
“k” and all its children.
The above stated calculations are statical and user independent and allow
us to estimate the knowledge which is required to successfully learn each learn-
ing object of the sequence. This necessary prerequisite knowledge can then be
compared to a user’s actual knowledge state.
5 Conclusion and Current Work

In this paper we discussed the use of ontologies for user and knowledge modeling
in open adaptive hypermedia systems (OAHS). Our approach settles on extract-
ing as much information as possible from standard metadata descriptions like
SCORM as resources for adaptation in OAHS. This avoids to burden the authors
of learning objects with too extensive descriptions of the objects. In addition we
expect these standard metadata to be widely used for describing learning ob-
jects. Exemplary we described improved retrieval of learning objects in open
learning repositories and the derivation of learning dependencies by investigat-
ing reading sequences. Current work concentrates on setting up an interface to
the EDUTELLA framework to provide personalized access to learning objects.
References
1. Lom: Draft Standard for Learning Object Metadata. http://www.imsglobal.org/.
2. K. M. Anderson and S. A. Sherba. Using open hypermedia to support information
integration. In Proceedings of OHS7 - the 7th International Workshop on Open
Hypermedia Systems, held in conjunction with Hypertext 2001, Danmark, 2001.
3. Ariadne: Alliance of remote instructional authoring and distributions networks for
europe, 2001. http://ariadne.unil.ch/.
4. P. Brusilovsky. Methods and techniques of adaptive hypermedia. User Modeling
and User Adapted Interaction, 6(2-3):87–129, 1996.
5. M. Campione and K. Walrath. The Java Tutorial. Addision-Wesley.
http://java.sun.com/docs/books/tutorial/.
6. L. Carr, S. Bechhofer, C. Goble, and W. Hall. Conceptual linking: Ontology-based
open hypermedia. In Proceedings of the Tenth International World Wide Web
Conference, Hongkong, May 2001.
7. Draft Standard for Learning Object Metadata.
http://ltsc.ieee.org/wg12/doc.html.
8. Edutella, 2001. http://edutella.jxta.org/.
9. K. Grønbæk and R. H. Trigg. From Web to Workplace: Designing Open Hyperme-
dia Systems. The MIT Press, 1999.
10. T. R. Gruber. A translation approach to portable ontology specifications. Knowl-
edge Acquisition, 5:199–220, 1993.
11. N. Henze and W. Nejdl. Extendible adaptive hypermedia courseware: Integrating
different courses and web material. In Proccedings of the International Conference
on Adaptive Hypermedia and Adaptive Web-Based Systems (AH 2000), Trento,
Italy, 2000.
12. N. Henze and W. Nejdl. Adaptation in open corpus hypermedia. IJAIED Special
Issue on Adaptive and Intelligent Web-Based Systems, 12, 2001.
13. J. Park, J. Gennari, and M. Musen. Mapping for reuse in knowledge-based systems.
In Eleventh Workshop on Knowledge Acquisition, Modeling and Management, 1998.
http://ksi.cpsc.ucalgary.ca/KAW/KAW98/KAW98Proc.html.
14. R. Rada. Interactive Media. Springer, 1995.
15. Scorm: The sharable content object reference model, 2001.
http://www.adlnet.org/Scorm/scorm.cfm.
Adaptive Navigation for Learners
in Hypermedia Is Scaffolded Navigation
Roland Hübscher1 and Sadhana Puntambekar2

1
Department of Computer Science and Software Engineering, Auburn University,
107 Dunstan Hall, Auburn, AL 36849-5347, U.S.A.,
[email protected],
http://iis.cse.eng.auburn.edu/˜roland/
2
School of Education, Program in Instructional Media and Technology, University of
Connecticut, U-64, 249 Glenbrook Ave., Storrs, CT 06269-2004, U.S.A.,
[email protected],
http://www.sp.uconn.edu/˜sadhana/
Abstract. Adaptive navigation support can be of great help in large

hypermedia systems supporting learners as well as users searching for
specific information. A wide variety of adaptive mechanisms have been
implemented in existing adaptive hypermedia systems that provide bet-
ter and better suggestions to the user what hyperlinks to follow. We
suggest that adaptive navigation support should scaffold a learner in an
educational hypermedia system to select the appropriate links. We show
that this implies that selecting a link is an educationally relevant activity
that should not always be reduced to a trivial task by powerful adaptive
mechanisms. It follows that learners require sometimes different kinds of
adaptive navigation support than users looking for information. Finally,
we will suggest how to extend current mechanisms to provide scaffolded
navigation support to learners.
1 Introduction
With the advent of the World-Wide Web, hypermedia systems have become
a widely used and dominating way of providing information and educational
content to users. Hypermedia has been recognized as having great potential
in providing content to learners because relationships between concepts can be
made explicit with hyperlinks, and the same material can be organized along
different dimensions presenting the material to be learned from different views
[1]. The non-linear nature of hypertext environments offers opportunities as well
as certain difficulties for learning, thus making the design of such systems both
complex and challenging. The flexible nature of hypertext makes it necessary
for designers to provide learners with some kind of navigational support. Re-
searchers believe that learning from hypertext puts a greater cognitive load on
learners [2]. Readers have to acquire specific strategies such as knowing where
they are, deciding where to go next and building a cognitive representation of the
network structure, in order to cope with the specific constraints of a non-linear

Adaptive Navigation for Learners in Hypermedia Is Scaffolded Navigation 185
presentation. Hyperlinks allow each individual learner to traverse and explore

the content in a way that fits his or her interests and learning goals at any par-
ticular time. In a hypertext system, the reader is actively engaged in creating
both meaning and structure.
The reader constantly makes decisions about where to go next. However, this
added flexibility, compared to books which are often read in a more or less linear
fashion, can also cause problems seriously impacting the pedagogical benefits of
a hypermedia system. It is quite easy for the learner to lose orientation and
therefore not knowing how the current page fits into the big picture and what
hyperlinked path to follow.
Adaptive hypermedia attempts to solve these problems by individualizing
the presentation of the content for each user and by providing personalized nav-
igation support. The goal of both approaches, personalized presentation and
navigation, is to reduce the cognitive load for the user. Learners with different
goals and knowledge may be interested in different information or prefer exam-
ples from different domains. They also may want navigate along different paths
depending on their goals and prior knowledge. This is especially important in
educational systems to make sure the learner sees those content relevant to his
or her learning goals. Furthermore, adaptive navigation can help prevent users
from getting lost in large systems.
Adaptive hypermedia systems (AHS) approach these problems with adaptive
presentation of the information and adaptive navigation. Adaptive presentation
mechanisms decide how the information is delivered most appropriately to the
user considering factors like user expertise, media preferences, learning goals,
interests, and so forth stored in a user model [3]. The user model is learned by
the AHS over time using various information sources like questionnaires, quizzes,
navigation behavior of the user, and others. The basic idea of adaptive navigation
is to constrain the options of where to go next to a smaller set of relevant choices
making sure the user will end up at a hypermedia page that is appropriate to
his or her task.
Examples of navigation support are the “next” or “continue” links which are,
according to the AHS, the best and possibly only choice [4], sorted lists of links
where earlier links are better choices [5], annotated links reflecting some impor-
tant status like ‘the user is ready to follow it or not’ or ‘it leads to redundant or
irrelevant material’ [6], or link hiding where only those links are made accessible
that the system considers relevant to the user [7].
Each of these mechanisms is based on a user model which is a description of
the user’s preferences, knowledge, skills, characteristics, learning goals, etc. and
a pedagogical model that suggests what the user should visit under what cir-
cumstances [8]. Too often, the pedagogical model is rather implicit and built into
the adaptive algorithm [9]. Much of the current research focusses on providing
better user models and better algorithms for providing appropriate navigation
support by ordering the content according to some scheme.
In this paper, we would like to revisit the assumption that the better sug-
gestions we can give to the user, the better the system serves its purpose. Of
186 R. Hübscher and S. Puntambekar
course, on the one hand, it is obvious that the better the adaptive mechanisms
are, the better the system is. On the other hand, it is not obvious that more ac-
curate advice is indeed better for a user—or, more specifically, a learner—under
all circumstances.
This paper is not about what can and cannot be done. It is about how adap-
tive navigation should be designed for specific users, especially for learners. We
do not need to propose new mechanisms—many good ones already exist. How-
ever, we want to make explicit that efficiently getting to the right page should
often not be the overriding goal of adaptive navigation support for learners.
Consider the following hypothetical AHS. Let’s assume that the user model
and adaptive mechanisms are so good that the system can almost always pro-
vide exactly one link to the page that is best for the user to visit next. In an
information-seeking task, this is of course great because the user will find im-
mediately what he was looking for, possibly without ever thinking about what
he was looking for at all. But is this also a good system for a learner? We will
come back to this question at the end of this paper.
2 Adaptive Navigation
Throughout this paper, we will consider two types of tasks—learning and infor-
mation seeking–that cover a large class of activities AHS are used for. A learner
who uses an AHS to learn about a certain concept executes a learning task. It is
the system’s goal to make sure that the learner will understand well enough the
concept and all the necessary prerequisite concepts. On the other hand, if the
user is only interested in finding the relevant information, then we consider the
task to be supported by the AHS an information-seeking task. Many systems
support both types of tasks, however, it is not necessarily the case that both
types of tasks should get the same adaptive navigation support.
Next, we will review existing adaptive hypermedia mechanisms. Then we will
discuss what navigation support information-seeking and learning tasks require.
We will argue that the objectively best suggestions the adaptive mechanism can
provide are not necessarily the ones from which learners benefit the most.
2.1 Adaptive Mechanisms
Adaptive navigation deals with the problem of the user having to select a link
among the many possibilities. Often, there are too many possibilities and it is
very difficult for the user to choose an appropriate link. Adaptive navigation
reduces the number of choices using various mechanims. Some of the most fre-
quently used ones are link ordering, link hiding, link annotation and the use of
the “next” link [10]. All these mechanisms constrain how and from how many
links the user can choose to go to the next page in the hypermedia system.
The most restrictive mechanism is the “next” or “continue” link that the
AHS recommends as leading to the most relevant next page. This allows the
user to turn to the next page as easily as in a book, i.e., there is no need to think
about where to go next. If this were the only link provided, the hypermedia
system would be reduced to an individual, though completely linear structure,
indeed even more linear than a textbook. However, in combination with other
mechanisms, as for instance in InterBook [11], the “next” link can be useful.
The previous idea can be somewhat relaxed and instead of suggesting just
one link, an ordered list of links can be provided where the first link is the most
relevant for the user, according to the system, and the last one the least relevant.
This approach provides some choice to the user, but in return, requires him or
her to make a conscious decision which link to choose. However, Guzdial found
evidence in WebCAMILE, a web-based collaboration system, that users tend to
choose the first item in a list simply because it is the first item in the list [12].
In this case, and no real decision has to be made. Thus, the sorted list may not
be that qualitatively different to “next” link approach.
Another widely used approach is link hiding where links that should not be
followed are simply hidden, e.g., on a WWW page by presenting the link text
as simple text [13]. This approach has the advantage that it does not impose an
external ordering on the links, yet the binary approach—a link is either hidden
or visible—does not allow for great expressiveness. Therefore, this approach was
extended resulting in annotated links, another way to support learners.
Annotated links are hyperlinks that are further tagged with some information
that can be taken advantage of by the user to decide whether to follow it or not.
For instance, the tag can provide further information about the difficulty of or
relevance of the information on the page the link leads to [6]. This allows the
user to make a more informed decision about which link to follow and it does
require less trial-and-error behavior.
All these adaptive mechanisms constrain the number of choices for the learner.
The less choice, the easier the decision will be for the user. The annotated-link
approach, however, does not just reduce the set of choices but provides further
information to the user supporting the decision making process to choose a good
next hypermedia page, i.e., a page relevant to the current goals.
2.2 Navigation Support for Information Seekers and Learners
Now, which one of those mechanisms is the best? This question needs to be
answered from two directions. First, empirical evidence needs to be collected
and second, the tasks the adaptive navigation is supposed to support need to be
analyzed to understand what kind of support they require. This paper focuses
on the latter of the two issues where we look at the tasks of learning and of
finding information.
An information seeker is a person who is interested in quickly and easily
finding relevant information. Adaptive hypermedia can be extremely helpful if
the user model is accurate. In this case it can make sure that the user is never
confronted with irrelevant information and thus, the search space that the user
has to explore to find the information can be made quite small. The more ac-
curate the user model, the more powerful the adaptive algorithms and the more
constraining the mechanisms, the better the system will support the user. This
is consistent with many AHS’s approaches.
The case where the user is a learner is more complicated, but extremely
important because many AHS target learners. Of course, many educational sys-
tems must also support information-seeking tasks. A learner not only needs to
understand concepts and principles but also has to place the new knowledge in
context, often requiring prerequisites that will enable her to understand the cur-
rent concepts. This definition can be extended such that it is not necessarily the
user’s goal to learn but whoever made the learner use the hypermedia system.
Supporting a learning task requires the learner to focus much more on the
process of acquiring the necessary knowledge and skills than just on finding
a certain location in the hypermedia system. The learning task, especially the
examples and practice exercises, should be as authentic as possible, i.e., relevant
to the learner’s interests. In addition to the concept to be learned, the learner also
needs to be provided with all the prerequisite knowledge that she has not learned
yet. Furthermore, the order in which these different concepts, examples and
exercises are visited is dependent on the employed teaching methodology, which
in turn may be dependent on the learner’s history and the material to be learned.
Although the learning process should not be dragged out forever, minimization
of time is not as important as when looking for information. Furthermore, the
learner needs to be challenged, yet nor frustrated by the difficulty of the material.
Of course, this will cause some decent cognitive load for the learner without which
learning could not occur. This does not imply that AHS need to be designed to
make them difficult to use. However, the learner must be encouraged to make
non-trivial decisions and reflect, for instance on what he or she is learning, how
this ties into previous knowledge, and why it is useful for accomplishing the
actual learning goal. In other words, adaptive navigation for learners should be
viewed as scaffolded navigation.
2.3 Scaffolded Navigation
Scaffolding in the context of learning has originally been defined by as an “adult

controlling those elements of the task that are essentially beyond the learner’s
capacity, thus permitting him to concentrate upon and complete only those
elements that are within his range of competence” [14]. Scaffolding has been
linked to the work of soviet psychologist Lev Vygotsky, although he never used
the term scaffolding. According to Vygotsky, a novice learns with an expert,
and learning occurs within the novice’s Zone of Proximal Development (ZPD).
ZPD is defined as the “distance between the child’s actual developmental level
as determined by independent problem solving and the higher level of potential
development as determined through problem solving under adult guidance and
in collaboration with more capable peers” [15]. Enabling the learner to bridge
this gap between the actual and the potential depends on the resources or the
kind of support that is provided. Instruction in the ZPD can therefore be viewed
as taking the form of providing assistance or scaffolding, enabling a child or a
novice to solve a problem, carry out a task or achieve a goal “which would be
beyond his unassisted efforts” [14].
Adaptive navigation can be viewed as a method to provide scaffolding to a
learner navigating through a large hypermedia system. Scaffolding in this context
then implies the following among other things [16, 17].
– The learner must be aware of and interested in the goal of the learning
activity.
– Continuous assessment of the learner needs to be used to calibrate the sup-
port.
– Scaffolding fades away over time and the learner must take control of the
task.
– The learner needs to be actively involved in the learning process.
We will now discuss each of these characteristics with respect to adaptive

navigation support in adaptive hypermedia systems.
The learner must be aware of and interested in the goal of the learning ac-
tivity. The learner needs to have a choice of selecting links that deal with more
authentic problems. For instance, it is important that the user selects the ex-
ample that he or she is more familiar with, if there is more than one example
illustrating the same point. An AHS can accomplish this by providing all the
examples illustrating the same point and the user can choose, or, if the user
model contains information about the preferences, it can simply provide the one
example it knows the learner is interested in. The latter solution is more efficient
whereas the former requires and allows the learner to make a decision. It is im-
portant that the learner has to make a decision if there is a potential educational
benefit in doing so.
Continuous assessment of the learner needs to be used to calibrate the support.
The difficulty of the decision which link to select needs to be adapted to the
learner so that it is in the ZPD. In other words, providing a set of possibly
annotated links needs to be viewed as posing a problem to the learner that the
learner should be able to solve, yet it must not be too simple to be pointless
activity. This implies that making it as simple to the learner as possible to select
a correct next page should not be the goal. It is important for a learner to figure
out what knowledge, information sources, case studies, etc. are relevant. For
instance, in Problem-Based Learning, finding out what the relevant knowledge
to learn is, is at the heart of this learning methodology [18, 19].
Scaffolding fades away over time and the learner must take control of the
task. A very important part of scaffolding is that it fades away as the learner
learns to select the proper next page. A user model that takes the learner’s
behavior into account can take care of this part automatically. Since the basic
idea of adaptive navigation is to constrain the learner’s options, it implies that
the constraints need to be relaxed as the learner gains more expertise. Of course,
these constraints must not be relaxed to violate usability guidelines.
The learner needs to be actively involved in the learning process. This implies
that the learner should not be spoon-fed with the correct choices, but should
be actively involved in the decision process of where to go next. This adds yet
another strong argument against over-constraining the choices for the user. The
learner should not just passively follow the lead of the system but should make
the decisions, or should at least, over time assume control.
Several important observations can be made. First, fading scaffolding maps
quite well to the adaptive nature of AHS. Second, selecting a link should be
regarded as a problem to be solved if the decision is educationally relevant.
And third, reducing the difficulty of selecting the right link as much as possible
is not always the right thing to do. Actually, it is important that the learner
is allowed to make mistakes for two reasons: it provides him- or herself some
valuable feedback and the AHS can use it to improve the user model based on
the learner’s apparent misconception.
The central observation is that selecting a link is sometimes an educationally
relevant task that needs to be scaffolded. Therefore, making it as easy as possible
is not adequate in such a situation. Selecting the link needs to require the learner
to reflect upon what is important to learn or read about next. How will this
information at the other end of the hyperlink help with respect to the learning
goals? We therefore suggest that the link-selection task needs to be put into a
context that supports the learner to make the right decision.
We suggest that the context consists of prompts and questions that are adap-
tively selected together with the set of links. Assume, the learner is working on
a problem where she has to figure out when an object hits the ground when
dropped from a certain height. Now, should we provide just links related to ve-
locity and acceleration or also links to mass? An AHS will recommend against
going to the pages about mass as they are irrelevant for this problem. However,
we provide both sets of links plus the contextual prompt “Does a heavy item fall
faster than a light one?”. Of course, the question needs to be phrased so that
the learner can answer it, i.e., it must be at the right level with respect to the
ZPD.
3 Conclusions
We have made the argument that adaptive navigation for learners in hypermedia
is scaffolded navigation. We have provided a largely theoretical argument whose
implications are backed up by the empirically well-established educational value
of scaffolding [20, 21].
Adaptive hypermedia systems use adaptive navigation mechanisms to sup-
port the users to find their way around in large hypermedia systems. They make
sure that the users find the relevant information and that learners are exposed
to all the relevant concepts to understand a certain goal concept. However, it
is important to keep in mind that a learner is not just interested in efficiently
getting to the page describing the goal concept. The learner needs work hard to
understand all the prerequisite knowledge and learn these concepts in some ped-
agogically appropriate order. Reducing the information seeker’s cognitive load is
great, however, doing the same for the learner is not always beneficial to him or
her. This implies that the use of too much navigation support can be detrimental
to the learner because it frees him or her up from thinking [22].
We therefore recommend that adaptive navigation support be separated into
two categories. The first category deals with navigation simply to get to a certain
location in the system as quickly and easily as possible, i.e., this is navigation
for navigation’s sake. The second category deals with educationally related deci-
sion problems where the learner needs to decide what information source, what
concept, what method, and so on, would be more relevant for the current task
at hand. In this case, the learner needs to be scaffolded so that he or she learns
to make the right decisions, i.e., chooses the relevant information sources, tools,
etc. We accomplish this annotating the adaptively selected links with appropri-
ate prompts and questions scaffolding the learner to make a good choice.
Let’s revisit the hypothetical system suggested earlier. We assumed that the
user model and adaptive mechanisms are so good that the system can almost
always provide exactly one link to the page that is best for the user to visit
next. This is potentially a great system for an information seeker. However, if
the system is supposed to support learners, then this hypothetical system boils
down to an individualized book that requires to be read in exactly one order. This
way we lose most of the educational advantages of hypermedia. Furthermore, the
learner is not required at all to reflect on why she is reading what she is reading,
she has no opportunity to make mistakes and recognize them as useful feedback.
In other words, a system that is optimal in some “adaptive sense” is surely not
that useful for a learner anymore.
References
1. Rand J. Spiro, Paul J. Feltovich, Michael J. Jacobson, and Richard L. Coulson.
Cognitive flexibility, constructivism, and hypertext: Random access instruction for
advanced knowledge acquisition in ill-structured domains. Educational Technology,
May:24–33, 1991.
2. P. Wright. Cognitive overheads and prostheses: Some issues in evaluating hyper-
texts. In R. Furuta and D. Slotts, editors, ACM Conference on Hypertext, pages
1–12. ACM Press, 1991.
3. Peter Brusilovsky. Adaptive educational systems on the World-Wide-Web: A re-
view of available technologies. In 4th International Conference in Intelligent Tu-
toring Systems, San Antonio, TX, 1998.
4. Elmar Schwartz, Peter Brusilovsky, and Gerhard Weber. World-wide intelligent
textbooks. In ED-TELECOM’96 - World Conference on Educational Telecommu-
nications, pages 302–307, 1996.
5. Peter Brusilovsky, Elmar Schwarz, and Gerhard Weber. Elm-art: An intelligent
tutoring system on World Wide Web. In C. Frasson, G. Gauthier, and A. Lesgold,
editors, Intelligent Tutoring Systems (Lecture Notes in Computer Science), volume
1086, pages 261–269. Springer Verlag, Berlin, 1996.
6. Gerhard Weber and Markus Specht. User modeling and adaptive navigation sup-
port in WWW-based totoring systems. In A. Jameson, C. Paris, and C. Tasso,
editors, User Modeling: Proceedings of the Sixth International Conference (UM97),
pages 289–300, Cagliari, Italy, 1997. Springer Verlag.
7. Paul De Bra. Design issues in adaptive web-site development. In 2nd Workshop

on Adaptive Systems and User Modeling on the WWW, pages 29–39, Banff, 1999.
8. Paul De Bra, Geert-Jan Houben, and Hongjing Wu. AHAM: a dexter-based refer-
ence model for adaptive hypermedia. In Proceedings of the 10th ACM Conference
on Hypertext and Hypermedia, pages 147–156, 1999.
9. Roland Hübscher. What’s in a prerequisite. In International Conference on Ad-
vanced Learning Technology (ICALT 2001), Madison, 2001.
10. Peter Brusilovsky. Efficient techniques for adaptive hypermedia. In C. Nicholas and
J Mayfield, editors, Intelligent hypertext: Advanced techniques for the World Wide
Web, volume 1326 of Lecture Notes in Computer Science, pages 12–30. Springer
Verlag, Berlin, 1997.
11. Peter Brusilovsky, Elmar Schwartz, and Gerhard Weber. A tool for developing
adaptive electronic textbooks on www. In WebNet-96, San Francisco, 1996.
12. Mark Guzdial, Cindy E. Hmelo, Roland Hübscher, Kris Nagel, Wendy C. Newstet-
ter, Sadhana Puntambekar, Amnon Shabo, Jennifer Turns, and Janet L. Kolodner.
Integrating and guiding collaborations: Lessons learned in computer-supported col-
laborative learning research at Georgia Tech. In Computer Support for Collabora-
tive Learning, Toronto, 1997.
13. Paul De Bra and L. Calvi. AHA! An open adaptive hypermedia architecture. The
New Review of Hypermedia and Multimedia, 4:115–139, 1998.
14. D. Wood, J. Bruner, and G. Ross. The role of tutoring in problem solving. Journal
of child psychology and psychiatry, 17:89–100, 1976.
15. L. S. Vygotsky. Mind in Society. Harvard University Press, Cambridge, 1978.
16. A. S. Palincsar. Keeping the metaphor of scaffolding fresh - a response to C.
Addison Stone’s “the metaphor of scaffolding: Its utility for the field of learning
disabilities”. Journal of Learning Disabilities, 31(4):370–373, 1998.
17. C. A. Stone. The metaphor of scaffolding: Its utility for the field of learning
disabilities. Journal of Learning Disabilities, 31(4):344–364, 1998.
18. Howard S. Barrows. How to Design a Problem Based Curriculum for the Preclinical
Years. Springer Verlag, New York, NY, 1985.
19. Roland Hübscher, Cindy E. Hmelo, N. Hari Narayanan, Mark Guzdial, and Janet L.
Kolodner. Mcbagel: A shared and structured electronic workspace for problem-
based learning. In Second International Conference on the Learning Sciences,
Evanston, IL, 1996.
20. A. L. Brown and A. S. Palincsar. Reciprocal teaching of comprehension strate-
gies: A natural history of one program for enhancing learning. In J. D. Day and
J. G. Borkowski, editors, Intelligence and Exceptionality: New directions for theory,
assessment, and instructional practice. Ablex, Norwood, NJ, 1987.
21. A. S. Palincsar and A. L. Brown. Reciprocal teaching of comprehension-fostering
and monitoring activities. Cognitition and Instruction, 1(2):117–175, 1984.
22. Roland Hübscher and Sadhana Puntambekar. Navigation support for learners in
hypertext systems: Is more indeed better? In Johanna D. Moore, Carol Luckhardt
Redfield, and W. Lewis Johnson, editors, Artificial Intelligence in Education: AI-
ED in the Wired and Wireless Future, pages 13–20. IOS Press, 2001.
Pros and Cons of Controllability: An Empirical Study
Anthony Jameson1 and Eric Schwarzkopf2,

1
DFKI / International University in Germany
2
DFKI, the German Research Center for Artificial Intelligence
Abstract. A key usability issue with systems that adapt to their users is control-
lability: the ability of the user to determine the nature and timing of the adaptation.
This paper presents an empirical study of the trade-offs involved in an attempt to
ensure a suitable degree of controllability. Within an adaptive hotlist for a con-
ference web site, two mechanisms for providing users with recommendations of
conference events were compared: automatic vs. controlled updating of recom-
mendations. In an experimental setting, each of 18 users worked with both variants
of the adaptive hotlist, as well as with a nonadaptive variant. The users differed
markedly in their responses to automatic vs. controlled updating. A number of
reasons for these differences could be found in the objective and subjective data
yielded by the study. The study illustrates how preferences for different forms of
user control can be influenced by factors ranging from stable individual differ-
ences to unpredictable features of a situation. General implications for the design
of controllable adaptive systems are discussed.
1 Is Maximal Controllability Always Best?

One of the main usability issues in connection with systems that adapt to their users
concerns controllability. For example, Norman [5] wrote in an influential article: “An
important psychological aspect of people’s comfort with their activities—all of their ac-
tivities, from social relations, to jobs, to their interaction with technology—is the feeling
of control they have over these activities and their personal lives” (p. 69). The relation-
ships between controllability and other usability issues with user-adaptive systems have
been discussed by Wexelblat and Maes [8] and by Jameson [3, section 4].
One plausible policy for designers of user-adaptive systems is to give users maximal
control over all aspects of system adaptation. For example, Kay [4] discusses many ways
in which learners can be given control over learner-adapted teaching systems. But as
Kay also points out (p. 121), simply maximizing control in all respects may not always
be the best policy: Some users may have generally less desire for control than others;
and making the user exercise too much control may lead to distraction and time-wasting.
Similarly, Trewin [7] discusses the controllability trade-offs involved with agents
that help users to configure aspects of an operating system or an input device such
as a keyboard. For example, a physically impaired user may not be able to operate a
keyboard even well enough to initiate and control the configuration process; so fully

The research described here is being supported by the German Ministry of Education and
Research (BMB+F) under grant 01 IW 001 (project Miau). Björn Mittelsdorf and Marie
Norlien made essential contributions to the empirical study.
194 A. Jameson and E. Schwarzkopf
Recommendations for Hotlist: Update

Sun DC Patrick Gebhard Enhancing Embodied Intelligent Agents View Session Remove
14:10−14:30 With Affective User Modelling
Mon Paper Neal Lesh, Charles Rich, Collaborating with Focused and View Session Accept or
11:00−11:30 Candace L. Sidner Unfocused Users Under Imperfect Reject
Communication
Mon Poster Piotr J. Gmytrasiewicz, Emotions and Personality in Agent View Session Remove
13:30−15:30 Christine L. Lisetti Design and Modelling
Mon Poster Detlef Küpper, Alfred Kobsa User-Tailored Plan Presentation View Session Accept or
13:30−15:30 Reject
Fig. 1. Example hotlist from the UM 2001 web site.

(The first and third entries were added by the user; the second and fourth, shown in the system
in a red font, are recommendations made by the system. Italicized words represent hyperlinks,
referred to in this article as “buttons”.)
autonomous system adaptation may actually give the user more control overall than if
she were required to control the configuration process.
Although there has been much discussion among researchers about controllability,
some of it quite heated, there is a dearth of systematically gathered evidence about
what users themselves think about these issues. The present study aims to provide such
evidence within the context of one particular adaptive hypermedia system. Section 2
introduces the system, Sect. 3 describes our empirical study, and Sect. 4 presents and
discusses the results.
2 The Hotlist and Its Recommendation Component

The Eighth International Conference on User Modeling, UM 2001, held in July of 2001,
was the latest in a biennial series of conferences concerning user-adaptive systems (see
[1]). It offered the following conference events: 3 invited talks, 3 tutorials, 19 full paper
presentations, 21 poster presentations, and 12 doctoral consortium presentations. The
conference web site (http://dfki.de/um2001) introduced a variety of adaptive features,
of which only the hotlist will be examined in the present study. The hotlist (see Fig. 1)
is basically a specialized bookmarking tool that helps a potential attendee put together
a list of personally relevant conference events. Once the user has explicitly added some
events to the hotlist, the system can insert a set of recommendations—essentially, a set
of similar events that this user might be interested in.
The recommendations are computed with a naive Bayes classifier (see, e.g., [6]),
using as features a set of 22 domain-specific key concepts such as Machine Learning.
If the user clicks on the View Session button for a recommended event to see its full
description, she will also see a simple “explanation” of the recommendation in terms of
the system’s estimates of the user’s interest in the individual key concepts associated with
the event. Further details concerning the recommendation mechanism must be omitted
here for reasons of space; they are not required for an understanding of the results that
will be presented below.
After evaluating a recommendation, the user can choose to Accept the recommen-
dation, making it into a normal hotlist entry; or to Reject it, causing it to disappear from
the hotlist.
Pros and Cons of Controllability: An Empirical Study 195
At various times, the system updates the set of recommendations: It removes any
recommendations currently in the hotlist and replaces them with a (perhaps overlapping)
set that is based on all of the user’s relevant actions so far. Different ways of controlling
this updating process were compared in our empirical study.
3 Empirical Study: Issues and Method

It would be possible to design a study to determine what type of control was really best
for users in the long run. But it is equally interesting to find out how users deal with and
respond to each system variant during an initial encounter of just a few minutes. After
all, users often briefly try out a system—or an option within a system—and decide on
the basis of a small sample of experience whether to continue using it.
Moreover, the goal of the empirical study is not to determine the accuracy or overall
utility of the hotlist recommendations. Instead, it is assumed (as will be confirmed) that
the recommendations have only modest accuracy, as is the case with many recommender
systems, because of the severely limited evidence on which they are based. The ques-
tion is: How much control do users want to have when dealing with these imperfect
recommendations?
Subjects
Subjects were 17 students and 1 recent graduate from Saarland University and the In-
ternational University in Germany. Only subjects were recruited whose major or minor
course of study had some affinity with the topic of user modeling (e.g., computer sci-
ence, information science, or psychology), so that the experimental task (to be described
below) would be motivating and manageable to them; but the large majority had little
or no specific knowledge of the field. The number hours per week that subjects reported
spending in the world-wide web averaged 12.9, with a standard deviation of 10.8. All
subjects were male. They received 15 German marks for their participation.
System Variants Studied

Three variants of the hotlist were used:
Controlled updating of recommendations. This is the variant shown in Fig. 1: The
user explicitly requests each update of the recommendations by clicking on the
Update button at the upper right.
Automatic updating of recommendations. In this variant there is no Update button;
the system updates the recommendations automatically whenever the user adds or
removes a hotlist event or accepts or rejects a recommendation.
No recommendations: In this variant, the user can use only the basic hotlist, adding
or removing events but receiving no recommendations.
In a within-subject design, each subject used all three variants of the hotlist, the order
of use being counterbalanced as is described below. This type of design was chosen over a
between-subject design because of (a) our expectation (confirmed during the study) that
individual differences would be very large; and (b) our desire to hear the comparative
Instructions (paraphrase): Example email message:

You are working as a research assistant at the From: Anna Reiter <[local email address]>
nearby research institute [name given]. A number Subject: What I’d like to see at UM 2001
of more senior researchers at this institute are
considering attending the UM 2001 conference, For me, the most important methodological approach in
which will take place 3 months from now. the area of user modeling is machine learning. Often,
methods from this category are applied in web-based
To spare these researchers the time of familiarizing systems, or in systems that select specific news stories
themselves with the conference site and program, for individual users. I’m not interested in these last two
the director has asked each of them to send you an types of application of machine learning.
email message in which they characterize the
topics that they are interested in. Anything that deals specifically with the improvement
of automobile safety would be especially interesting to
For each such message, your job is to build up a me.
list of relevant conference events, which you will
email back to the researcher in question. On the Another thing I’m interested in is systems that model
basis of this list, the researcher will decide whether some type of psychological state of the user, such as
he or she considers it worthwhile to attend the emotions or stress.
conference. Best regards, Anna Reiter
Fig. 2. Paraphrase of the key instructions (left) and one of the three fictitious email messages used
as a sketch of an interest profile (right).
comments of subjects who had experienced all three variants. Learning effects could
not be avoided with this design. But the counterbalancing measures described below
ensured that such learning effects could not lead to overall differences in the results for
the three variants; and we will also see that the observed differences among subjects are
not explainable in terms of learning effects.
Each subject spent only a limited amount of time with each variant: about 4 minutes
of introduction plus 7 minutes of measured use. A serious conference visitor might spend
considerably more time constructing a personal conference schedule. On the other hand,
the shorter amount of time seems typical of the time that a user might spend trying out
the hotlist recommendations before deciding whether to continue using them to create
a complete schedule.
Material
The experimental task assigned to the subjects was designed to overcome two obstacles:
1. Subjects have considerably less familiarity with the topic of the conference than
a potential conference visitor would typically have.
2. Because of the within-subject design, each subject has to search the conference
site with respect to three different configurations of interest.
The left-hand side of Fig. 2 summarizes the way in which the experimental task was
introduced to each subject; the right-hand side of the figure shows one of the three
fictitious email messages employed. Each of the three messages had a similar style and
structure, and it described interests for which it was approximately equally easy to find
relevant conference events. The interests expressed were in part strongly related to the
hotlist recommender concepts, but for the most part subjects had to look at the detailed
information about an event in order to decide whether it was really relevant. This situation
appears to be typical of the way in which real potential conference visitors use the hotlist.
Orders of Presentation
Each of the 6 possible orders of the 3 system variants was employed equally often (i.e.,
for 3 of the 18 subjects). Each of the 3 fictitious interest profiles was used equally often
in the 1st, 2nd, and 3rd temporal position and equally often together with each system
variant.
Procedure
Each subject participated individually with the guidance of an experimenter. In an intro-
ductory phase that lasted between 20 and 25 minutes, the experimenter explained that
the investigators had developed various methods for searching for information in a con-
ference web site and that they were interested in evaluating and improving them with a
view to possible use in other sites. The experimenter then summarized some basic ideas
of the field of user modeling and explained the fictitious situation. Using an example
email, the experimenter gave an explanation of the web site and the hotlist, frequently
stopping to allow the subject to try out the system’s functions.
In each of the three main trials, the subject first read one of the emails from a
hypothetical colleague and then was allowed 7 minutes to build up a hotlist for that
colleague, starting with the system initialized for a new user (with an empty hotlist).
In the system’s log files, a record was kept of all pages visited and all actions taken in
relation to the hotlist. The experimenter took notes on other observable aspects of the
subject’s behavior. At the end of the 7 minutes, the experimenter saved the hotlist to disk
in its printable form.
After the three main trials, the subject typed in answers to a number of questions
about his use of the system, some of which are discussed below. Finally, subjects were
asked for further comments during a debriefing.
Despite their lack of knowledge about user modeling, subjects reported no major
difficulties in understanding the fictitious interest profiles or in evaluating individual
events with regard to these profiles.
4 Results
4.1 Quantitative Objective Results
Although the key variable of interest is subjects’ subjective evaluation of the two types
of updating, some objective results will give us a general picture of the way in which
they used the hotlist.
Figure 3 shows that subjects were just about equally (un)successful with all three
variants in finding relevant events to add to the hotlist. The rather small overall number
of events found is consistent with the limited amount of time that subjects had available
to process each interest profile.
Figure 4 shows the differences between the two system variants that included rec-
ommendations, in terms of how the system presented recommendations and how users
responded to them. It is not surprising that automatic updating led to about 4 times as
8
Number of Entries
Highest
6 Legend
Mean
4
Lowest
2
No Automatic Controlled
Updating Updating Updating
0
Correct Entries Incorrect Entries
Fig. 3. Mean numbers of correct and incorrect entries in the final hotlists produced by subjects
with the three system variants.
(The maximum numbers of correct entries for the three interest profiles were 12, 14, and 18
respectively, but some correct entries were difficult to identify as such. The upper and lower ends
of the vertical line segment in the middle of each bar indicate the highest and lowest values,
respectively, that were found among the 18 values obtained for the 18 subjects.)
8 [Highest = 13]
6
Number
Updates Accepted Rejected Unprocessed

Recommendations Recommendations Recommendations
Fig. 4. Objective results concerning the appearance and processing of recommendations with
automatic and controlled updating.
(Legend as for Fig. 4.)
many updates of the hotlist, since it involves utilizing just about every opportunity for
an update.
Although subjects using controlled updating experienced only 30% as many hotlist
updates as those using automatic updating, they received 59% as many recommendations:
By the time they had gotten around to requesting an update, it was likely to contain more
new recommendations than a typical automatic update.
On the average, subjects accepted and rejected exactly the same number of recom-
mendations using the two variants. The big difference is that with automatic updating,
many more recommendations were never responded to explicitly at all (this difference,
shown in the right-hand pair of bars in Fig. 4, is highly significant by a Wilcoxon rank-sum
test: Z = −3.03, p < 0.01). As the logs confirm, in many cases these recommendations
were swept away by an automatic update after the subject had made some change to
another aspect of the hotlist.
Very
eager PQ L F A
to use
Quite
With automatic updating?
willing M GH B Legend
to use
Probably T Subject who worked first
would R N C with automatic updating
use
Would
not IJ
care T Subject who worked first
Probably with controlled updating
would DE
not use
Would
not O
use
Would
definitely K
not use
Would Would Probably Would Probably Quite Very

definitely not would not would willing eager
not use use not use care use to use to use
With controlled updating?
Fig. 5. Scattergram of subjects’ expressed preferences for two updating methods.

(Each letter represents the responses of one subject. With respect to each method, the question was:
“If you had to perform more searches, what would you prefer to do: use the recommendations (that
is, have them displayed, and follow up on at least some of them); or not use the recommendations
(that is, turn them off or not request them in the first place)?”.)
4.2 Attitudes Toward the System Variants

Figure 5 gives an overview of the subjects’ responses to two questions that were designed
to reveal (indirectly) their preferences for controlled vs. automatic updating. Given the
emphasis in previous literature on the importance of controllability, we might expect to
see a statistically significant tendency for subjects to prefer controlled updating. (With
a within-subject design involving 18 subjects, a moderately strong tendency could have
been detected.) Instead, the most important conclusion to be drawn from Fig. 5 is that
users responded in very different ways to the questions. Given a sufficiently large sample
of subjects, we could no doubt find a statistically significant preference for one type
of adaptation or the other. But it is more important to understand the reasons for the
differences in users’ responses, using all of the available types of data: their responses
to the rating scales in the questionnaire; the verbal comments that they typed into the
questionnaire and made during their work or during the debriefing; and the detailed
records of the interaction that can be found in the system logs. This type of analysis
inevitably involves qualitative interpretation.
Subjects Who Preferred Controlled Updating

Subject K (cf. Fig. 5) is typical of users who have a strong general desire to remain in
control of the interaction with a system. He wrote “I am used to updating information
manually” and “I hate having the information appear automatically”. K’s behavior, as
it is revealed by the log files, is consistent with his attitude: With controlled updating,
he requested 1 update of the recommendations and proceeded to accept or reject each
of the 3 recommendations that appeared. With automatic updating, he received 8 rec-
ommendations in 4 updates, and he was able to follow up on only 4 of them (1 per
update).
Subject A—the most successful subject of all in terms of the number of relevant events
found—showed an attitude and a strategy similar to that of K with controlled updating.
But unlike K, he was able to follow up on the recommendations equally thoroughly in the
automatic updating condition, accepting or rejecting 11 out of the 12 presented—simply
because the system happened to present only about 1 new recommendation after each
update. Consistent with this result, A expressed an equally strong willingness to work
with both system variants. He mentioned two advantages of automatic updating that will
be discussed below, and he stated that his true preference would be to switch back and
forth at will between the two variants.
Subject O had quite a different reason for preferring controlled updating: On the
whole he found the recommendations to be of little value, accepting only 1 of the
total of 5 that he received. Accordingly, his attitude toward both of the variants with
recommendations was relatively negative (cf. Fig. 5). But he was especially critical
of the variant with automatic updating, saying that the burden of having to read the
recommendations may be even greater than that of reading through the detailed event
descriptions. Note that a reasonable strategy is for the user to start paying attention
to the recommendations only when the user has reason to believe that the system’s
model has achieved a reasonable level of accuracy. In both system variants, the user can
indeed always decide whether to follow up on the recommendations; but with automatic
updating, the user pays a price for the recommendations even when he or she is ignoring
them, in terms of screen clutter and longer system response times.
Subjects Who Preferred Automatic Updating

The clearest preference for automatic updating was shown by subject P (see Fig. 5).
He volunteered the comment that “If you are not accustomed to press the update button
periodically or after a decision you just made, you’ll miss topics.”
Similarly, Subject L commented spontaneously on two advantages of automatic up-
dating: First, he found it “too time-consuming to press the button each time”. Second,
L appreciated the fact that the automatically generated recommendations always repre-
sented the system’s most up-to-date model of his interests.
Subject Q illustrated a somewhat different drawback of controlled updating: The
danger that the user may forget about updating entirely. Indeed, while using the variant
with controlled updating he had completely forgotten about recommendations, using
instead just the basic hotlist, as he himself noticed later.
Table 1. Summary of the potential advantages of each variant of the hotlist recommender that
came to light in the empirical study.
Potential advantage Precondition(s) for advantage to apply
Controlled updating:
1. The user’s feeling of control over the The user has a general desire to control interactions.
interaction with the system is enhanced.
2. The user can follow up on more than one The user receives relatively large, nonoverlapping sets
recommendation in a given set. of recommendations.
The user pursues the strategy of looking at all of the
recommendations in each set.
3. System response times can be faster Technical conditions make system response time an
because of less frequent updating. important factor.
The user would not choose to request an update at every
opportunity.
4. The user can restrict updates to situations The user can assess the likely accuracy of the system’s
in which the system’s model of her interests user model.
is assumed to have useful accuracy.
5. A smaller amount of irrelevant text The user finds recommendations distracting although
appears in the hotlist. they are clearly distinguishable from normal hotlist
entries − perhaps because of limited available screen
space.
Automatic updating:
1. The user is regularly reminded that new The user’s strategy does not provide for regular
recommendations are available. consideration of the recommendations.
The user has not yet learned that hotlist actions
typically result in new recommendations.
2. The user is spared the effort of clicking on The user’s hotlist−related actions are sufficiently
a button to obtain new recommendations. numerous that new recommendations are frequently
available.
3. The recommendations displayed always The accuracy of the system’s user model tends to
reflect the system’s most complete model of improve significantly with each modification to the
the user’s interests. hotlist.
4. The user cannot overlook the availability The user is not yet accustomed to using
of the recommendation feature. recommendations.
5 Discussion
Whenever a choice between controlled and automatic adaptation arises, each solution
is likely to have its own potential advantages over the other one. The specific poten-
tial advantages of automatic and controlled updating that emerged from our study are
summarized in Table 1.
As this table illustrates, the relative importance of each of these advantages may
depend on various types of conditions:
1. The nature of the application and of the adaptation involved.
2. Individual differences among users in terms of preferences, experience, and ways
of approaching the tasks in question.
3. Relatively stable contextual factors such as the speed of an internet connection.
4. Essentially random situational factors such as the nature of the information re-
trieved during a small number of search attempts.
One general design implication is that an attempt to deal with the controllability prob-
lem should begin with an analysis of the reasonably stable, predictable conditions that
are likely to be relevant. For example, Trewin [7] discusses different controllability
mechanisms that are appropriate for different types of configuration task.
A second approach to providing suitable controllability is to allow users to choose
the type of control that they desire (see, e.g., [8, “Issue 4”]). For example, if our hotlist
included a button for toggling between automatic and controlled updating, those users
who had a clear, strong preference for one type of updating might be quickly satisfied.
But a user cannot in general be expected to be able or willing to take into account all of
the relevant considerations (e.g., the entire set listed in Table 1).
To a certain extent, the factors identified as relevant can be taken into account by the
system itself. For example, our hotlist recommender could compute at any moment the
expected utility of an automatic update, taking into account factors such as the length
of the delay that would be caused by the update and the number of recommendations
in the hotlist that the user has not yet processed. The user could then be allowed to
set an expected utility threshold that must be exceeded before an automatic update is
performed. (A similar approach was realized in the Lumière prototype; cf. [2].)
Given the nature of the factors that tend to be involved, neither the designer nor
the user nor the system—nor all of them working together—will in general be able to
ensure that the right degree of controllability is available all of the time. It should be
anticipated that frustrations like those experienced by our subjects with respect to both
of the adaptive variants will in some cases occur; and the possibility should be taken
into account that they may cause a user to abandon a system entirely.
Although this last point sounds discouraging, taking into account the limited pre-
dictability of users’ behavior and responses may be an important step toward an adequate
solution of the problem of giving users appropriate control over adaptation.
References
1. Mathias Bauer, Piotr Gmytrasiewicz, and Julita Vassileva, editors. UM2001, User Modeling:
Proceedings of the Eighth International Conference. Springer, Berlin, 2001.
2. Eric Horvitz, Jack Breese, David Heckerman, David Hovel, and Koos Rommelse. The Lumière
project: Bayesian user modeling for inferring the goals and needs of software users. In Gre-
gory F. Cooper and Serafin Moral, editors, Uncertainty in Artificial Intelligence: Proceedings
of the Fourteenth Conference, pages 256–265. Morgan Kaufmann, San Francisco, 1998.
3. Anthony Jameson. Adaptive interfaces and agents. In Julie A. Jacko and Andrew Sears, editors,
Handbook of Human-Computer Interaction in Interactive Systems. Erlbaum, Mahwah, NJ,
2002. In press.
4. Judy Kay. Learner control. User Modeling and User-Adapted Interaction, 11:111–127, 2001.
5. Donald A. Norman. How might people interact with agents? Communications of the ACM,
37(7):68–71, 1994.
6. Michael Pazzani and Daniel Billsus. Learning and revising user profiles: The identification of
interesting web sites. Machine Learning, 27:313–331, 1997.
7. Shari Trewin. Configuration agents, control and privacy. In Proceedings of the ACM Conference
on Universal Usability, pages 9–16, Arlington, Virginia, U.S., 2000.
8. Alan Wexelblat and Pattie Maes. Issues for software agent UI. Unpublished manuscript,
available from http://wex.www.media.mit.edu/people/wex/, 1997.
Personis: A Server for User Models
Judy Kay, Bob Kummerfeld, and Piers Lauder

University of Sydney, Australia, 2006
{judy, bob, piers}@cs.usyd.edu.au
Abstract. A core element of an adaptive hypertext systems is the user

model. This paper describes Personis, a user model server. We describe
the architecture, design and implementation. We also describe the way
that it is intended to operate in conjunction with the rest of an adaptive
hypertext system.
A distinctive aspect of the Personis user model server follows from our
concern for making adaptive systems scrutable: these enable users to
see the details of the information held about them, the processes used to
gather it and the way that it is used to personalise an adaptive hypertext.
We describe how the architecture supports this.
The paper describes our evaluations of the current server. These indi-
cate that the approach and implementation provide a workable server for
small to medium sized user collections of information needed to adapt
the hypertext.
Keywords: Server for user profile/model, Security and Privacy of User

Models, User Modelling, Personalisation, User Control
1 Introduction
Adaptive hypertext relies on the availability of information about the user as

a foundation for its adaptivity. This paper describes our approach to building
a server for such user models. The motivation for such a server follows from
the nature of the information in a user model, the difficulties associated with
building good user models and the way that such models might be used in a
range of applications.
First consider the nature of user modelling information. Because it consti-
tutes personal data, it needs to be treated rather differently from other parts of
an adaptive hypertext system: it is subject to far tighter requirements for secu-
rity of the information. For systems to move out of the laboratory, it will have
to meet legal requirements such as the European Community Directive on Data
Protection2 It is in the spirit of such legislation that users be able to access and
control their own data. A server makes sense for the provision of the required
security at the same time as ensuring user access and control.
2
http : //www.doc.gov/ecommerce/eudir.htm (visited Jan 2002)

204 J. Kay, B. Kummerfeld, and P. Lauder
Another important problem for user modelling is that it takes considerable

time and effort to build up a detailed user model. When users first comes to an
adaptive hypertext system, they either have to accept a generic interface initially
or they have to provide information about themselves. A server should enable
the reuse of the user model across applications. In particular, suppose the user
explores one adaptive hypertext to do some substantial activity such as learning
how to program in C [1]. When they move to another adaptive hypertext system
that teaches Java, it would be useful for that system to be primed with the user
model that has already been built up.
For the most part, work on personalisation has placed the user model within
an application. For example, there have been several user modelling shells:
GUMS [2,3] and its successor, GUMAC [3]; BGP-MS [4] UMT [5] and in the
area of student modelling, TAGUS [6] These systems explored many issues in
building generic tools for managing user models. They did not provide a server
for reuse of user model information across applications. A recent review of gen-
eralised support for user modelling [7,8] concluded that we have yet to see a user
model server that addresses the needs for ensuring the user’s privacy, control
and ability to scrutinise their user model and the processes for personalisation.
There has been some work on user model servers. Orwant built a system [9]
with user models in a Lisp-like language. Orwant took care to encrypt the user
model during transmission. Paiva also built a server [10] which could support
multiple teaching agents. A rudimentary form of user model server is provided
by Hailstorm3 . This initiative is an indication of the recognition of the value of
a user model.
Interestingly, Hailstorm is described with a focus on user control: ‘It puts
users in control of their own data and information, protecting personal informa-
tion and providing a new level of ease of use and personalisation.’ It allows a
registered user to store personal data in a standard form with standard access
methods on a central server provided by Microsoft. Access to the information
is provided through a standard set of services using the Simple Object Access
Protocol (SOAP) and is based on the earlier ‘Passport’ system for storing user
names and passwords. It seems to be intended for broad tasks: address book,
email, diary, documents, device settings. It is claimed to give users control over
the data while retaining privacy. However, the architecture presents both a single
point of failure and a single point of security vulnerability. It is described as a
simple data store, without inferencing ability.
Several of these systems, including BGP-MS, UMT and TAGUS, had inter-
faces for the use of the developer to scrutinise the models as they built and
debugged them. However, the representation and reasoning mechanisms were
not designed for the user to scrutinise them.
The goal of the Personis project is to explore ways to support powerful and
flexible user modelling and at the same time to design it, from its foundations,
to be able to support user scrutiny and control. Our underlying representation
of the user model [11,12] collects evidence for each component of the user model.
3
http : //www.microsof t.com/net/hailstorm.asp (visited Jan 2002)
Personis: A Server for User Models 205
In large scale field testing [12] we have demonstrated that many users can and
do scrutinise their user models. In a relatively short eight-week study, some users
scrutinised the full range of details of the user model, including the meanings of
the components of the model, the details of evidence about each one, the details
of the evidence sources and the reasoning used to infer component values. In
this paper, we describe the way that the Personis server builds upon this foun-
dation to provide a user model server that can support scrutability for adaptive
hypertext systems.
In Sect. 2, we describe the architecture of the Personis server and the way that
the server fits into the architecture of an adaptive hypertext system. Section 3
gives an overview of its application in a simple recommender application. We
report results of some evaluations of the current Personis server in Sect. 4 and
Sect. 5 has discussion and conclusions.
2 Architecture of a Personis-Based
Adaptive Hypertext System
2.1 High Level Architecture of Several

Adaptive Hypertext Systems
Figure 1 shows the way that a Personis server can support reuse of the user
models over a series of adaptive hypertext systems. This illustrates four main
ideas that we now describe: the server itself; generic scrutiny tools that enable the
user to see and control their own user model; a collection of adaptive hypertext
applications; and the views which are the conceptual, high level elements shared
between the server and each application.
Personis
User Model
Server
view N
view 2
generic scrutiny view 1

interface AHA N scrutiny
interface
...
AHA 2 scrutiny
interface
AHA 1 scrutiny
interface
User
Fig. 1. Personis Server Architecture

In keeping with our previous work, our architecture includes provision for
tools for the user to scrutinise their user model. This might include an interface
to explore the details of each component of the model, as in the um toolkit
[12] and we envisage additional tools for exploring large user models [13] We do
not focus on the range of these tools in this paper. We do, however, note that
such tools are important to our approach. The figure shows the user interacting
with the generic scrutiny interface and this, in turn, interacting with the server.
In fact, as the figure suggests, interfaces that support user scrutiny of the user
model operate as a special type of adaptive interface. (These tools may well be
adaptive hypertext applications in their own right.)
Figure 1 shows each adaptive hypertext application with two parts: the core
of the adaptive hypertext which enables the user to do some task such as learn
to program; and, in addition, we show a scrutiny interface associated with that
adaptive hypertext application. This structure is important. If a user model
server were to be in practical use, we would expect that the user model for
each individual would steadily grow to be quite substantial. Although we are
committed to supporting the user’s scrutiny of that model, we expect that users
will generally want to explore their model in the context of their interaction with
a particular application. So, for example, the user might be using an adaptive
hypertext that teaches the programming language, C. As they do so, they might
wonder why it presented information in a particular way. They might also see
a friend using the same system and if its adaptation for that friend is different,
our user might want to explore why. In this type of user-scrutiny, the answers
to their questions will typically involve the interaction of the adaptive hypertext
application and the user model. So, it makes sense to provide support for the
user to scrutinise the adaptivity within the context of the adaptive hypertext
application.
Issues of scale and comprehensibility give another reason for supporting
scrutiny of the user model within the adaptive hypertext application. For the
case of the C hypertext, the user would probably be primarily interested in
those parts of the user model that are used by that application. Since this will
be a small part of a full user model, it is a more manageable and relevant as-
pect to explore. Our architecture requires that scrutability be supported in the
application.
The last element of Fig. 1 is the views of the user model available to each
adaptive hypertext application. For example, the leftmost application in the fig-
ure might need just a few components of the user model. Our architecture allows
the definition of a view that defines just these components. Another application
will typically use a different view. The application writer would define those
parts of the user model needed by their application and these would be defined
in views established for that application.
Importantly, these views have an interaction with the design of the access
control for the server. Personis allows the user to define just which applications
are allowed to see each part of the user model. The user can also control the
information sources that should be made available to each applications. So, for
example, the user model may contain evidence from several sources about the
user’s knowledge of programming. That user could decide to make only the
information from certain sources available to an application. Another user might
make a different decision. Access control information is stored with the user
model in the object database.
In particular, suppose an application like AHA1 in the figure teaches about
C and it collects data from the user’s answers to quiz questions. It provides
this to the user model as evidence about the user’s knowledge. Further, suppose
that AHA2 teaches about a somewhat related subject, Java programming. The
user can control whether AHA2 is allowed to access user model evidence that
was provided to AHA1. The user can also control just which components of the
model are available to AHA2. This means that if AHA2 requests information
about the user’s knowledge of Java, this will only be provided if the user has
made it available to AHA2.
2.2 Internal Architecture of Server
The internal architecture of a Personis server is depicted in Fig. 2. The user

model information is held as an object database. This stores the user model in
a representation that has the same conceptual foundation as the um toolkit [12]
The basic element is the component and each has an associated list of evidence.
Each piece of evidence is tagged with information about its source. The object
database also manages the structuring of user models into contexts which give
a hierarchical structuring of the component namespace. It also holds objects
which define the views of Fig. 1. Any context can define a view which includes
components from any part of the user model context hierarchy.
Client
resolver
Ask/Tell Client
res
Personis
olv
XML−RPC/SSL
OODB User Model
er
Server
HTTPS Management
interface
Fig. 2. Personis Server Architecture
The object database also holds the access control information. The establishes
access rights to for applications and users. Each aspect can be controlled at the
common levels of readable and/or writable.
The server needs to handle large numbers of clients simultaneously. To achieve

this, the current implementation is an object layer over a relational database.
Internally, objects in the database are referenced using an Object Identifier that
is resolved to a server address. This allows parts of user models to reside on
different servers. It is envisaged that users would want to keep most of their
personal model on a local system under their control. They can then choose to
allow selected information to be stored on servers outside their direct control.
As shown in Fig. 1, a view defines a collection of components and an appli-
cation can ask for the values of these components or send evidence about these
to the user model. As shown in Fig. 2, this communication between server and
client operates with remote method calls using the XML-RPC4 protocol over
SSL (secure socket layer).
In addition, the server provides a management interface implemented with
HTML, HTTP and SSL protocols and accessible from any web browser. The
management interface for the user model server has several functions. It provides
a status display for the running server and allows reconfiguration. It also allows
a suitably privileged user to create and manipulate user model definitions as
well as create new user models for individuals according to a previously created
definition. It also provides interfaces for setting and altering access control.
The user model server is designed to be scalable to a large number of users,
each with large models. Access to model objects by an application is extremely
simple with only one line of program code typically required to acquire a com-
plete set of component values corresponding to a view.
The remaining element of Fig. 2 is the resolvers. These follow the approach
of the um toolkit where the generic user model simply holds the uninterpreted
collection of evidence for each component. It is only at runtime that the appli-
cation uses a resolver to interpret the evidence available to it and conclude the
value of a component. A default resolver is available but specialised resolvers can
be associated with an application. So, as shown in Fig. 2, one client adaptive
hypertext system might use one resolver. Another client might use another and
so interpret the same component differently. To take a simple example, Resolver
A might treat the user’s self-assessment of their knowledge as highly reliable. Re-
solver B might give higher reliability to the user’s performance on quizzes. Then,
suppose Rebecca is a the user who is quite knowledgeable about C control struc-
tures as evidenced by quiz performance. Suppose that she lacks confidence and
rates her knowledge as low in this area. An application which uses Resolver A
will treat her as not knowing C control structures. One that uses Resolver B
will treat her as knowing them. Note that, Rebecca could decide that evidence
derived from her quiz results was not to be made available to either of these
applications: in that case, the server would not present evidence to the resolvers
and associated applications would operate as if Rebecca’s user model had no
quiz results.
4
www.xmlrpc.org
2.3 Application View of Personis

The basic API for the user model server is simple and elegant. Remote method
calls using the XML-RPC protocol (and possible future use of SOAP) provides
user model access to any application with a client side XML-RPC implementa-
tion. The basic client API consists of three calls:
um = access(odbname, user, password)
The user, password and odbname are strings. The system maps the odbname
to a server address.
components = um.ask(context, view, resolverident)
The context is a list of context names giving a path to the required context.
The view is either a simple string indicating a view name, or a list of names of
components. The resolver ident is the name of a resolver located at the server
to resolve the values. If the resolver ident is omitted a default is used.
um.tell(context, component, evidence)
The evidence is a list containing the type of the evidence, an optional expiry
time, and the value of the component.
An application can collect resolved values for a complete set of components,
as defined by a view, using an ask statement. Combined with a statement to
connect to the server and one to close the connection the entire interaction with
the server is three lines of code.
3 Overview of User Model Server

We now illustrate the architecture in an example application: a “Personal Jazz
Channel” that provides users with personalised jazz programmes.
Like many such systems, it asks the user to prime its user model by answering
a small set of carefully chosen question. This screen is shown in Fig. 3.
From this, the system sends a collection of evidence to Personis. That evi-
dence is tagged as being given by the user, via the Personal Jazz Channel query
interface.
Then, the PJC application makes inferences about other CDs, styles and
artists. Each of these generates a piece of evidence which is tagged as inferred
and added to the user model by the Personal Jazz Channel application. This
very simple process enables the system to have a rich set of preference data from
a small set of questions.
At this point, a personal streaming audio channel is created with a mix of
tracks conforming to the user’s modelled preferences. During normal interactions,
the Personal Jazz Channel adds new pieces of evidence about tracks the user
allows to play, those that are skipped and so on. Each time the user comes to
the Personal Jazz Channel, they log in and the system sets up a connection with
Fig. 3. Example rating screen for a personalised jazz channel
the Personis server and asks for the user model view it needs to perform its
personalisation.
The Personal Jazz Channel application interacts with the Personis server to
gather relevant model information. For example, to gather all the resolved values
for jazz styles the following lines of code connect to the server, retrieve the values
and close the connection:
um = access(odbname, user, password)
styles = um.ask(context=["music","jazz","styles"])
um.close()
In this case the context is ‘music->jazz->styles’ and the default view for
the context is all the components in that context. The um.ask call will return a
dictionary of resolved values. Each element of the dictionary is a tuple containing
the resolved value, the name of the resolver used and the time it was resolved.
The Personal Jazz Channel operates in the way illustrated in Fig. 1. In ad-
dition to its job as a typical customised application, it also has an interface to
support scrutability. At any time, the user is able to examine and change the
personal information held by the system by selecting the Profile button at the
top of the screen. This brings up a screen like that shown in Fig. 4.
From the point of view of our architecture, there are some important aspects
to point out. Firstly, we note that it is an essential aspect of the architecture that
the Personis server and each of its associated adaptive applications is loosely re-
lated. The authors of the application are responsible for it. Different applications
will be created by different people and will work differently. The user needs to
explicitly allow an application to access relevant views in the user model. The
user may decide that the security of some systems is effective enough that it is
acceptable to allow them access to substantial amounts of the user model. On
the other hand, a user may be less happy with the security protection in another
Fig. 4. Profile information: the user has selected styles from the list at the left and see
sees the user model components for styles of music.
system: that application might be authorised to access only a limited part of the
user model and only information derived from a few of the evidence sources.
4 Evaluation
The current server has had very little work to optimise performance but still
performs at an acceptable level. A test application performing 10000 accesses
achieved the following number of transactions per second:
tell (new value each time) 26/sec
ask (single component) 17/sec
ask (view with 2 components) 13/sec
Each of the ask operations used a default resolver (most recent given evidence
value).
The server was running on an 850Mhz Duron processor with 512MBytes
memory. The user model being accessed had approximately 50 components
arranged in 15 (sub)contexts. The database would have cached the complete
model after the first access and so the figures quoted show performance for the
client/server protocol and the internal data structure search and overhead.
Investigation has shown that the XML-RPC protocol is a major overhead.
This is mainly due to the poor performance of the general purpose XML parser
used in the implementation. We feel that significant performance gains can be
made with a hand crafted parser. We feel strongly that the use of a standard
protocol such as XML-RPC is warranted since it provides access to the Personis
server from a wide range of programming languages.
From the application programmers point of view the Personis server is very
easy to use. Only a handful of lines of code are required to retrieve the resolved
values of a sets of components. In a similar way, it is very straightforward to add

evidence to existing components.
5 Conclusion
The underlying design of the Personis user model server is based upon the pri-
mary requirement that users have access to their user model and control over
it. In addition, it has been designed to provide support for user modelling with
an elegant but powerful programmer interface. It is novel in its design being
explicitly focussed on user control and scrutability.
References
1. Kay, J., and Kummerfeld, R.J.: An individualised course for the C program-
ming language. Online Proceedings: http://www.ncsa.uiuc.edu/SDG/IT94/Pro-
ceedings/Educ/kummerfeld/kummerfeld.html , Elsevier (1994)
2. Finin, T. W.: GUMS - a general user modeling shell. In: Kobsa, A., Wahlster. W.
(eds.): User models in dialog systems. Springer-Verlag, Berlin (1989) 411–431
3. Kass, R.: Building a user model implicitly from a cooperative advisory dialog. User
Modeling and User-Adapted Interaction. 1 (1991) 203–258
4. Kobsa, A., and Pohl, W.: The user modeling shell system BGP-MS. User Modeling
and User-Adapted Interaction 4 (1995) 59–106
5. Brajnik, G., and C Tasso, C.: A shell for developing mon-monotonic user modeling
systems. International Journal of Human-Computer Studies 40 (1994) 36–62
6. Paiva, A., and Self, J.: TAGUS - a user and learner modeling workbench. User
Modeling and User-Adapted Interaction 4 (1995) 197–228.
7. Kobsa, A.: Generic User Modeling Systems. User Modeling and User-Adapted
Interaction - Ten Year Anniversary Issue 11 (2001) 49–63
8. Fink J., Kobsa, A.: A Review and Analysis of Commercial User Modeling Servers
for Personalization on the World Wide Web. User Modeling and User-Adapted
Interaction - Special Issue on Deployed User Modeling 10 (2000) 209–249
9. Orwant, J.: Heterogenous learning in the Doppelganger user modeling system. User
Modeling and User-Adapted Interaction 4 (1995) 59–106
10. Machado, I., Martins, A., Paiva, A.: One for all and all for one: a learner modelling
server in a multi-agent platform In: Kay, J (ed): User Modeling: Proceedings of
the Seventh International Conference, UM99. Springer Wien, New York (1999)
11. Kay, J.: Accretion representation for scrutable student modelling. In: Gauthier, G.,
Frasson, C., VanLehn, K. (eds.) Intelligent Tutoring Systems (2000) 514–523
12. Kay, J., The um toolkit for cooperative user modelling. User Modeling and User-
Adapted Interaction 4 (1995) 149–196
13. Uther, J., On the visualisation of large user models in web based systems. Phd
Thesis, Department of Computer Science, University of Sydney (2001)
The Munich Reference Model
for Adaptive Hypermedia Applications
Nora Koch and Martin Wirsing
Ludwig-Maximilians University of Munich

www.pst.informatik.uni-muenchen, Germany
{kochn,wirsing}@informatik.uni-muenchen.de
Abstract. Although adaptive applications are increasing in popularity, there are

only a few approaches that focus on their generalization or the specification of a
reference model. Trying to fill this gap, this paper presents a reference model
for adaptive hypermedia applications, similar to AHAM. The main novelty of
our approach is an object-oriented specification written in UML (Unified
Modeling Language) which integrates both an intuitive visual representation
and a formal unambiguous specification in OCL (Object Constraint Language).
Our reference model is defined as an extension of the Dexter Hypertext
Reference Model including user modeling aspects and rule-based adaptation
mechanisms.
Keywords. Adaptive Hypermedia, Reference Model, Visual Modeling, UML,
Formal Specification, Constraint Language, OCL.
1 Introduction
An adaptive hypermedia system is a set of nodes and links that allows one to navigate
through the hypermedia structure and that dynamically “adapts” (personalizes)
various visible aspects of the system to individual user’s needs, preferences or
knowledge [2]. These applications include an explicit representation of properties of
the user. This paper presents a reference model for these adaptive hypermedia
applications. The contribution of the paper is twofold. Firstly, we provide an object-
oriented formalization for such a reference model. Secondly, we include a graphical
representation of this model.
The objective of a reference model is to find common abstractions to the current
systems and to provide a basis for the development of these applications. It is named
Munich Reference Model, continuing with the tradition of choosing names of places
for the reference models related to the hypermedia field, such as the Dexter Model
[4], the Amsterdam Model [5] or the Dortmund Family of Hypermedia Models [9].
Adaptive hypermedia systems are first of all hypermedia systems, therefore our
reference model is based on the widely used Dexter Model for hypertext systems. It
includes the same three layers, but enhanced with adaptation functionality. The key
aspects of the Munich Reference Model are inclusion of a user model and an
adaptation model as part of the Storage Layer, the dynamic acquisition of user
behavior, a dynamic rule-based adaptation and a user behavior triggered Run-Time
session. To our knowledge there is only one other reference model for adaptive
applications: AHAM [3,12], which is semi-formally defined with tuples. The Munich

214 N. Koch and M. Wirsing
Model takes an object-oriented software engineering point of view whereas AHAM

takes more a database point of view. Our architecture is similar to the architecture of
the AHAM reference model, where user models are always structured as tables. The
AHAM adaptive engine is included in the adaptation model of our reference model as
data and functionality are integrated in the object-oriented approach. An important
contribution of AHAM is the adaptation rule language.
Our focus is, as already mentioned, the formal and visual description of the
reference model. The Dexter Model was formalized by Halasz and Schwartz [4] in the
specification language Z, early in the nineties. Since then, the use of object-oriented
methodologies gained in dissemination and importance. In addition, more emphasis is
now put on visual modeling languages making models more intuitive. These were our
motivations to select the Unified Modeling Language (UML) – standard for object-
oriented modeling – for the formalization of the Munich Reference Model. On the
one side, the UML [10] provides the notation and techniques (diagrams) for the visual
representation. It has the advantage of showing the relevant concepts at a glance, how
they are organized and how they are related to each other. This augments the intuitive
comprehension. On the other side, the Object Constraint Language (OCL) which is
part of the UML, is used to supplement the semi-formal graphical representation with
formally written semantics information.
The Munich Reference Model constitutes the basis for the UML-based Web
Engineering (UWE) approach that focus on development of adaptive hypermedia
applications [6,7]. UWE includes a design method and the description of a
development process that covers the whole life-cycle of these applications. This
reference model was used in the development of SmexWeb, a framework for
implementing adaptive learning systems on the Web [1]. SmexWeb supports all type
of dynamic adaptations, i.e. content, link and presentation adaptation.
This paper is an outline of the specification of the Munich Reference Model; the
complete version is included in [6]. It is organized as follows: Section 2 gives an
overview of the Munich Reference Model. Section 3 presents the specification of the
domain. Section 4 introduces the extensions to include user modeling and adaptivity.
Section 5 briefly presents the specification of a hypermedia session management.
Finally, in the last section some conclusions are outlined.
2 An Overview of the Reference Model
The Munich Reference Model preserves the three-layer structure of the Dexter Model
describing the network of nodes and links and the navigation mechanism. It extends
the functionality of each layer to include the user modeling and adaptation aspects.
The Run-Time Layer, the Storage Layer and the Within-Component Layer are
represented as UML subsystems as it is illustrated in Fig. 1.
• The Run-Time Layer contains the description of the presentation of the nodes
and links. It is responsible for user interaction, acquisition of user behavior and
management of the sessions.
• The Storage Layer has more functionality than just storing information about
the hypermedia structure. To support adaptation the Storage Layer is divided
into three sub-models:
The Munich Reference Model for Adaptive Hypermedia Applications 215
− The Domain Meta-Model that manages the basic network structure of the
hypermedia system in terms of mechanisms by which the links and nodes
are related and navigated. The nodes are treated as general data containers.
− The User Meta-Model manages a set of users represented by their user
attributes with the objective to personalize the application.
− The Adaptation Meta-Model consists of a set of rules that implement the
adaptive functionality, i.e. personalization of the application.
• The content and structure within the
hypermedia nodes are part of the Within-
Component Layer, which is not further Run-time Layer
detailed as its structure and content depend

on the application.
Presentation
The functionality of adaptive hypermedia systems Specification
is specified by three types of operations included Interface
in the classes of the reference model: Storage Layer
• Authoring operations are needed by

adaptive hypermedia systems to update Domain
Meta-Model
User
Meta-Model
components, rules and user attributes, e.g.
to create a link or a composite component,
to create a rule, to add an user attribute to Adaptation
the model, to delete components or rules. Meta-Model
• Retrieval operations are required to access

the hypermedia domain structure and the
User Model, e.g. to get a component, to get Anchoring
Interface
all rules triggered by a user’s behavior or
another rule.
• Adaptation operations are used to dyna- Within-Component
mically adapt the User Model content to the Layer
user behavior and to adapt the presentation

to the current state of the User Model, e.g. Fig. 1. Architecture of
the adaptive resolver, the constructor or the Adaptive Hypermedia
rule executor. Applications
The remainder of this paper presents the visual
specification (slightly simplified) of the layers of the reference model and includes a
few constraints of the formal specification out of a total of seventy constraints that
comprise the complete specification of the Munich Reference Model [6].
3 Specification of the Domain Model
The Domain Meta-Model describes the structure of a hypermedia as a finite set of

components together with three main operations, a resolver, an accessor and a
constructor. These concepts are modeled by a class Domain and a class Component.
Every component has a globally unique identity (class UID). With the operations
resolver, accessor and constructor it is possible to “retrieve” and “construct” adaptive
components. The accessor operation allows one to “access” a component given its
UID. UIDs “resolve” to a component. This way UIDs provide a guaranteed

mechanism for addressing any component in the hypermedia domain. As in the
Dexter Model, this addressing is accomplished in a indirect way based on the entities
called anchor (class Anchor) consisting of two parts: an anchor ID (class AnchorID)
and an anchor value (class AnchorValue). The anchor value is an arbitrary value that
specifies some location within a component. The anchor ID is an identifier that
uniquely identifies the anchor within the scope of the component.
Domain
resolver(cs): Set (UID)

accessor(uid): Component
constructor (atoms):Page
linksTo(uid): Set (UID)
linksToAnchor(uid,aid):
... Set (UID)
components 1..*
component * Attribute
Component 1
value:Value
1 attributes
1
consistency(c,c): Boolean 1 * 1 1
1..* Within-
anchors Anchor 1 anchor component
children {ordered} Value Layer ::
AnchorValue
Prerequisite
1 anchorID
Concept OnPage
Concept AnchorID
Relationship
...
anchorSpec 1
1 Link
0..1
/fragments
Composite Atom {ordered}
1..* pres
1 1 specifiers 2..* Spec
1
Specifier 1 presSpec
1 PresentSpec
content 1 direction = enum specifier 1 1
Page Within- (TO,FROM,
component NONE,BIDIRECT) 1 1 ComponentSpec 1..* 1..* UID
Layer :: uids
1..*
Content compSpec resolvesTo 1 uid
constructs
accessTo
Fig. 2. View of the Domain Meta-Model of the Munich Reference Model
The visual specification of the Domain Meta-Model of the Munich Reference

Model is represented by an UML class diagram. Part of it is shown in Fig. 2. Note
that for classes belonging to another package, the name of the package is included
before the class name, e.g. Within-Component-Layer::AnchorValue. As examples, the
classes Component and Domain are described with some more detail below.
Component. A component is an abstract representation of an information item

from the application domain. It is represented by an abstract class Component. A
component can either be a concept (class Concept) or a concept relationship (class
ConceptRelationship). A concept, in turn, can either be an atom (class Atom) or a
composite (class Composite). A concept relationship can be a link (class Link) or a

prerequisite (class Prerequisite), or a is-part-of relation (class On page), etc. This
inheritance hierarchy is shown in the UML class diagram (Fig. 2). The component
information consists of attributes (class Attribute), a presentation specification (class
Present Spec) and a sequence of anchors (class Anchor). The UML visual
specification is insufficient to model the “type consistency” between components.
Therefore, the following OCL constraint is added to the specification to express that
two components are “type consistent” if both are of the same type.
context Component :: consistency (c1:Component, c2:Component):Boolean

post: result = c1.oclIsTypeOf(Composite) and c2.oclIsTypeOf(Composite)
or c1.oclIsTypeOf(Link) and c2.oclIsTypeOf(Link)
or c1.oclIsTypeOf(Atom) and c2.oclIsTypeOf(Atom)
or .....
Domain. The domain is represented by a class Domain, which is a composition of

objects of type Component. The class Domain includes two operations for links and
anchors ensuring the navigation functionality of the hypermedia system. These are the
linksTo and the linksToAnchor operations. The linksTo operation returns the set of
links that resolve to a specific component. The linksToAnchor obtains the set of links
that resolve to a specific anchor. The following is the OCL specification of the pre-
condition and post-condition of linksTo. The post-condition expresses that result
consists of the set of all Link identifiers such that one of the component specifications
of the corresponding Link resolves to the given UID.
context Domain :: linksTo ( uid : UID ) : Set (UID)

pre: components → exists ( c : Component | accessor (uid) = c )
post: result = UID.allInstances → select ( lid : UID |
Component.allInstances → exists (link :Component |
link.oclIsTypeOf (Link) and link = accessor (lid)
and ComponentSpec.allInstances → exists ( cs :
ComponentSpecs | link.specifiers.compSpecs → includes (cs)
and uid = resolver (cs) ) ) )
4 Modeling Adaptive Hypermedia Applications
The Munich Reference Model includes adaptation and user modeling functionality.
The User Meta-Model defines the user attributes and attribute-values that are relevant
to the adaptive application. The adaptive mechanisms are specified in the Adaptation
Meta-Model and they are responsible for adaptive content, adaptive links and
adaptive presentation. The presentation specification builds pages out of page
fragments, taking into account the adaptive mechanisms.
4.1 The User Meta-Model
The User Meta-Model describes the structure of the individual models of each user
and how these models are administrated. User modeling comprises initialization,
updating and retrieval of the current state of a User Model. The User Meta-Model is
modeled as a subsystem that consists of a class UserManager and a set of Users and
operations initializer, updater and an evaluator. Fig. 3 depicts the classes of the User
Meta-Model subsystem and its relationship to the Domain Model.
A user of an adaptive hypermedia application is modeled by a class User, which is
related through an aggregation association to a UserIdentification and to a set of User
Attributes. The user ID identifies the user uniquely in the universe of the application.
With the user attributes the system provides a representation of the user’s
characteristics that are relevant for the application. One can distinguish different types
of information contained in user models: user’s knowledge, user’s preferences, user’s
background experience, user’s tasks, etc., summarized in two categories: “user
knowledge related to the domain components” and “user general characteristics”.
The first group includes domain dependent attributes while those of the second
group are domain independent. The second group includes knowledge not related to
the components, such as background knowledge and preferences. Classification like
this can be found in Hynecos [11] and SmexWeb [1]. We model these two groups of
user’s characteristics with Class DependentAttr and Class IndependentAttr. The
separation has the advantage that the domain independent attributes can be shared
with other applications. The following constraint defines the invariant for a domain
independent User Model, i.e. all user attributes are independent of the domain.
User Meta-Model
UserManager
initializer(ui,name,attr): User
evaluator(ui,attr):
UserAttributeValue
updater(ui,attr, val)
...
Domain Meta-Model
1.. users
*
User
Domain
user
username: String userIDs
UserID resolver(cs)
email: String userID domain
* accessor(uid)
...
1..
* userAttrs
UserAttribute
attname: String
components 1..
*
dependAttrs
IndependentAttr DependentAttr *
comps
Component
*
{xor}
attrVal UserAttrValue
value: Value
Fig. 3. View of the User Meta-Model of the Munich Reference Model

context User
inv domain independent user model:
userAttrs → forAll ( uat:: UserAttribute |
uat.oclIsTypeOf (IndependentAttr) )
Let us mention here only the formalization of one functionality related to the User
Meta-Model subsystem: the registration of a new user. We define an initializer
operation that creates a new instance of class User for each new user that registers to
the adaptive hypermedia application and assigns a given set of attributes to this user.
context UserManager :: initializer (userIdentification:UserID, n:String,

defaultAttrs: Set(UserAttribute)) : User
post: result.oclIsNew and users = users@pre → including (result)
and result.userID = userIdentification and result.username = n
and result.userAttrs = defaultAttrs
4.2 The Adaptation Meta-model
The adaptation is performed using a set of rules, such as in most adaptive hypermedia
applications; typical examples of rule-based adaptation is supported by the
frameworks AHA [3] and SmexWeb [1]. These rules determine how pages are built
and how they are presented to the user. The Adaptation Meta-Model is specified by a
UML class diagram, which is depicted in Fig. 4.
The core elements used to model the adaptation are the class Adaptation and the
class Rule. The class Adaptation includes three main operations: an adaptation
resolver, a finder and a trigger. The first one “resolves” a component specification
into a UID of an appropriate component that builds an adapted page. The second one
implements a trigger mechanism that returns all the rules triggered by one given rule,
i.e. the rules to be used at a given time. The first rule to be used is triggered by the
user behavior. The executor operation of the class Rule allows the system to select the
appropriate components, and to perform content-adaptation, presentation-adaptation
and link-adaptation as well as to update the User Model. These operations play the
role of the adaptive engine in AHAM.
The specification of the Adaptation Meta-Model is supplemented with a set of
OCL constraints. For example, the following OCL invariant assures the dynamic
update of the User Model: For at least one user attribute there exists a rule that
modifies an attribute value of the User Model.
context Adaptation
inv dynamic update of the user model:
Rules.allInstances → exists ( r:Rule | r.oclIsTypeOf (AcquisitionRule)
and r.action.elements → exists (m: ModelElement |
m.values.oclIsTypeOf (UserAttributeValue)
and m.modified ) )
An object of class Rule consists of one condition (class Condition), one action
(class Action) and attributes, such as phase and propagate proposed by De Bra et. al
[3]. phase determines whether rules are applied before or after the User Model is
updated while propagate with a value true allows the system to trigger other rules.
Conditions and actions are expressions containing model elements and operators.
ModelElements are defined by two attributes: an element identifier (elementID)

and a Boolean value (modified) which indicates whether the model element is being
modified in the actual action. Only certain types of model elements, i.e. User Model
attribute values and presentation specifications can have a modified value true.
Our formalization of rules is very general, thus our reference model does not
prevent problems with confluence and termination of rule-based systems. Depending
on the chosen rule language, rule applications may be non-terminating and non-
confluent. These problems can be analyzed in each case using different approaches,
such as those used in rewriting systems or in the active database field.
Adaptation
finder (uid,ub): Rule

trigger (rule): Set(Rule)
adaptationResolver
(cs,ui,ub): UID
...
Construction
trigger Rule Content
rules Adapter
1 ..*
Rule
1..* propagate: Boolean
Adaptation Link
phase: enum
{PRE,POST}
Rule Adapter
1 ..*
executor (uid,ui,ub)
...
triggers
Presentation
Acquisition Adapter
1..* behaviours condition
Rule
action
User ..*
0
Condition Action
Behaviour
elements elements
1 ..* 1..*
ModelElement values User Meta-Model::
elementID: String UserAttributeValue
modified: Boolean
Fig. 4. View of the Adaptation Model of the Munich Reference Model
Rules are classified according to their objectives into: construction rules,

acquisition rules and adaptation rules [6]. Adaptation rules adapt content, links or the
presentation of the application. They differ in the executor method. The different
types of rules are represented as a hierarchy of rules as it is shown in the class
diagram of the Adaptation Model (see Fig. 4).
5 Session Management
The Run-Time Layer manages different sessions for the users generating and
presenting the instances of pages. The Run-Time Layer describes how the
components are presented to the user. This presentation is based on the concept of
instantiation of a component, i.e. a copy of the component is cached to the user. The
Session
User Meta-Model 1..* Adaptation
*
:: UserID userIDs Meta-Model:: adaptation sessions instants (iid):Instantiation
Adaptation openComponent (specs,pspecs):Set(Inst)
realizer (inst): Component
1.. rules runTimeResolver(cs,ui,uop): UID
* ...
Adaptation
Meta-Model::
Rule
1..
*
Domain *
Adaptation behaviour Operation history
Meta-Model:: Meta-Model:: * opn = enum ( History
Domain ÓPEN, CLOSE,
UserBehaviour operations
*
EDIT, SAVE, ...) operations
{ordered}
components 1..
* * iids
Domain Domain
Meta-Model::
accessTo
Meta-Model :: * IID
instUid iids
Component UID
inst
LinkMarker *links{ordered} Instantiation
linkAnchor (lm): AnchorID
Fig. 5. View of the Run-Time Layer of the Munich Reference Model
copy receives an instantiation identifier (class IID). It should be noted that more than
one instantiation for a component may exist simultaneously and that a user may be
viewing more than one component.
Instantiation of a component also results in instantiation of its anchors. An
instantiated anchor is known as a link marker. These concepts are modeled with the
classes Instantiation, IID, and LinkMarker. In order to keep track of all these
instantiations the Run-Time Layer uses an entity session (class Session) as shown in
Fig. 5. A session can be open or closed and in a session the user can perform
operations, such as open a component that results in the creation of an instantiation,
edit an instantiation and follow a link. All these operations that result from the user
interactions are recorded in a history which constitutes the basis of the observation of
the user behavior and the adaptation mechanism. As example a constraint for the
instantiator operation is shown. Given an UID of a component, the function returns
an instantiation of the component that is part of the session. The presentation
specification is a primitive in the model, which contains information about how the
component is to be presented by the system during instantiation (see Figures 5 and 2).
context Session :: instantiator (uid: UID, ps: PresentSpec): Instantiation
pre: adaptation.domain.components → includes (accessor(uid) )
post: result = iids.inst → select (ins:Instantiation |
ins.presSpec = ps and ins.iid.instUID = uid ) → asSequence → first
Figure 5 depicts part of the Run-Time Layer for adaptive hypermedia systems. The
UML class diagram shows how core classes of the Domain Meta-Model, the User
Meta-Model and the Adaptation Meta-Model collaborate with classes of the Run-
Time Layer.
In this paper we introduced an object-oriented reference model for adaptive

hypermedia applications – called Munich Reference Model. It constitutes the basis of
our UWE engineering approach for adaptive Web applications [6,8]. This model was
defined in parallel to the development of SmexWeb. SmexWeb is a framework for
adaptive Web learning applications [1]. The architecture of this model is similar to the
AHAM architecture, i.e. it extends the Dexter model by including an user model and
an adaptation model in the Storage Layer as well as adapting the domain model and
the Run-Time Layer.
We presented an integrated visual and formal specification of the reference model.
The model is visually represented using UML notation and is formally specified in
OCL. UML was chosen as it is the de facto standard modeling language. A visual
representation is missing when other kind of specification, e.g. Z or VDM are selected
[4]. OCL is part of the UML and is used for the specification of invariants for the
model elements and for the specification of pre-conditions and post-conditions on
operations describing the adaptive functionality.
Acknowledgment. We thank the reviewers for their valuable feedback and their
requests for additional explanations.
References
1. Albrecht F., Koch N. and Tiller T. (2000). SmexWeb: An Adaptive Web-based Hyperme-
dia Teaching System. Journal of Interactive Learning Research. Kommers P. &
Mizoguchi R. (Eds.).
2. Brusilovsky P. (1996). Adaptive Hypermedia: An attempt to analyze and generalize.
Proceedings of First International Conference on Multimedia, Hypermedia and Virtual
Reality 1994. Brusilovsky P. & Streitz N. (Eds.) LNCS 1077, Springer, pp. 288-304.
3. De Bra P., Houben G.-J., and Wu H. (1999). AHAM: A Dexter-based Reference Model of
Adaptive Hypermedia. Proceeding of the ACM Hypertext Conference, pp. 147-156.
4. Halasz F. and Schwartz M. (1990). The Dexter Hypertext Reference Model. NIST
Hypertext Standardization Workshop.
5. Hardman L., Bulterman C. and Rossum G. (1994). The Amsterdam Reference Model.
Communications of the ACM 37(2).
6. Koch N. (2000). Software Engineering for Adaptive Hypermedia Systems: Reference
Model, Modeling Techniques and Development Process. PhD. Thesis, Uni-Druck.
7. Koch N. (2002). An Object-Oriented Hypermedia Reference Model. In Information
Modeling for Internet Applications, van Bommel P. (Ed.), to appear.
8. Koch N. and Wirsing M. (2001). Software Engineering for Adaptive Hypermedia
Applications? Third Workshop on Adaptive Hypertext and Hypermedia at the UM´2001.
9. Tochtermann K. and Dittrich G. (1996). The Dortmund Family of Hypermedia Systems.
Journal of Universal Computer Science.
10. UML: The Unified Modeling Language. Version 1.3. (1999). http://www.omg.org/uml
11. Vassileva J. (1994). A Practical Architecture for User Modeling in a Hypermedia-based
th
Information System. Proceeding of the 4 International Conference on User Modeling.
12. Wu H., De Bra P., Aerts A. and Houben G.-J. (2000): Adaptation Control in Adaptive
Hypermedia Systems. Proceedings of the Adaptive Hypermedia and Adaptive Web-based
Systems. Brusilovsky P, Stock O., Strapparava C. (Eds.). LNCS 1892, Springer, pp. 250-
259.
Tracking Changing User Interests
through Prior-Learning of Context
Ivan Koychev
FhG - FIT.ICON
D-53754 Sankt Augustin, Germany
phone: +49 2241 14 2194, fax: +49 2241 14 2146
[email protected]
Abstract. The paper presents an algorithm for learning drifting and recurring
user interests. The algorithm uses a prior-learning level to find out the current
context. After that, searches into past observations for episodes that are relevant
to the current context, ‘remembers’ them and ‘forgets’ the irrelevant ones. Fi-
nally, the algorithm learns only from the selected relevant examples. The ex-
periments conducted with a data set about calendar scheduling recommenda-
tions show that the presented algorithm improves significantly the predictive
accuracy.
1 Introduction
Recently, many systems have been developed that recommend information, products
and other items. These systems try to help users in finding pieces of information or
other objects in which the users could be interested [8]. In a similar way, adaptive
hypermedia systems build a model of the goals and preferences of each user and use
this model to adapt the interaction to the needs of the user [3]. Many of those systems
use machine learning methods for learning from observations about the user [15].
However, user interests and preferences can change over time. Some of the systems
are provided with mechanisms that are able to track drifting user interests [1, 5, 9, 11,
among others]. The problem of learning drifting user interests is relevant to the prob-
lem known as concept drift in the area of machine learning. The next section dis-
cusses different approaches about learning drifting concept and their applications for
learning about users.
In this paper it is assumed that the user interests do not only change, but also pos-
sibly recur. The user interests can be quite wide and the user can currently focus her
attention on a small subset of her broad interests. For example, the whole set of user
interests in the case of Internet browsing can include interests that are relevant to her
job, as well as her hobbies, etc. Even the user's job related interests could be quite
extensive and interdisciplinary. A system that assists the user in web browsing should
be flexible enough to recognize what her current interests are and provide her with
relevant recommendations. A possible approach is to learn about current user interests

224 I. Koychev
from a time window that includes recent relevant observations only. However, if the
current user interests often change, a precise user profile cannot be learned from a
small set of relevant recent observations only. Hence, the system can search for past
episodes where the user has demonstrated a similar set of interests and try to learn a
more precise description of the current user interests, ‘remembering’ relevant and
‘forgetting’ irrelevant observations.
This paper presents such an algorithm for tracking changing user interests and
preferences in the presence of changing and recurring context. First, the algorithm
learns about current context. Subsequently, it selects past episodes that are relevant to
this context and eventually it learns concept descriptions from the selected examples.
The next section discuses different approaches for tracking changes developed in
areas of machine learning and user modeling. Section 3 presents a two-level learning
algorithm that is applicable to learning changing and recurring user interests and
preferences. Section 4 presents experiments of the designed algorithm with real data
about calendar scheduling preferences as well as with an artificial data set.
2 Related Works
This section briefly introduces different approaches developed for tracking changing
(also known as shifting, drifting or evolving) concepts. Such systems use different
forgetting mechanisms to cope with this problem. Usually it is assumed that if the
concept changes, then the old examples become irrelevant to the current period. The
concept descriptions are learned from a set of recent examples called time window.
For example, a software assistant for scheduling meetings is described in Mitchell et
al. [11]. It employs induction on a decision tree to acquire assumptions about individ-
ual habits of arranging meetings. The learning method uses a time window to adapt
faster to the changing preferences of the user. A system that learns user's interest
profiles by monitoring web and e-mail habits is described in Grabtree and Soltysiak
[15]. This research shows that user's interests can be tracked over time by measuring
the similarity of interests within a time period.
An improvement of the time window approach is the use of heuristics to adjust the
size of the window. Widmer and Kubat [17] use a time window with a flexible size,
which is adapted dynamically. The window size and thus the rate of forgetting is
supervised and dynamically adjusted by heuristics that monitor the learning process.
Klingenberg and Renz [17] investigate the application of such an approach in the area
of information retrieval.
Maloof and Michalski [10] have developed a method for selecting training exam-
ples for a partial memory learning system. The forgetting mechanism of the method
selects extreme examples that lie at the boundaries of concept descriptions and re-
moves from the partial memory examples that are irrelevant or outdated for the learn-
ing task. The method uses a time-based function to provide each instance with an age.
Examples that are older than a certain age are removed from the partial memory.
Nevertheless, pure time window approaches totally forget the observations that are
outside the given window, or older than a certain age. The examples which remain in
Tracking Changing User Interests through Prior-Learning of Context 225
the partial memory are equally important for the learning algorithms. This is abrupt
and total forgetting of old information which in some cases can be valuable.
System use different approaches to avoid loss of useful knowledge learned from
old examples. The CAP system [11] keeps old rules till they are competitive with the
new ones. The architecture of FLORA systems [17] assumes that the learner main-
tains a store of concept descriptions relevant to previous contexts. When the learner
suspects a context change, it will examine the potential of previous stored descrip-
tions to provide better classification.
An intelligent agent called NewsDude that is able to adapt to changing user inter-
ests is presented in Billsus, and Pazzani [1]. It learns two separate user models: one
represents the user's short-term interests and the other represents the user's long-term
interests. The short-term model is learned from the most recent observations only. It
represents user models that can adjust more rapidly to the user's changing interests. If
the short-term model cannot classify the story at all, it is passed on to the long-term
model. The purpose of the long-term user model is to model the user's general prefer-
ences for news stories and compute predictions for stories that could not be classified
by the short-term model. This hybrid user model is flexible enough to consider
changes in user interests and keeps track of long-term user interests as well. Chiu and
Webb [4] have used a similar approach - a dual student model for handling concept
drift.
Webb and Kuzmycz [14] suggest a data aging mechanism that places an initial
weight of 1 on each observation. In a similar way Koychev and Schwab [9] have used
a gradual forgetting function that provides each observation with a weight according
to its appearance over time.
An approach for tracking changing concepts that employs two-level learning algo-
rithms is presented in [16]. The assumption is that the domain provides explicit clues
as to the current context (e.g. attributes with characteristic values). A two-level learn-
ing algorithm is presented that effectively adjusts to changing contexts by trying to
detect (via meta-learning) contextual clues and using this information to focus the
learning process. Another two-level learning algorithm assumes that concepts are
likely to be stable for some period of time [6]. This approach uses batch learning and
contextual clustering to detect stable concepts and to extract hidden context.
The approach presented in this paper also employs a two-learning level. However,
it does not assume that the attributes represent current context explicitly. It starts from
the assumption that the recent observations are able to provide information about
current context. The recent relevant observations cannot be sufficient to learn an
accurate description of the concept, but the learned description is accurate enough to
be able to distinguish the past episodes that are relevant to the current context. Then
the algorithm constructs a new training set, ‘remembers’ relevant and ‘forgets’ irrele-
vant examples. Finally, the concept description is learned from this set of examples.
226 I. Koychev
3 Tracking Changes through Prior-Learning of Context
When the concept drifts and possibly recurs, we can use time window based forget-
ting mechanisms. However, the recent examples that represent the current context can
be insufficient for learning accurate descriptions. Therefore, if the context recurs, then
remembering the ‘old’ examples that are relevant to the current context should
enlarge the size of the training set and thus improve the predictive accuracy. How-
ever, the context is frequently hidden and explicit indicators about its changes and
recurrences cannot be discovered easily. Hence, in such cases the aim should be to
learn more about the current context and then to search for old observations that were
made in a similar context. An algorithm that makes use of this idea consists of the
following three steps:
1. Learning about current context. A relatively small time window is used
to learn a description of the current context (e.g. learning a description
of the user interests based on the recent observations about the user).
2. Remembering relevant past episodes. The learned description in step 1.
is tested against the rest of the training set. The episodes that show a
predictive accuracy that is greater than a predefined threshold are se-
lected (i.e. selecting the episodes that are relevant to the current con-
text).
3. Learning from context-related examples. The new data set selected in
step 2. is used for learning a new description of the current user inter-
ests, which is expected to be more accurate.
Let’s call this algorithm COPL (COntext Prior Learning algorithm). The COPL al-
gorithm requires a predefinition of the following settings:
• The size of the time window used in step 1. This time window should be long
enough to allow a sufficiently accurate description of the current context to be
learned, as well as short enough to be able to track fast changing user interests.
Some enhancements like adaptive time window [17] can be employed aiming at
improving predictive accuracy.
• The episode selection criterion for step 2. This criterion should be able to distin-
guish the episodes that are relevant to the learned context in step 1. The criterion
should be resistant to noise in the sequence of examples.
• The threshold for the episode-selecting criterion in step 2. After the episode selec-
tion criterion has been established, a suitable threshold should be defined, which
should assure as much as possible that only the relevant old examples be selected.
• The learning algorithms used in steps 1. and 3. The same or different learning
algorithms can be used in those steps.
Those settings should be defined empirically and based on preliminary investiga-
tion of the application domain. The implementation of the algorithm described in the
next section gives an example of such definitions.
The next section presents the results from experiments that compare the designed
algorithm where the main idea is to extend the set of examples by recovering relevant
past examples as opposite to the CAP and FLORA approaches where the model was
extended by past rules.
4 Experiments
This section present results from experiments with the COPL algorithm. Two data
sets are used in the experiments. The first one contains data from a real use of a cal-
endar manager tool aiming at helping the user to scheduling meetings [Mitchell et al.
[11]. The second one is an artificial data set [13] that is used in many papers in the
area of Machine Learning dedicated to concept drift (e.g. [10, 17], etc.)
Mitchell et al. [11] have developed a software assistant that helps schedule a par-
ticular user's calendar: a calendar manager called CAP (Calendar APprentice). CAP
learns the users' scheduling preferences through routine use, enabling it to give cus-
tomized scheduling advice to each user. It can be considered as an analogy to a hu-
man secretary who might assist someone in managing a calendar. CAP employs in-
duction on decision tree to acquire assumptions about individual habits of arranging
meetings. The learning method uses a time window to adapt faster to the changing
preferences of the user. The newly generated rules are merged with old ones. The
rules that perform poorly on the test set drop out of the list.
The user's scheduling preferences depend very much on a hidden context. Some of
this context can be assumed and explicitly presented and used for improving predic-
tive accuracy (e.g. academic semesters, etc.). However, there are many other events
and conditions that can influence the meeting schedule and which cannot be explicitly
represented by an attribute space (e.g. room availability, the schedule preferences of
other participants of a meeting and many others). Under this condition, the predictive
accuracy of the system can oscillate with very high amplitude. A more comprehen-
sive investigation and analysis of the specifics of the domain can be found in Mitchell
et al. [11].
The section below presents the results from experiments conducted with the CAP
data set1. The attributes used for describing the calendar events in the current experi-
ments are listed in Table 1. The task is to predict the following meeting characteris-
tics:
• Duration - the duration of the meeting in minutes e.g. 30, 60, 90, etc. (number of
values legal - 13);
• Day-of-week - the day of the week of this meeting; e.g. Monday, Thursday, etc.
(number of legal values - 6);
• Location – the place where the meeting is held; e.g. weh5409 (number of legal
values - 142);
1 http://www.cs.cmu.edu/afs/cs.cmu.edu/project/theo-5/www/cap-data.html
228 I. Koychev
• Start-time - the time at which the meeting begins, in military time; e.g. 930
(9:30am), 1400 (2pm), etc.) (number of legal values - 21);
Table 1. The list of features that are used for describing calendar events.
Third-most-common-time-last-60-days-this-meeting-type
Third-most-common-time-last-60-days
Second-most-common-time-last-60-days-this-meeting-type
Second-most-common-time-last-60-days
Most-common-time-these-attendees-last-60-days
Most-common-time-these-attendees
Most-common-time-last-60-days-this-meeting-type
Most-common-time-last-60-days
Most-common-day-these-attendees-last-60-days
Most-common-day-these-attendees
Duration-of-next-meeting-with-these-attendees
Duration-of-last-meeting-with-these-attendees
Day-of-week-of-next-meeting-with-these-attendees
Day-of-week-of-last-meeting-with-these-attendees
Required-seminar-type
Required-course-name
Required-speakers
Single-person?
Action
CMU-attendees?
Group-attendees?
Position-attendees
Department-attendees
Sponsor-attendees
Known-attendees?
Duration
Day-of-week
Location
Start-time
The settings of the algorithm listed in the previous section are defined for the con-
ducted experiments as follows:
• The size of the time window: Preliminary experiments show that for different
prediction tasks the size of the window that produces best predictive accuracy can
be quite different. For the given data set the best accuracy is reached for the win-
dow of the following size: Location - 200; Duration - 350; Start-time - 350; Day-
of-week - 400.
• The episode selection criterion for step 2. The criterion used in this implementation
selects the examples e j for the new data set taking into account the average pre-
dictive accuracy in its neighborhood. In particular, a small episode around the ex-
ample which includes the previous two and next two examples, is used. An event
will be selected for the new training set e j ∈ S new if the average predictive accu-
racy for this episode is greater than or equal to a predefined threshold τ .
• The threshold for the episode-selecting criterion in step 2. is set up to τ = 0.6 in
all experiments.
• The learning algorithm used in steps 1. and 3. is Induction on Decision Tree (aka
ID3) [12]. This algorithm was used in CAP, which makes the comparison between
different approaches clearer. This algorithm produces an explicit user profile (e.g.
set of rules) that is understandable for the user. This is an important advantage
from the viewpoint of user modeling.
Table 2. Comparison of predictive accuracy for the User.
Prediction task CAP ID3-FM COPL (ID3)

Location 64% 58% 67%
Duration 63% 71% 79%
Start-time 34% 39% 48%
Day-of-week 50% 52% 66%
Average 53% 55% 65%
Table 2 presents the results from experiments with data for User 1. In this experi-
ment a new description of user preferences is learned after each 10 meetings. The
learned description at each step is tested on the next 10 meetings. The line in the table
presents the accuracy of prediction for different learning tasks. The results are com-
pared with the CAP. The average predictive accuracy of the ID3 with full memory
(ID3-FM) to some extent outperforms the CAP. This is slightly surprising, because
CAP is designed to track changing user preferences better than a simple learning
algorithm. An explanation of this phenomenon is that some implementation details
like attribute selection criteria and used pruning method can change the outcome of
the algorithm. The use of one level time window, even with an adaptive size, does not
improve the predictive accuracy because the user preferences alternate very often and
with high amplitude. The comparison between full-memory learning algorithm (ID3-
FM) and the presented two-level learning algorithm is fully compatible because the
same implementation of the basic learning algorithm is used. The results from the
experiments show that the context-learning algorithm is able to improve the average
predictive accuracy for each feature. All those improvements are significant (using t-
test with α = 0.01 ).
Figure 1 shows the results from experiments for the predicted features. It can be
seen that the user's preferences can change abruptly, which leads to a dramatic de-
crease of the predictive accuracy. The presented two-level algorithm tracks changes
better than the basic algorithm and produces a significantly improved average accu-
racy.
Experiments with this data set, which use the Winnow and Weighted-Majority al-
gorithms, were reported in Blum [2]. The Winnow with a large feature set reaches the
best average accuracy, which is equal to that reached by the algorithm in the pre-
sented experiments. However, these algorithms are not suitable for producing explicit
user profiles, which is considered to be important in the area of user modeling.
230 I. Koychev
Day-of-Week Duration
100% 100%
90% 90%
80% 80%
70% 70%
Predictive Accuracy
Predictive Accuracy
60% 60%
50% 50%
40% 40%
30% 30%
20% 20%
10% 10%
0% 0%
200 400 600 800 1000 1200 1400 1600 200 400 600 800 1000 1200 1400 1600
Events Events
Context Learning ID3-FM Contex learning ID3-FM
Location Start Time
100% 100%
90% 90%
80% 80%
70% 70%
Predictive Accuracy
Predictive Accuracy
60% 60%
50% 50%
40% 40%
30% 30%
20% 20%
10% 10%
0% 0%
200 400 600 800 1000 1200 1400 1600 200 400 600 800 1000 1200 1400 1600
Events Events
Context learning ID3-FM ID3-FM Context Learning
Fig. 1. The improvement in predictive accuracy for the predicted features.
To compare the presented approach with FLORA3, which is able to recover ‘old’
rules learned in a similar context [17], experiments were conducted also with
STAGGER data set [13]. The instance space of a simple blocks world is described by
three attributes size = {small, medium, large}, color = {red, green, blue}, and shape =
{square, circular, triangular}. There is a sequence of three target concepts (1) size =
small and color = red, (2) color = green or shape = circular and (3) size = (medium
or large). 120 training instances are generated randomly and classified according to
the current concept. The underlying concept is forced to change after every 40 train-
ing examples: (1)-(2)-(3). A concept description is learned from initial n examples.
After each learning phase the predictive accuracy is tested on an independent test set
of 100 instances. The result are averaged over 10 runs. The concept recurrence is
simulated by generating this sequence three times: (1)-(2)-(3)-(1)-(2)-(3)-(1)-(2)-(3)
[17].
The parameters for the COPL algorithm in this experiment are set up as follows:
the size of the time window used at step 1 is 18; the episode selection criteria and the
related threshold remain the same as above; and the used learning algorithms at step 1
and 3 is Naïve Bayes Classifier (NBC) to demonstrate the ability of the presented
two-level algorithm to work with other learning algorithms.
Table 3 compares the presented algorithm with FLORA3 [17]. On the basic data
set (1-120) the FLORA3 produces a slightly better accuracy (i.e. non significant dif-
ference). On recurring concepts (i.e. examples 121-360) both algorithms perform
better than the ones that do not recover the context (e.g. FLORA2 [17]- 81.5%). The
COPL (NBC) algorithm benefits from the recurrence of context better than FLORA3
(see columns 121-240 and 241-360 of Table 3). Moreover, the predictive accuracy of
the presented algorithm increases when context recurs, which shows that it really
takes advantage of context recurrence. For example, on second recurrence of the

concept (see column 241-360 of Table 3) the COPL algorithm produces a signifi-
cantly better (using t-test with α = 0.01 ) average accuracy than FLORA3.
Table 3. Comparison between FLORA3 and COPL (NBC) on recurring context.

Algorithm\Examples: 2-120 121-240 241-360
FLORA 3 context 85.9% 85.4% 83.5%
COPL (NBC) 85.3% 85.6% 87.1%
5 Conclusion
The paper describes a two-level learning algorithm that is able to track changing user
interests and preferences through prior-learning of context. The algorithm benefits
from the recurrence of the context by remembering the relevant observations and
forgetting the irrelevant ones. The presented approach provides a general framework
for dealing with changing and recurring user interests that can be used with different
machine learning algorithms. Conducted experiments with recommendations about
calendar scheduling demonstrate that the approach is able to improve the predictive
accuracy significantly. Additional experiments conducted with an artificial data set
demonstrate that the presented algorithm really makes use of context recurrence and
increases the predictive accuracy when the context recurs. Further investigations of
the episode selection criterion and designing a mechanism for its threshold detection
are expected to improve the predictive accuracy of the algorithm additionally.
The presented two-level learning algorithm can be embedded in any type of adap-
tive hypermedia system where some observations during the interaction with the user
have been collected and then used to learn about the user. The knowledge learned
about the user can then be used to adapt the interaction to the needs of that user. Pro-
viding the user with adequate recommendations in the presence of fast changing
user’s interests and preferences is, for example, vital for many contemporary recom-
mendation systems. Future applications of the algorithm are expected to provide fruit-
ful ideas for the development of mechanism for dynamical adaptation of the algo-
rithm parameters.
References
1. Billsus, D., and Pazzani, M. J.: A Hybrid User Model for News Classification. In
Kay J. (ed.), UM99: Proceedings of the Seventh International Conference on User
Modeling, Lecture Notes in Computer Science, Springer-Verlag (1999) pp. 99-108.
2. Blum, A.: Empirical Support of Winnow and Weighted-Majority Algorithms: Re-
sults on a Calendar Scheduling Domain. Machine Learning 26 (1997): 5-23.
3. Brusikovsky, P. Adaptive Hypermedia. User Modeling and User-Adapted Interaction
11 (2001) 87-110.
232 I. Koychev
4. Chiu, B. and Webb, G.: Using Decision Trees for Agent Modeling: Improving Pre-
diction Performance. User Modeling and User-Adapted Interaction 8 (1/2) (1998)
131-152.
5. Grabtree, I. and Soltysiak, S.: Identifying and Tracking Changing Interests. Interna-
tional Journal of Digital Libraries vol. 2 (1998) 38-53.
6. Harries, M. and Sammut, C. Extracting Hidden Context. Machine Learning 32
(1998) 101-126.
7. Klingenberg, R. and Renz, I.: Adaptive information filtering: learning in the presence
of concept drift. AAAI/ICML-98 Workshop on Learning for Text Categorization, TR
WS-98-05, Madison, WI, (1998).
8. Kobsa, A., Koenemann, J. and Pohl, W.: Personalized Hypermedia Presentation
Techniques for Improving Online Customer Relationships. The Knowledge Engi-
neering Review, 16(2) (2001) 111-155.
9. Koychev, I. and Schwab, I.: Adaptation to Drifting User's Intersects - Proceedings
ECML2000/MLnet workshop: ML in the New Information Age, Barcelona, Spain,
(2000) pp. 39-45.
10. Maloof, M. and Michalski, R.: Selecting examples for partial memory learning. Ma-
chine Learning 41 (2000) 27-52.
11. Mitchell, T., Caruana, R., Freitag, D., McDermott, J. and Zabowski, D.: Experience
with a Learning Personal Assistant. Communications of the ACM 37(7) (1994) 81-
91.
12. Quinlan, R.: Induction of Decision Trees. Machine Learning 1 (1986) 81-106.
13. Schlimmer, J. and Granger, R.: Incremental Learning from Noisy Data. Machine
Learning 3, Kluwer Academic Publishers (1986), 317-357.
14. Webb, G. and Kuzmycz, M.: Feature-based modelling: a methodology for producing
coherent, consistent, dynamically changing models of agents' competencies. User
Modeling and User-Adapted Interaction 5(2) (1996) 117-150.
15. Webb, G. Pazzani, M. and Billsus, D. Machine Learning for user modeling. User
Modeling and User-Adaptive Interaction 11 (2001) 19-29.
16. Widmer, G.: Tracking Changes through Meta-Learning. Machine Learning 27 (1997)
256-286.
17. Widmer, G. and Kubat, M.: Learning in the presence of concept drift and hidden
contexts: Machine Learning 23 (1996) 69-101.
Prediction of Navigation Profiles
in a Distributed Internet Environment
through Learning of Graph Distributions
Dirk Kukulenz
Kiel University, Institute of Computer Science,

Preusserstr. 1-9, 24105 Kiel, Germany
[email protected]
Abstract. Collaborative filtering techniques in the Internet are a means

to make predictions about the behaviour of a certain user based on the
observation of former users. Frequently in literature the information that
is made use of is contained in the access-log files of Internet servers stor-
ing requested data objects. However with additional effort on the server
side it is possible to register, from which to which data object a client
actually navigates. In this article the profile of a user in a distributed
Internet environment will be modeled by the set of his navigation deci-
sions between data objects. Such a set can be regarded as a graph with
the nodes beeing the requested data objects and the edges being the de-
cisions. A method is presented to learn the distribution of such graphs
based on distance functions between graphs and the application of clus-
tering techniques. The estimated distribution will make it possible to
predict future navigation decisions of new users. Results with randomly
generated graphs show properties of the new algorithm.
1 Introduction
In many applications in the field of Internet research it is important to estimate

the relevance of data objects available in the Internet for a specific user or a
group of users. In the field of content-based learning the estimation is based on
the behaviour of a specific user. Collaborative filtering techniques make it possible
to learn from former usages of other users in order to make predictions for a new
user. These estimations can e.g. be used to make navigation on an Internet site
easier, to improve the quality of web sites or to find groups of consumers or
interest groups.
In [10] a procedure is presented to apply a collaborative filtering technique
for the creation of index lists, i.e. new web pages containing lists of hyperlinks
relevant for a certain topic, that are based on sets of requested data objects. In
[2] a navigation support system is presented that learns from search words and
browsing decisions of users, applying a reinforcement learning technique. In [14]
and [12] techniques for presending documents on the WWW are described that
apply different kinds of Markov-learning techniques.

234 D. Kukulenz
The information that is known about a specific user in the case of [10] and
[14] is the log data of Internet servers. Each request is stored in the so-called
access-log file containing information about the time of a request, the IP-address
of a client and the (IP-address of the) requested data object. However, different
caching strategies are used in the Internet, with the intention to reduce net
traffic and to increase the speed of requests. As a consequence, not all requests
of clients actually reach the original server. Thus only a subset of requested data
objects of a specific client is known on the server’s side.
In [14] the actual navigation path is estimated using the access-log informa-
tion. Here we will however use an idea presented in [2] to register the actual
set of navigation decisions of a client on the server’s side. A specially developed
proxy server in the connection between server and client modifies each requested
web page in a way that all hyperlinks point to that proxy server. The new links
contain additional information like the originally requested page, the page where
the link is located and and an id-number assigned to the client. By this means
navigation decisions of Internet users on the considered website can be registered
on the server’s side. This method makes it also possible to register the number
of navigation decisions in a distributed Internet environment, i.e. the navigation
decisions between data objects on a number of Internet servers.
Our collaborative filtering procedure is based on these sets of navigation de-
cisions of users. In the field of data mining algorithms are presented to find sets
with high frequencies [1]. Related to that, in [6] an algorithm is presented to
find frequent navigation sequences in the Internet. The approach described here
is based on the distances between patterns. A set of navigation decisions can be
regarded as a set of directed edges between data objects. These edges constitute
a graph structure with vertices being the requested data objects and the edges
being the decisions. In the field of pattern recognition different distance functions
between graph structures are presented e.g. in [9], [4], [13]. We will use one of
these functions together with an application of nearest neighbourhood cluster-
ing [7] to estimate the shape of the distribution of the graph profiles. Knowing
this distribution simple classification procedures can be applied to classify a new
profile and thereby to predict future decisions. The advantage of this technique
compared to Markov models, as presented e.g. in [14], is that we don’t have to
consider the order of a Markov process. Such a predefinition may cause classifi-
cation errors or otherwise cause an unnecessary increase of complexity.
In the next section the technique for the estimation of graph distibutions
and the prediction technique of future navigation decisions will be described. In
Sect. 3 some estimation examples with randomly generated graphs are presented
showing properties of the distribution estimation and the prediction technique.
Section 4 gives a summary and mentions further research issues.
Prediction of Navigation Profiles in a Distributed Internet Environment 235
2 Estimation of Graph Distributions

2.1 Definitions and Model
As described in the introduction, the information that can be acquired about a
specific Internet user on the server’s side with the described method is the set
(or at least a subset) of his navigation decisions. These navigation decisions take
place between certain data objects being available on the considered Internet
site, like web pages, images, scripts, etc . Let ’D’ denote the set of data objects
in the Internet site, having an (own) URL address.
A user profile, measured by the agent, is then a graph structure:
Definition 1. A (profile-) graph or navigation profile is a 4-Tupel G=(V,E,µ,ν).
V is a set of nodes and E ⊆ V × V is a set of edges. Function µ : V → LV ⊂ D
assigns labels to the nodes. Function ν : E → LE assigns labels to the edges.
Let < G > be the set of all graphs following the preceding definition. This set
will be denoted as ’graph space’ based on D. Let {G} ⊆< G > denote a set of
graphs.
LV is a subset of D or a set of pointers to D. The edges considered here are in
the most common case hyperlinks that are present on certain web pages, Java
applets or scripts. However, with the help of a search engine, the user can get
from one data object to possibly any other object.
In the following sections the definitions of a subgraph, a graph and sub-
graph isomorphism, graph-edit operations and an error correcting subgraph iso-
morphism are used that are common in the field of graph theory or artificial
intelligence and that are presented e.g. in [3] and [9].
2.2 Characterizations of Graph Distributions

It is our aim to classify a new profile graph according to a set of former profiles
supplied by users. For this purpose it is helpful to know the distribution of graph
profiles or at least to get an idea of the shape of this distribution. It is possible
to regard < G > as a discrete set and to assign a probability value to each
element depending on the relative frequency. However, graphs may be similar
according to certain aspects which may not be taken into account by the discrete
formulation.
It is very likely that people having the same question in mind produce sim-
ilar navigation profiles that are however slightly distorted because of Internet
caching, different starting points or different searching strategies. Vice versa,
similar profiles are likely to result from similar questions or intentions of users
which is the main assumption we make [5], [8]. We therefore assume that the
profiles are distributed in a way that one or a number of profiles in some ’places’
in the graph space have a high likelihood and the other profiles being more and
more distant from one of these ’central’ profiles have a decreasing likelihood with
respect to a distance function that will be defined in Sect. 2.3. This distribution
can then be characterized by a function:
Charac1 : {G} −→ {1, ..n}
236 D. Kukulenz
Here, every profile is associated with one of the clusters. Another method is to
consider the centers of the clusters and to take into account some characteristics
of the inner cluster structure. A characterization of the graph distribution is
then the set of these cluster properties:

Charac2 : {(µi , σi , Ai )},
i=1,..n
where µi is the center graph of cluster i, σi is a measure for the distribution

within the cluster, e.g. the mean value of the distances of the elements in the
cluster i from the center element µi and Ai is the number of elements in the
cluster. The center values µi can easily be found from Charac.1 by determining
the element in the cluster with the smallest sum of the distances to all the other
elements in the same cluster.
In the following we will use a simplification of the graph distribution charac-
terization Charac2 by taking only the center elements into account.
We define: Charac2 : i=1,..n {µi }.
2.3 Graph Metrices

The ’shape’ of the graph distribution as being characterized by Charac1 or
Charac2 depends strongly not only on the data elements but also on the distance
measure between graphs. Several definitions of graph distances are known from
the field of pattern recognition.
A simple idea to define such a distance function is to count the number of
identical nodes. To achieve a better segmentation of the set of graph profiles
however, the structure of the connections, i.e. edges in the graphs, should be
taken into account, too. A measure for such a structural similarity is the size
of the largest common subgraph. In [4] it was shown that for two non-empty
graphs G1 und G2 and the largest common subgraph lcS(G1 , G2 ) the function
|lcS(G1 ,G2 )|
Definition 2. d(G1 , G2 ) := 1 − max(|G1 |,|G2 |)
has the mathematical properties of a metrics ( |.| denotes the number of nodes in
a graph). A similar graph distance was defined in [13]. The disadvantage of this
metrics is that possible similarities between different nodes can’t be taken into
account. Such similarities between the type of nodes that are considered here,
i.e. data objects, have been examined for textual data in the field of information
retrieval [11]. They are important for the automatic indexing of web pages for
the realization of search engines. One well-known distance measure is the tfidf-
Norm, in which text pages are converted into vectors of weights of words that
can be compared with the help of the cosine between the vectors.
A distance measure for two graphs G1 and G2 making it possible to take such
similarities into account is the following function, where ∆ is a set of graph-edit
operations and C is a cost function for the edit oprations as described in [9]:
Definition 3. d(G1 , G2 ) := min∆ { C(∆) | there exists an error-correcting-
subgraph-isomorphism f∆ from G1 to G2 }
The distance function in definition 3 is not symmetric. In order to create a

symmetric distance function it is possible to take the minimum of d(G1 , G2 ) and
d(G2 , G1 ).
In the following we will however work with the distance in definition 2 which
is easier to implement and faster to compute.
2.4 Estimation of a Graph Distribution
The previously defined metrices or distance functions can now be applied to

estimate the shape of a graph distribution considering the distribution charac-
terization Charac2 in Sect. 2.2. The navigation graphs can be clustered using a
common clustering technique like nearest neigbourhood clustering as described
in [7] and by using one of the distance functions given in Sect. 2.3. Further in-
vestigations concerning the shape of the inner cluster distributions according to
Charac2 in Sect. 2.2 can then be made.
In order to measure the quality of such a distribution estimation it may be
helpful to determine the distance between a real distribution that is known in
advance and an estimation of this distribution. Let G1 , ..Gn be the elements
in {G}, H1 , ..Hm (m ≤ n) be the real cluster centers characterizing the graph
distribution and d(G1 , G2 ) be the distance between two graphs according to
one of the definitions in Sect. 2.3. Let δ(G) := min{d(G, Hj )|j = 1, ..m} with
G ∈ {G}.
Definition
4. Given an estimation of the cluster centers Ĥ1 , ..Ĥm , let err :=
i=1,..m δ( Ĥ i ).
Obviously, err decreases, if the estimation result gets better i.e. the estimated
cluster centers move towards the real ones.
Knowing the estimated distribution of navigation graphs we can describe a
prediction technique to find future navigation steps of a specific user if we assume
that the new profile follows the same distribution as the former ones. One way
is to compare the new navigation profile to the estimated cluster centers and
to find the closest center. Given the estimated cluster centers Ĥ1 , ..Ĥm and the
new profile G, in this method d1j := d(G, Ĥj ) has to be minimized in j where
d(G, Ĥj ) is a distance of G to the cluster center Ĥj as defined in Sect. 2.3. This
center element Ĥj can then be expected to have a high relevance for the user.
A further possibility is to take into account the absolute probability that a
user profile belongs to a cluster. This probability can be estimated by the relative
number of elements in the cluster. The minimization of d2j := d(G, Ĥj ) 1+A1j /A in
j takes this absolute probability into account, where A is the number of observed
profiles, Aj is the number of patterns in cluster j. These functions will be tested in
the following section. The basic steps of the estimation and prediction algorithm
are:
• Data acquisition
238 D. Kukulenz
Distribution estimation (offline)
• Computation of the distance matrix

• Clustering procedure
• Distribution estimation
Prediction (online)
• Registration of a new (partial) user profile

• Computation of distances to the estimated cluster centers
• Classification of the new profile according to the estimated distribution and
a classification function
• Prediction of future navigation decisions according to the classification re-
sults
The distribution estimation as described above can be done offline. For most of
the applications like navigation support, the prediction step has to be done in
real-time.
3 Simulation Results with Randomly Generated Graphs
It is our aim to show some of the properties of the described distribution esti-
mation and classification with randomly generated navigation profiles where the
distribution (i.e. Charac2 in Sect. 2.2) of the original data is known in advance
and can be compared to the estimation results. The simulation process starts
by defining a graph space < G > as defined in Sect. 2.1. A number of graphs
will then be computed randomly with equal distribution, the number of nodes
being identical and a fix number of edges. These graphs represent the real center
graphs. Then a sequence of graphs will be computed presenting the simulated
graph data. Each graph is obtained by randomly choosing one of the real center
graphs and a number for the label errors. The error value is chosen according to
a discrete Gaussian N (0, σ) distribution. The simulated graph is computed by
changing a number of node labels of the center graph, equal to the number of
label errors.
In Fig. 1 the dependence of the estimation quality according to definition 4
on the number of graphs in the sequence of navigation profiles is shown. The
number of elements in D is 30, the number of nodes in each graph is 25, with 30
edges. The graphs were computed from 2 original graphs (m=2), constituing the
real distribution characterization. The number of identical simulations was 10.
In Fig. 1 each value is the mean value of the estimation errors in the identical
simulations. The graph metrics applied here for the clustering and the estimation
quality measurement is the subgraph metrics in definition 2. As can be expected,
the estimation error decreases, when the number of graphs increases since more
information about the distribution is available for the estimation process.
In a second experiment we examined the prediction quality supposing that
the distribution characterization is already known. A number of profiles were
0.35
0.3
0.25
average estimation error
0.2
0.15
0.1
0.05
0
0 2 4 6 8 10 12 14
number of navigation graphs
Fig. 1. Dependence of the estimation error on number of graphs
1
generated, following this distribution as described above. The percentage ( × 100 )
of missclassifications was determined, denoted as ’classification error’.
Figure 2 shows the classification error based upon the minimization of d1
(•) and d2 (+) in Sect. 2.4. In the experiment the deviation of label errors is
changed. As can be seen, the prediction based upon minimization of d2 shows
better results for higher values of the label error. This result was expected since
more information about the shape of the distribution is used in the case of d2.
4 Conclusion and Further Aspects

In this article an estimation technique was presented that applies clustering of
a set of graphs based on a definition of a distance between graphs. This pro-
cess provides a characterization of the distribution of graphs which is difficult to
describe directly. This characterization can then be applied for a relevance esti-
mation presuming that the new navigation graph follows the same distribution.
Some characteristics of the algorithm like the convergence for an increasing
number of patterns were shown by means of randomly generated graphs. The
advantage of the use of simulated data is the knowlege about the distribution
that can’t be known for real data.
Compared to Markov modelling this estimation method has the advantage
that a multi-step-prediction can easily be done and that not only sequences of
navigation steps but also navigation graphs i.e. sets of navigation steps can be
taken into account. A graph modelling of user decisions can be of advantage if
e.g. caching strategies in the Internet cause distorted navigation profiles or if the
actual set of navigation decisions has to be considered.
One problem to discuss when recommending navigation decisions is the so-
called ’snowball effect’. If the system learns a wrong path and presents it to other
240 D. Kukulenz
0.6
0.5
0.4
classification error
0.3
0.2
0.1
0.1
0 2 4 6 8 10 12 14
stand.dev. label errors
Fig. 2. Classification experiment of user profiles by minimizing d1 (•) and d2

(+) in Sect. 2.4.
users, they may also follow this wrong path and the system will learn again the
wrong path. This problem however becomes only important if a high percentage
of users actually use the support system. The registration of navigation decisions
described in Sect. 1 is also possible for users who don’t apply the support system.
There are more refined methods conceivable to describe a distribution of
graphs. A first improved method is given in definition 2.2, however further im-
provements should be developed. Different and more refined graph distances can
be defined e.g. taking into account node distances as described in definition 3.
Additionally the prediction quality has to be examined closely for real data. The
time requirements of the prediction algorithm are very important because this
step has to be done in real-time if the prediction result is used e.g. for a naviga-
tion support tool. Further improvements of the system with respect to learning
from additional information about a user or the Internet site are of interest.
References
1. R. Agrawal, T. Imielinski, and A. Swami. Mining association rules between sets of
items in large databases. In Proc.of the ACM SIGMOD Conference on Management
of Data, 1993.
2. R. Armstrong, D. Freitag, T. Joachims, and T. Mitchell. Web watcher: A learning
apprentice for the www. In AAAI Spring Symposium on Information Gathering
from Heterogeneous, Distributed Environments, pages 6–12, 1995.
3. Bollobas. Graph theory. Springer, 3 edition, 1999.
4. H. Bunke and K. Shearer. A graph distance metric based on the maximal common
subgraph. In Pattern Recognition Letters, volume 19, pages 255–259, 1998.
5. E. Carmel, S. Crawford, and H. Chen. Browsing in hypertext: A cognitive study.
In Transactions on System, Man and Cybernetics, volume 22, pages 865–883, 1992.
6. M. Chen, J.S. Park, and P.S. Yu. Data mining for path traversal patterns in a web
environment. In Proc. of the 16th ICDCS, volume 16, pages 385–392, 1996.
7. B.S. Everitt. Cluster Analysis. Edward Arnold, 3 edition, 1993.
8. C. Hoelscher and G. Strube. Web search behavior of internet experts and newbies.
In World Wide Web Conf, volume 9, 2000.
9. B. Messmer and H. Bunke. Efficient graph matching algorithms for preprocessed
model graphs. PhD thesis, Bern University, 1996.
10. Mike Perkowitz and Oren Etzioni. Towards adaptive web sites: Conceptual frame-
work and case study. Artificial Intelligence, 118(1–2):245–275, 2000.
11. G. Salton. Developments in automatic text retrieval. Science, 253:974–979, 1991.
12. R. Sarukkai. Link prediction and path analysis using markov chains. In Intern.
World Wide Web Conf., 2000.
13. W.D. Wallis, P. Shoubridge, M. Kraetz, and D. Ray. Graph distances using graph
union. In Pattern Recognition, volume 22, pages 701–704, 2001.
14. I. Zukerman, D.Albrecht, and A.Nicholson. Predicting users’ requests on the www.
In UM99 – Proceedings of the Seventh International Conference on User Modeling,
1999.
Educo - A Collaborative Learning Environment
Based on Social Navigation
Jaakko Kurhila1 , Miikka Miettinen2 , Petri Nokelainen2 , and Henry Tirri2

1
University of Helsinki, Finland
[email protected]
2
Complex Systems Computation Group
Helsinki Institute for Information Technology, Finland
{Miikka.Miettinen, Petri.Nokelainen, Henry.Tirri}@hiit.fi
Abstract. Web-based learning is primarily a lonesome activity, even

when it involves working in groups. This is due to the fact that the ma-
jority of web-based learning relies on asynchronous forms of interacting
with other people. In most of the cases, the chat discussion is the only
form of synchronous interaction that adds to the feeling that there are
other people present in the environment. Educo is a system that tries to
bring in the sense of other users in a collaborative learning environment
by making the other users and their the navigation visible to everyone
else in the environment in real-time. The paper describes Educo and
presents the first empirical evaluation as Educo was used in a university
course.
1 Introduction
When Dourish and Chalmers introduced the concept of social navigation, they
stated it to be “navigation because other people have looked at something” [4].
The concept has evolved since then (see e.g. [8] for an overview of the topic),
and various categories of social navigation have emerged (direct – indirect [3],
intended – unintended [6]). Today, many of the systems incorporating social
navigation use collaborative filtering. It means that these “systems provide the
user with recommendations of their likely interest in data items on the basis
of ‘interest matches’ derived from ratings from the set of users” [5]. Examples
of such recommender systems include various web-stores, where the customer is
recommended a product based on the actions of previous customers.
In the area of web-based learning, recommender systems based on collabo-
rative filtering can have a positive impact on the overall learning process. How-
ever, these systems do not address the problem of the feeling of being alone
in a web-course. Commercial or even research-level course delivery systems [1]
have rarely taken this into consideration. There are various collaborative virtual
environments [2] that include the “feeling” of other users, but the solutions are
not necessarily directly applicable to web-based learning.

Educo - A Collaborative Learning Environment Based on Social Navigation 243
Educo is a system that visualizes other live users currently present in the
learning environment. Navigation and initiating synchronous or asynchrornous
discussions have been made as simple as possible. The movement from one docu-
ment to another in the environment is updated for every participant in real-time,
thus adding to the feeling of truly live action. Research on workspace awareness
has notified this as an important issue in groupware [12].
The Educo system has been tested in one advanced university course. A
detailed description of the system, the study setting and the results are discussed
in the subsequent sections.
2 Educo
Before going into the system description of Educo, a few concepts should be
clarified. A user is a learner participating in a course in a web-based learning
environment using Educo. A document is an HTML-file within Educo that is
visible to the users. Documents have visual representations on the screen, and
they can be grouped into document clusters. An instance of Educo means a
fixed set of document clusters, i.e. a unique course within the Educo learning
environment. An administrator of Educo is a person responsible for an instance
of Educo. It is typical that the administrator of Educo is the teacher of the
course. Only an administrator can add users or documents to an instance of
Educo. The document clustering is also conducted by the administrator. The
administrator can also assign group information to the users, thus forming var-
ious user groups. The group information and the document clustering are static
unless the administrator makes the required changes to the system.
The user interface of Educo consists of six views of which only one is visible
at a time. The views are map, chat, search, alarm, preferences and help. The
screen layout when using Educo is presented in Fig. 1. The six views of Educo
are presented in a tool resembling a handheld computer (upper-left corner in
Fig. 1, now in “map” view). The largest area is reserved for documents gathered
into an instance of Educo (right-hand side of the web-browser in Fig. 1). The
space below the Educo tool is for the comments provided by the users.
“Map view” presents the document clusters of an instance of Educo. Every
document is visible, and the clusters are distinct. Documents are presented as
paper-icons. The users in an instance of Educo are presented as coloured dots.
The dot is located next to the document the user is currently viewing. When
a user places the mouse pointer on top of a document or a dot representing a
user, a tool tip text appears showing the name of the person or the document.
In Fig. 2 the pointer is on a document called “Where did all the people go?”.
Double clicking a document opens the document into the right-hand side of the
browser window and moves the dot representing the user to a corresponding
location on the map view of every user in the Educo instance.
The rectangle at the bottom of the map view in Fig. 2 is a magnifying glass
included in the map view. The purpose of the magnifying glass is to make it
easier for the users to click on the dots or documents while navigating.
244 J. Kurhila et al.
Fig. 1. The user interface of Educo.
Fig. 2. The “map” and “chat” views of Educo.

The colours of the dots indicate different group memberships or types of user
profile. The groups are assigned according to some metric the administrator of
the Educo instance wishes to choose. For example, the groups can be assigned
based on the students’ interest in various topics within the course topics.
The documents change their colour on the map depending on how much they
have been read in relation to the other documents. The colours range from bright
(heavily read) to dimmed (not heavily read), as presented in Fig. 2. This way
the user can get the “footprint” information at a glance and does not have to
stay online and constantly watch where the other users navigate.
When in map view, clicking a user or a document symbol once selects it
for further use. The further use can, for example, be a “Chat”, which is the
second view (Fig. 2). Any user can easily initiate a chat discussion with other
users simply by clicking the corresponding user symbol and then clicking the
“Connect” button in the chat view. The chat can be initiated with one to n
other users. The restriction is that one person can be involved in only one chat
channel at a time.
Chat is a form of synchronous communication in Educo. Depending on the
situation, asynchronous discussions might sometimes be more useful. Therefore,
every user in Educo has the possibility to write a comment when viewing a
document. The comment is visible to users navigating to that document, i.e. the
comments are document-specific. Other users can comment on the comment,
thus continuing the chain of comments as illustarted in Fig. 1.
The third view is “Search”. Users can search for other users and documents
in an instance of Educo. When a user searches another user or a document, the
results are shown textually in search view (Fig. 3) and graphically in map view
by highlighting the corresponding user or document (the same effect as clicking
a user or a document once in map view). The highlighting is illustrated in Fig. 2
where the magnifying glass is on a highlighted document. In addition to finding
documents on the map, the operation of the search makes it easy to initiate a
chat with a specific user.
“Alarm view” gives users the possibility to set up an alarm that is triggered
if the requested condition occurs. For example, if a user seeks another user who
is also interested in a certain document or topic, he or she can tell the system
to give a notifying message when someone else arrives at the document. Alarm
is a versatile tool, since the user can make different combinations of the three
possible triggering events: users, group members and documents. Figure 3 shows
the alarm view where the user is about to set the alarm to trigger if a user named
“Demo user” or someone from group 2 arrives at the document named “Adaptive
and Intelligent Technologies”. Other highly used combinations include notifying
if a certain user enters the system or if someone arrives to a specific document.
The last two views are “Preferences” and “Help”. While viewing “Prefer-
ences”, the user is allowed to change personal settings in the system. The set-
tings include preferred nickname within the chat (in case the user wants to stay
anonymous in the chat), and whether the user is visible to other users. If the
user chooses to make her own navigation invisible in Educo, it automatically
Fig. 3. The “search” and “alarm” views of Educo.
means that she cannot see the other users. Help view provides information about
the system in general (Fig. 4)
Fig. 4. The “preferences” and “help” views of Educo.

3 Educo Architecture
From a technological point of view, Educo consists of a socket server, a Java
applet for every user and several CGI-scripts. The most important task of the
server is to keep track of the state of the distributed system and inform the
clients as changes occur. The changes include navigation from one document to
another. If one of the users moves to another page, the new location has to be
sent to everyone currently present in Educo. The implementation of this kind
of communication scheme without delays requires that the clients maintain an
open socket connection to the server throughout the session.
Besides managing connections as well as user and document information,
the server forwards chat messages to their recipients and takes care of various
bookkeeping activities related both to its own functioning and logging the users’
actions for research purposes.
To avoid copyright issues and to make the use of Educo simpler for the
administrator (course teacher), we have taken the approach that the documents
(HTML-files) for a particular instance of Educo do not need to be copied to the
Educo server. Instead, they can be located anywhere on the Web. To operate
properly, the server still needs to know which document the user is reading to
be able to send that information to all the other users in that instance. The
operation has to work even when the users navigate along the hyperlinks in the
documents and are not using the map view by double-clicking the document
symbols. We have solved this problem by using the Educo server as a proxy. It
means that the documents are routed through the server instead of being sent
to the client directly from their actual location. This requires two additional
operations: clients are informed about the new location of the user and all of
the links in the document are changed so that they point to their destination
indirectly through the proxy. If the user then clicks one of these links, the same
procedure is repeated in a recursive fashion.
Commenting the documents in an instance of Educo is technically based on
the use of ordinary HTML-forms. Each document has an associated comment
file, which is opened to its designated frame every time a user navigates to that
document. The server keeps track of modifications and the visits of individual
users. This way the documents that have been commented after the last visit
can be distinguished visually from those that contain only comments the user
has already seen.
As mentioned above, the documents change their colour on the map view
depending on how much they have been read in relation to the other documents.
The total time all users have spent reading each document is recorded by the
server on an hourly basis. The change in the colour of an individual document
is determined by the distance of its moving average for the last 24 hours from
the same average for all the documents. Heuristic methods have been devised
to make the colouring of the documents operate sensibly at the very beginning
of the course and during the silent periods (weekends, holidays etc.). It is also
appropriate to eliminate the disturbing effect of very long reading times by
setting an upper bound after which the additional time is ignored. The value for
the upper bound can be the same for all documents or adjusted according to the
differences in the lengths of the documents.
4 Study Setting
The course structure. The first empirical evaluation of Educo was conducted
during a course entitled “Web-based learning” given at the University of Helsinki,
Finland. The course is an advanced course in Computer Science studies. Twenty-
four students participated in the course, some of them adult learners with varying
backgrounds and degrees but most of them were CS majors. The type of the
course was a “seminar” which means that the students have to pick a topic,
prepare a 10-page paper on the topic and present it to the teacher and other
students in the course. In addition, there were some short weekly assignments
to complete.
There were only two face-to-face meetings during the course. The first was
an initial meeting where the structure and requirements for the course were ex-
plained and the Educo system was introduced. The second face-to-face meeting
was the final meeting where the students presented their papers. Everything else
between the initial and final meeting was conducted on-line using Educo.
Only the real-time social navigation of the system was studied and not the
use of the “footprint” information. Because of the small student population
participating in the course, we fixed a primary time slot to make sure that there
would be people in Educo at the same time. However, the time slot was not
restrictive in any way.
Forty-three documents were first gathered into the instance of Educo to
serve as a starting point for the topics in the course. The documents were clus-
tered according to the six general areas to be covered: history of web-based
learning, society and web-based learning, research findings, teaching and study-
ing in a web-based course, course delivery systems, and providing adaptation in
educational systems.
The data set for the study was gathered in three stages: (1) a pre test after
the start of the course measured motivational level and learning strategies, (2)
users’ actions were logged during the course, and (3) a post test after the course
measured how students’ expectations met the reality.
Pre test. Motivational profiling in this study is based on the Motivated Strate-
gies for Learning questionnaire (MSLQ), which is developed on the basis of a
motivational expectancy model [7]. MSLQ measures both motivational factors
and learning strategies. The motivation section (A) of MSLQ consists of 17 items
that were used to assess students’ value for a course, their beliefs about their
skill to succeed in the course, and their anxiety about tests in the course. A
5-point Likert-scale ranging from 1 (”Not at all true of me”) to 5 (”Very true of
me”) was used for all items.
The theoretical model of motivation [10] is constructed out of six factor solu-
tion: (1) Intrinsic goal orientation, (2) Extrinsic goal orientation, (3) Meaning-
fullness of studies, (4) Control beliefs, (5) Efficacy beliefs, and (6) Test anxiety
Fig. 5. A sample of the filtered Educo log file.
[11]. We expected to find a similar structure in the sample data and thus to be
able to construct sensible motivational groups.
Users’ actions during the course. The user log from Educo (time stamp, user
id, action) was recorded during the course from September 24 to November 20,
2001. The filtered log file of 1832 recorded actions (Fig. 5) was analysed in order
to find what effect Educo’s visual social navigation information had on the
users’ navigation behaviour.
The main level problem was operationalised into the following two sub-level
propositions: firstly, did the users in general prefer occupied (someone else at the
document, “occupied” = 1 in Fig. 5) documents over unoccupied (the document
is “free” from other users, “occupied” = 0) ones? Secondly, did the users based
on their pre test motivational group membership (”id group”) prefer occupied
documents over unoccupied ones?
Post test. An email survey consisting of 17 open propositions was conducted two
weeks after the course in December 2001. Propositions measured users’ experi-
ences and expectations towards web-based education together with attributes
related to Educo (usability issues, user interface, functionality etc.).
5 Results
Pre test. The analysis of the “A” section of the motivational pre test ques-
tionnaire was carried out with a Bayesian dependence modeling tool named
B-Course1 [9]. The results indicated that the theoretical model of six factors [11]
was a viable solution for this small number data set. Based on the motivational
level scores on six dimensions, respondents were divided into three groups:
– Group 1 (“Blue”, N=10) characteristics: (2) Extrinsic goal orientation, (6)

test anxiety and (3) meaningfullness of studies.
– Group 2 (“Green”, N=8) characteristics: (5) Efficacy beliefs, (1) intrinsic
goal orientation and (3) meaningfullness of studies.
– Group 3 (“Red”, N=6) characteristics: (4) Control beliefs and (1) intrinsic
goal orientation.
1
B-Course URL http://b-course.cs.helsinki.fi
The classification accuracy of the theoretical model [10] was confirmed with both
a linear and nonlinear discriminant analysis (87.5% of original and 75.0% of
cross-validated grouped cases were correctly classified). There was no statistical
signifigance between the group memberships of male and female respondents.
The group descriptions with clear explanations were published on the course
web-site for all the participants, so that the students were able to use the infor-
mation when completing weekly assignments or choosing a study partner.
Users’ actions during the course. After filtering out the entry document from
the log file, the analysis of the data (total number of logged events = 1832)
indicated that the users preferred unoccupied documents (943 requests, 51.5%)
over occupied ones (889 requests, 48.5%). The number of simultaneous readers in
occupied documents varied from one to six with the following request frequencies:
one reader (501 requests, 56.5%), two readers (268 requests, 30.2%), three readers
(75 requests, 8.5%), four readers (35 requests, 3.9%), five readers (7 requests,
0.8%), and six readers (1 request, 0.1%). The results revealed no gender-related
differences.
The results proved (χ2 = 13.29, p=0.01) that the respondents selected docu-
ments based on their pre test motivational group membership (Fig. 6). Members
of the “Blue” group preferred unoccupied documents (N=496, 55.5%) over oc-
cupied ones (N=398, 44.5%). Students belonging to the “Green” group made no
distinction between documents. This result is in balance with the group charac-
teristics that emphasize intrinsic goal orientation and efficacy beliefs. Members
of the “Red” group preferred occupied documents (N=234, 55.1%) over unoccu-
pied ones (N=191, 44.9%) indicating a tendency towards social navigation.
The third part of the log file data analysis focused on the reading times per
document. The values of “Time” variable were categorised into six classes: 0–2,
2–5, 5–10, 10–30, 30–60 and 60–90 minutes. Reading sessions that lasted over 90
minutes were excluded from the analysis. This part revealed interesting group
specific results. Members of the “Green” group spent the least time (x=2.7 min)
per document compared to other groups (χ2 = 19.38, p=0.04). This result sup-
Fig. 6. Social navigation preferences based on pre test motivational group membership.
ports the “result-oriented” label of the group members. There was no difference
in reading times between the “Blue” (x=3.2 min) and the “Red” (x=3.6 min)
group.
Post test. The third phase of this study was to analyse the propositions of the
post test (the total number of propositions was 17). The total number of answers
to the post test questionnaire was 17 (71%) out of 24. The sample data consisted
of five female and twelve male students.
Results of the post test show that EDUCO was seen as a useful tool in
the matters like adaptation to respondents learning, cognitive and motivational
strategies, and means to implement collaborative actions.
“It was very useful to see what documents other users were reading, it
gave me many hints and saved time.”
“It was truly nice to be able to see what the most interesting document
at the moment is and who is reading it.”
“Actually, in several cases I wanted to start a chat conversation with
someone reading the same hyperdocument with me . . . I guess this is
social navigation?”
The presence of EDUCO increased task-related participation and was a valued
tool for those who had difficulties to participate in face-to-face meetings:
“The learning material was easy to access.”
“EDUCO gives more flexibility to the studying process.”
“It was possible for me to participate in this course and carry out all
those tasks regardless of my domicile.”
The real time interaction of EDUCO also elicited negative comments:

“EDUCO hindered the formation of REAL social contacts!”
“Chat never beats traditional face to face meetings.”
EDUCO’s tools for seeking work mates (group membership, search function)
were truly useful for most of the respondents:
“I was in a blue group, and when another blue was looking for a mate,
I replied instantly. He had already chosen an article, I glanced at it and
found that it was suitable for me too.”
“I had a group proposal via email message. As my forthcoming work mate
had the same colour as I did, it was easy to make the decision to start
collaboration. Afterwards I thought that I agreed so quickly because of
the same motivational group, normally it takes more consideration. But
to be honest, the topic was the most important factor.”
6 Conclusions
Educo, a system based on social navigation for web-based learning, has been
beneficial for some users. Moreover, using a system like Educo opens up other
important issues as well. The teacher of a web-course has an opportunity to

re-think the learning process as a whole. Recent pedagogical approaches such
as just-in-time learning can benefit, for example, from the possibility in Educo
to form ad hoc study groups based on the motivational group information and
the interest expressed for a certain topic. The students themselves are in control
of forming the groups, thus shifting the emphasis towards peer-assisted and
learner-centered education.
References
1. Brusilovsky, P. and Miller, P.: Course Delivery Systems for the Virtual University.
In T. Tschang and T. Della Senta (Eds.): Access to Knowledge: New Information
Technologies and the Emergence of the Virtual University, pages 167-206. Amster-
dam: Elsevier Science (2000).
2. Churchill, E., Snowdon, D. and Munro, A.(Eds): Collaborative Virtual Environ-
ments. Springer (2001).
3. Dieberger, A.: Social Navigation in Populated Information Spaces. In A. Munro,
K. Höök and D. Benyon (Eds.), Social Navigation of Information Space, pages
35-54. London: Springer (1999).
4. Dourish, P. and Chalmers, M.: Running Out of Space: Models of Information
Navigation. In Proceedings of HCI’94 (1994).
5. Dourish, P.: Where the Footprints Lead: Tracking Down Other Roles for Social
Navigation. In A. Munro, K. Höök and D. Benyon (Eds.), Social Navigation of
Information Space, pages 15-34. London: Springer (1999).
6. Forsberg, M., Höök, K. and Svensson, M.: Footprints in the Snow.
Position paper for 4th ERCIM Workshop User Interfaces for All.
http://ui4all.ics.forth.gr/UI4ALL-98/forsberg.pdf (1998).
7. Garcia, T. and Pintrich, P.: Regulating Motivation and Cognition in the Classroom:
The Role of Self-Schemas and Self-Regulatory Strategies. In D.H. Schunk and
B.J. Zimmerman (Eds.), Self- Regulation of Learning and Performance: Issues and
Educational Applications. Hillsdale, N.J.: Erlbaum (1994).
8. Munro, A., Höök, K. and Benyon, D.: Footprints in the Snow. In A. Munro, K.
Höök and D. Benyon (Eds.), Social Navigation of Information Space, pages 1–14.
London: Springer (1999).
9. Nokelainen, P., Silander, T., Tirri, H., Tirri, K and Nevgi, A.: Modeling Stu-
dents’ Views on the Advantages of Web-Based Learning with Bayesian Networks.
Proceedings of the 10th International PEG2001 Conference, pages 202–211 (2001).
10. Ruohotie, P.: Conative Constructs in Learning. In P. Pintrich and P. Ruohotie
(Eds.) Conative Constructs and Self-Regulated Learning, pages 1-30. Saarijärvi:
Learning and Change Series of Publications (2000).
11. Ruohotie, P., Nokelainen, P., Tirri, H. and Silander, T.: Modeling Student Mo-
tivation and Self-regulated Learning with Bayesian Networks. In P. Ruohotie, P.
Nokelainen, H. Tirri and T. Silander (Eds.) Modeling Individual and Organiza-
tional Prerequisites of Professional Growth, pages 174–195. Saarijärvi: University
of Tampere (2001).
12. Wang, H. and Chee, Y.S.: Supporting Workspace Awareness in Distance Learn-
ing Environments: Issues and Experiences in the Development of a Collaborative
Learning System. Proceedings of the International Conference on Computers in
Education (ICCE2001), pages 1109–1116. Seoul, Korea: Incheon National Univer-
sity of Education (2001).
GOOSE: A Goal-Oriented Search Engine
with Commonsense
Hugo Liu, Henry Lieberman, and Ted Selker
MIT Media Laboratory

20 Ames St., E15-320G
{hugo, lieber, selker}@media.mit.edu
Abstract. A novice search engine user may find searching the web for informa-
tion difficult and frustrating because she may naturally express search goals
rather than the topic keywords search engines need. In this paper, we present
GOOSE (goal-oriented search engine), an adaptive search engine interface that
uses natural language processing to parse a user’s search goal, and uses “com-
mon sense” reasoning to translate this goal into an effective query. For a
source of common sense knowledge, we use Open Mind, a knowledge base of
approximately 400,000 simple facts such as "If a pet is sick, take it to the vet-
erinarian" garnered from a Web-wide network of contributors. While we can-
not be assured of the robustness of the common sense inference, in a substantial
number of cases, GOOSE is more likely to satisfy the user's original search
goals than simple keywords or conventional query expansion.
1 Introduction
The growth of available content on the World Wide Web makes it necessary for eve-
ryone to use tools, not experience, to find things. Major search engines like Google
and Yahoo have made great progress in indexing a large percentage of the content on
the web so that it is searchable. However, the user interface to the search process is
usually just a text input box waiting for input. The user interfaces in most of today’s
search engines still rely on a grammar of set operators and keywords, and for good
results, the user is expected to be able to fill the box with the right keywords and in
the right combination. This situation prompts the question: instead of having the
user conform to the search engine’s specifications, why not make the search engine
adapt to how the user most naturally expresses his/her information needs, so that even
inexperienced users can perform an effective search?
1.1 An Experiment
To learn some qualities an intuitive search engine interface should have, we asked
four search engine novices and four experienced search engine users to perform sev-

254 H. Liu, H. Lieberman, and T. Selker
eral tasks using the Yahoo search engine. Whereas experienced users chose precise
keywords likely to isolate the types of web pages they were looking for, novice users
reverted to typing their search goal into the keyword field in natural language. For
example, one search task that users were asked to perform was to find people on the
web who shared the user’s own interests. One novice user submitted the query: “I
want to find other people who like movies,” and obtained many irrelevant and un-
wanted search results on the topic of movies. In contrast, a more experienced user
formed the query: “ +‘my homepage’ +‘my interests’ +‘movies’ ” and was able to
get many relevant results. The experienced user chose not only a keyword (“mov-
ies”) on the topic of the search, but also two keywords (“my homepage”, “my inter-
ests”) differentiating the context in which the topic keyword should appear. In choos-
ing these keywords, the experienced user used her expertise to guide a series of infer-
ences from the search goal. In interviewing the user, we learned that the inference
chain, or thought process, that she went through looked something like this:
I want to find people online who like movies.

Movies are a type of interest that a person might have.
People might talk about their interests on their homepage
People’s homepages might start with “my homepage”.
This prompted us to further reasoning. As with all of the inference chains used by
the four experienced users, this inference chain has the following property: Most of
the steps in the inference chain are statements that arguably fall under the “common
sense” knowledge domain, things that most people know to be true (in this case, only
the last step is somewhat domain-specific knowledge); however, the knowledge of
how to connect these commonsense facts to infer a good search query is where search
engine expertise is required.
Even the experience of these few subjects point to out that novice searchers are
confusing the search engine with the approach that they naturally communicate with.
Simple improvements might be to: 1) allow the user to formulate the search query as
a statement of the user’s search goal, and from that, the search engine must make the
necessary inference to arrive at the appropriate keywords, and 2) allow the user to
express the search query in natural language.
To meet the second criterion, the search engine needs to have natural language
parsing capabilities. The first criterion is trickier. If we are to expect the search en-
gine to assume the burden of performing the inference, we might give it knowledge
about the world that most people know (commonsense), and also some knowledge
about what a good search query is, something that experienced search engine users
know (expertise).
GOOSE is a goal-oriented search engine organized around the concept of a search
goal. Enriched with commonsense knowledge, search engine expertise, and natural
language parsing capabilities, it assumes the burden of translating a user’s search goal
into a good query.
In this paper, we will first present some background on this project, followed by
descriptions of the GOOSE user interface and internal mechanism, a sample user
GOOSE: A Goal-Oriented Search Engine with Commonsense 255
scenario, and preliminary user test results. We will then proceed to discuss some
future work of personalizing the commonsense and conclude.
2 Background
Previous approaches to query improvement have for the most part employed three
techniques: 1) expanding the topic keyword using thesauri and co-occurrence lists [7,
10] 2) relevance feedback [11], and 3) using hand-crafted question templates [1].
Though the first approach shows promise for queries that return limited results, ex-
panding keywords does not necessarily improve the relevance of the search results.
The second approach does a better job of improving relevance, but complicates the
task model by adding additional search steps. In addition, neither the first nor the
second approaches address the weaknesses of keywords as the basis of the user inter-
face. The third approach, as used by Ask Jeeves [1], offers the user a more intuitive
natural language interface, but answerable questions must be anticipated in advance
and a template for each question must be handcrafted. For this reason, we don’t be-
lieve that this approach is easy to scale.
Our approach is significantly different from all the aforementioned approaches. In
our system, the original query is a natural language statement of the user’s search
goal, and the reformulation step involves natural language parsing of this statement,
followed by inference to generate the query that will best satisfy this goal. Unlike
thesauri-driven keyword expansion, our system is not merely adding new keywords,
but is actually performing inference and composing an entirely new search query that
would best fulfill the user’s goal. Compared with relevance feedback, the user inter-
face we propose is automatic, and does not require additional steps in the task model.
Finally, unlike handcrafted question templates, we believe that our approach of using
a freely available, ever-growing, and vast source of commonsense knowledge to per-
form reasoning over the original query is a more scalable approach, and allows for
many levels of inference, compared to the fixed, one-level of inference associated
with question templates.
2.1 Source of Commonsense
The idea of using commonsense reasoning to improve user interfaces has been exten-
sively explored by Minsky [5]. The commonsense knowledge used by our system
comes from the Open Mind Commonsense Project [8] – an endeavor at the MIT Me-
dia Laboratory that aims to allow a web-community of teachers to collaboratively
build a database of knowledge using diverse representations, and to explore ways to
use this knowledge to make computer applications more intelligent and context-
aware. Using the Open Mind Commonsense website, web collaborators input simple
facts about the world, expressed as simple English sentences, which are organized
into an ontology of commonsense relations.
2.2 Ordinary Commonsense vs. Application-Level Commonsense
When we refer to the commonsense knowledge used in GOOSE, we mean two things.
The first is ordinary commonsense, which encompasses the things that people nor-
mally consider to be known by everyone, such as “sugar tastes sweet,” or “if someone
hits you, you may feel pain.” The second is application-level commonsense, that is,
knowledge specific to a domain, and considered to be commonsense in that domain.
An example of application-level commonsense in our web search engine domain is:
“espn.com is a website which provides news about sports.” Both types of common-
sense can be easily solicited through the Open Mind website interface because each
piece of knowledge is expressed in simple English. In addition, some application-
level commonsense can be mined from the World Wide Web.
3 User Interface
Fig. 1. A screenshot of the current User Interface for GOOSE, where search goals must be
manually disambiguated.
Arguably the most intuitive interface would simply allow the user to type the entire
search goal as a sentence in natural language. Our system must then understand that
goal and have the expertise to know how to reformulate the goal into a good query.
Unfortunately, the expertise in our system is currently not complete enough to be able
to interpret arbitrary goals, so instead we have created some templates that encapsu-
late search engine expertise for the common categories of goals.
GOOSE’s user interface (Fig. 1) asks the user to select the goal of his/her search
from a pull-down menu, and enter a query that completes the sentence begun by the
search goal. Currently five search goals exist, and they are:
1. “I want help solving this problem:”
2. “I want to research…”
3. “I want to find websites about…”

4. “I want to find other people who…”
5. “I want specific information about the product/service…”
Because the knowledge associated with each goal category is modular, it would be
relatively easy to add new search goals. Without an extensive usability study, it is
unclear exactly how much coverage these categories of goals provide, and how many
may eventually be needed; however, we believe that some of the categories listed
above are generic enough to support any goal. Therefore, the issue of scaling up is not
likely to be limited so much by the number and types of available search goals as by
the diversity and coverage of the commonsense knowledge available for inference.
One limitation of this type of interface is the lack of ability to state multiple goals and
overlapping goals, but this is addressable if we allow more than one goal to be active
at a time, and devise a method for combining search results obtained through multiple
goals.
4 Mechanism
Given a search goal and search query, GOOSE performs four major internal steps
before results are returned to the user: 1) Parsing the query into a semantic frame [6];
2) classifying the query into a commonsense sub-domain; 3) reformulating the query
through commonsense inference guided by expertise templates; 4) and re-ranking
results using commonsense concepts.
4.1 Parsing to Frames
After the user executes the search query, GOOSE parses the query to fill the slots of a
semantic frame, which provides a concise, stereotyped representation of the original
query. Representing the original query as a frame makes commonsense reasoning
easier because the most important features of the query are extracted. An example is
given in Table 1.
Table 1. An example of a filled semantic frame for the goal, “I want help solving this prob-
lem:” and the query, “My cat is sick and wheezing”
Slot Name Slot Value

Problem Attribute [Sick, Wheezing]
Problem Object [Cat]
As suggested by this example, each search goal needs its own unique set of seman-
tic frames. This is true because different aspects of each query are useful to accom-
plishing different search goals. In the above example, identifying the problem attrib-
ute and problem object is most useful to identifying a solution through commonsense
reasoning. The set of all semantic frame templates represents a part of the expertise
that experienced users possess. Currently, each of the five search goal categories has
one semantic frame, but as the system scales, more frames are likely to be added.
It is worth pointing out that our system’s parsing of the natural language query dif-
fers from the ways in which other search engines handle unstructured text input. A
typical search engine throws out a list of stop words and treats the remaining words as
keywords [9], but our approach tokenizes, part-of-speech tags, parses the entire
query, and translates the parse tree into a filled semantic frame.
4.2 Classification
In addition to parsing the original query to frames, a classifier examines the original
search goal and determines the commonsense sub-domain that can provide the most
applicable knowledge when performing commonsense reasoning. Examples of sub-
domains for the “I want help solving this problem” search goal include “personal
health problems,” and “household problems.” Each sub-domain contains both ordi-
nary and application-level commonsense knowledge.
Classification is performed in a relatively straightforward way. Each sub-domain
is described by the commonsense concepts it covers. A search goal is classified into
all sub-domains, which match the concepts contained in it. Multiple sub-domain
matches can be safely merged.
We have chosen to group the commonsense used for GOOSE into sub-domains to
help disambiguate certain words and concepts. Another advantage of sub-domains of
commonsense is the savings in the run-time of the inference, a benefit of a smaller
search space.
4.3 Reformulation
In this step, we take the filled semantic frame and apply reasoning over the chosen
commonsense sub-domain. In our current implementation, reasoning takes place as
an inference chain, implemented as a depth-first search, guided by heuristically moti-
vated rules that help direct the inference so as to avoid unnecessary searching. Infer-
ence terminates when an application-level rule has fired. Again, application-level
rules are a component of search engine expertise. When the inference terminates, we
will have the reformulated search terms that we need.
Once the query has been successfully reformulated, it is submitted to a commercial
search engine and the result is captured for further refinement.
4.4 Re-ranking
Using GOOSE’s concept vectors, a list of weighted words and phrases representative
of the concepts contained within the search query, GOOSE re-ranks the search results
so that the hits most relevant to the search query are given higher priority.
This concept-based re-ranking step is similar to the query expansion approach pro-
posed by Klink [3], except that in our case, it is only used as a refinement of existing
search results.
Where commonsense inference fails to infer any query from the search goal, key-
words from the search goal are extracted and passed to the commercial search engine.
In such cases, query refinement with commonsense concept vectors can serve as a
back-up mechanism, because such refinement may still lead to improved results over
the baseline where GOOSE is not used at all.
5 A Scenario
Having explained the GOOSE user interface and mechanism, let us imagine a typical
user scenario. Suppose that a novice user has a sick pet and wants to find ways to
remedy the problem. She does not know how to form a good search query, so she
decides to try her search on GOOSE. She chooses the goal, “I want help solving this
problem:” and types in the query, “my golden retriever has a cough.” Using the se-
mantic frame defined for this particular goal, GOOSE fills the frame as follows:
Problem Attribute: [cough]
Problem Object: [golden retriever]
The classifier examines the query and determines that the commonsense sub-
domain to be used is “animals.” Performing inference over the “animals” sub-
domain, the following inference chain is invoked:
1. A golden retriever is a kind of dog.
2. A dog may be a kind of pet.
3. Something that coughs indicates it is sick.
4. Veterinarians can solve problems with pets that are sick.
5. Veterinarians are locally located.
The first three steps in the inference chain are ordinary commonsense, while the
last two steps are application-level commonsense. GOOSE takes the result of the
inference and submits the reformulated query, “Veterinarians, Cambridge, MA” to a
commercial search engine. The locale that was added is a personalization, and must
be obtained through a user profile. After search results are returned, commonsense
concept vectors are used to refine the results so that search hits containing the con-
cepts closest to “veterinarian” appear higher in the search results.
The user finds what she was looking for in the first page of results and never had to
explicitly choose the keywords that brought her to what she was looking for.
6 Preliminary User Tests
We conducted preliminary user tests asking four search engine novice users to form
queries for a few simple search tasks. Due to the limited search goal categories avail-
able in the current implementation, we focused on the categories and commonsense
sub-domains that the system knew how to handle. The query being inputted into the
GOOSE UI was sent to both GOOSE as well as directly to the Google commercial
search engine. Users were then asked to rate the relevance of the first page results on
a scale of 1 (most irrelevant) to 10 (most relevant). In cases where commonsense
inference failed to infer a search query from the search goal, commonsense concept
vectors were still used to reorder the search results. Table 2 presents some of the
results.
Table 2. Preliminary user test results. Participants formed 2 queries each for each search task.
Search Task # successful Avg. score Avg. score
inferences GOOSE Google
Solve household problem 7/8 6.1 3.5
Find someone online 4/8 4.0 3.6
Research a product 1/8 5.9 6.1
Learn more about 5/8 5.3 5.0
Our test results suggest that for novice search engine users, GOOSE on average
produced more effective first-page results than Google, a leading commercial search
engine. The problem solving goal category is where inference showed the most
promising results, as demonstrated by the search task “solve household problem.”
However, the high rate of failure of the inference in producing a query suggests that
GOOSE is still very brittle in the current implementation and only works well under
very constrained domains for which organized commonsense knowledge exists, such
as personal health and household problems. Many more domains and goals must be
supported before any extensive user tests can performed, but these initial results are
encouraging.
One fundamental limitation of using the commonsense knowledge for inference is
illustrated by the result for the “research a product” search task. In this task, some
users chose to search for trademark names of products, such as “the Total Gym”, and
“TurboTax”. In such cases, commonsense inference will not be of help because
trademark product names are not part of the knowledge base. This task, however,
seems to be particularly suitable to a keywords approach, as Google received rela-
tively high marks on this task. GOOSE received similar marks because although it
could not be helpful in this case, it did not hurt the results either.
The results of this preliminary user test are promising. In future user tests, we
hope to measure the intuitiveness of the proposed user interface, the usefulness of
GOOSE to already experienced users, and have a head-to-head comparison against
other query enrichment techniques such as keyword expansion and relevance feed-
back.
7 Conclusion
In this paper, we presented an adaptive search engine interface that can use common-
sense to perform inference over user’s search goal, in order to generate an effective
query. While the commonsense inference is not complete, it can still be useful.
GOOSE is a fail-soft application, in that, in the case it fails to produce a better query
for the user, it will just produce the same results the user would have obtained any-
way. So the argument can be made that GOOSE can be useful, even if the common-
sense reasoning is brittle, because GOOSE can help some of the time, and it won’t
hurt the rest of the time.
As we continue to scale the commonsense coverage of GOOSE, we face several
pointed issues. First, classification into commonsense sub-domains becomes less
accurate as the number of sub-domains increases. Second, it will become increas-
ingly difficult to define commonsense sub-domains that are of the right size and that
do not overlap with existing sub-domains. One radical solution to the two above-
mentioned problems would be to not only allow overlapping sub-domains, but to go
so far as fostering many diverse and competing representations of commonsense,
each with strengths and weaknesses, which will compete with each other in their
reasoning of a particular query. In this model as suggested by Minsky [5], common-
sense coverage will be increased, and reasoning will be more robust because it will
exploit the complementary strengths of different representations.
Third, as the number of commonsense statements increase, inference will take
combinatorially longer and be more prone to noise because the search space will have
increased. To overcome these problems, we need to give the inference process more
guidance via pruning techniques, and give it the ability to recognize when it is on the
wrong path to the goal, or when it does not know enough to reach the goal. One
possibility for guided inference is to valuate candidate inference chains that will result
from different inference paths, much like a chess-playing program valuates board
positions that result from different moves. However, this approach assumes that it is
feasible to devise good ways to valuate an inference chain, which is a non-trivial
problem.
From the preliminary user tests, we have learned more about the fundamental limi-
tations of using commonsense to help the user compose search queries. First, the
commonsense knowledge in Open Mind contains only about 400,000 facts, which do
not assume equal distribution of the knowledge over diverse topics. Minsky estimates
that somewhere on the order of 15 million pieces of knowledge may be needed in
order to be comparable to what humans possess. Obtaining and organizing knowl-
edge on that scale will be a huge challenge, not to mention efficiency issues that that
scale creates. The second major limitation of commonsense is that it will probably
not tell GOOSE about all the specific topics needed to perform inference, such as
trademarked products, what specific companies do, etc. It can only help to reason
about concepts and problems that we encounter in everyday life. Without speculation
of the difficulty of doing so, if we can mine specific knowledge from other resources
or the web, it may be possible to connect this knowledge to the inference mechanism.
8 Future Work
As of now, GOOSE is not yet robust or helpful enough although it has the potential to
be. In addition to the scaling issues discussed above, we are working toward two
goals: personalizing the commonsense, and automatic detection of goal categories.
One way to consider the role “common sense” plays in the system is to think of it
as a generic user model, because it represents knowledge in everyone’s head (every-
one within a particular culture). We can customize the user model by adding personal
commonsense to the system such as “Mary is my sister.” Personalizing commonsense
is logical, because the notion of what “common sense” is varies from one person to
another. GOOSE may be able to utilize personal commonsense to better interpret the
user’s search goals and produce more relevant results. An example of personalization
currently used by the system is the placement of the locale keyword in the query to
accompany a local business. “Veterinarian Cambridge, MA” is one example. How-
ever, we can also imagine subtler examples of how personal commonsense can influ-
ence inference. For instance, if a user has a broken VCR, she might want search re-
sults for either do-it-yourself resources or electronics repair shops. Depending on the
type of person she is, she might want one type of result or the other or both.
Personal commonsense can be stated as simple English sentences, so it is easy to
add to the system. The real challenge is in devising a way to collect the information.
For instance, we can imagine with the broken VCR example, that the user may be
shown two sets of search results, and her preferring one set or the other set may then
enter the appropriate piece of personal commonsense into the system. Other ways to
enter personal commonsense can include an interview wizard, mining information
about the user from a homepage, or getting shared information from some other con-
text-aware application that is also learning personalizations about the user. ARIA [4],
a photo agent also being developed at the MIT Media Lab, is an example of such
application.
The second goal is to eliminate the explicit goal selection task by automatically
classifying queries into goal categories. This may first require the coverage of the
goal categories to be validated, and may necessitate more robust natural language
processing in order to parse the unconstrained input. Alternatively, we may be able to
apply shallow IR techniques such as support vector machines to perform the classifi-
cation.
In the end, we hope to create a much more intuitive and personalized search ex-
perience for all web users, and to utilize the lessons learned here about commonsense
reasoning so as to be able to apply its benefits to other domains and applications.
References
1. Ask Jeeves, Inc..: Ask Jeeves home page. (2002). http://askjeeves.com/.

2. Belkin, N.J.: Intelligent information retrieval: Whose intelligence? In: ISI '96: Proceedings
of the Fifth International Symposium for Information Science. Konstanz: Universtaetsver-
lag Konstanz. (1996). 25-31.
3. Klink, S.: Query reformulation with collaborative concept-based expansion. Proceedings of

the First International Workshop on Web Document Analysis, Seattle, WA (2001).
4. Lieberman, H., Liu, H.: Adaptive Linking between Text and Photos Using Common Sense
Reasoning. In Proceedings of the 2nd International Conference on Adaptive Hypermedia
and Adaptive Web Based Systems, Malaga, Spain (2002).
5. Minsky, M.: Commonsense-Based Interfaces. Communications of the ACM. Vol. 43, No. 8
(August, 2000), Pages 66-73
6. Minsky. M.: A Framework for Representing Knowledge. MIT, (1974). Also, In: P.H.
Winston (Ed.): The Psychology of Computer Vision., McGraw-Hill, New York, (1975).
7. Peat, H. J. and Willett, P.: The limitations of term co-occurrence data for query expansion
in document retrieval systems. Journal of the ASIS, 42(5), (1991), 378--383.
8. Singh, P.: The Public Acquisition of Commonsense Knowledge. AAAI Spring Symposium,
Stanford University, Palo Alto, CA, (2002).
9. Shneiderman, B., Byrd, D., and Croft, B.: Sorting out searching: A user-interface frame-
work for text searches, Communications of the ACM 41, 4 (April 1998), 95-98.
10. Voorhees, E.: Query expansion usin lexical-semantic relations. In Proceedings of ACM
SIGIR Intl. Conf. on Research and Development in Information Retrieval. (1994) 61-69.
11. Xu, J. and Croft, W.B.: Query Expansion Using Local and Global Document Analysis. In
Proceedings of the Nineteenth Annual International ACM SIGIR Conference on Research
and Development in Information Retrieval, (1996). pp. 4-11.
On Adaptability of Web Sites
for Visually Handicapped People
Mercedes Macías, Julia González, and Fernando Sánchez
Escuela Politécnica. Universidad de Extremadura.

Avda. Universidad s/n, 10071, Cáceres, Spain
{mmaciasg, juliagon, fernando}@unex.es
Abstract. Currently, the great majority of content published on the Internet is

inaccessible to visually impaired users. Although designers have guidelines that
guarantee the accessibility of pages constructed as well as software tools to fa-
cilitate this task, it is necessary to consider the user's perspective too, allowing
him/her to participate in the restructuring or presentation process of contents.
There are few software tools which are able to do this. In this paper we present
KAI (Accessibility Kit for the Internet) that considers both the user and the de-
signer. It classifies the different components of a published Web page and pre-
sents them to the user according to his/her needs. KAI is based on a new lan-
guage, BML (Blind Markup Language) that helps authors to develop better
structured pages. It provides two levels of independence: original Web code
and user browser or navigation platform. KAI includes a mixed audio/touch
browser (WebTouch) that enables selective reading of contents. The proposed
accessibility kit uses several accessibility metrics to ensure that those pages
transformed by KAI are more accessible than the original ones. In this paper we
give an overview of the overall system.
1 Introduction
When the World Wide Web service was created and HTML became its main support,
only some people could guess what today constitutes one of the most valuable re-
searching or working instruments of present society. Some of the best qualities that
this service offers are availability and immediate diffusion of information published
on the Internet. These characteristics are supposed to be especially useful for users
with some types of disability.
However, the terms availability and accessibility are not synonyms. Anybody can
obtain the information published on the net but not everyone has the ability to obtain
it in the same way or in the same quantity or quality. Most of the time, users with
disabilities must overcome an endless number of accessibility obstacles in order to
reach the same level as users without disabilities. In the case of visually handicapped
people, a conventional computer cannot be used in a conventional way. First of all,
these users need technical assistance to use hardware (Braille lines, Braille printers)
or software (screen magnifiers, screen readers). Second, the majority of browsers

On Adaptability of Web Sites for Visually Handicapped People 265
used to surf the net use a graphical approach that hinders the work of the screen read-
ers mentioned above. The few specific text-only or voice browsers which exist do not
contain the information provided by the structure of the original page. Finally, the
Web page itself may be badly organised or incompletely or incorrectly coded in
HTML, making it impossible for the user to access to its content.
Conscious of these problems, several companies and associations are working on
guidelines and recommendations for software and hardware engineers, Web browser
and authoring tool developers and Web page designers. Fortunately, governments are
also trying to minimise these difficulties in public websites.
The accessibility research group of the University of Extremadura is developing a
project for people with visual handicaps. It improves the accessibility of information
published on the net. The aim of this paper is to present an overview of this project.
The rest of the article is as follows: in Sect. 2 related works are considered. In
Sect. 3, the main ideas of our proposal are outlined. Finally, in Sect. 4 we summarize
and present ideas for future works.
2 Motivation
When one surfs the Internet with certain regularity and visits different Web pages,
one realises that there are almost as many ways to present the information, as there
are Web page designers. With such variety, authors have something in common: they
try to structure the Web page content in such a way that it invites Web surfers to read
it.
Structuring different contents of a Web page goes far beyond the mere aesthetic
function. Structures help users to perceive the information better, understand more
quickly the ideas behind them, and make conclusions. Many times, the designer un-
consciously chooses a determinate structure for a certain type of content. For in-
stance, a group of statistical data is usually assigned to a table, since this structure
better enables its comprehension. Despite the importance of the Web page content’s
structure, very few books about Web design and publication on the net refer to this
aspect. However, all of them explain how to construct different structures from a
technical point of view, generally in HTML.
As a reference of the variety of Web pages and designs, the Web Accessibility Ini-
tiative (WAI) of the World Wide Web Consortium (W3C) has compiled some Guide-
lines of Accessibility for Web page authors [1], authoring tools creators [2] and Web
browser designers [3]. Their final objective is to make Web pages more accessible.
These guides constitute excellent resources that every author should be familiar with,
although it is impossible to force anybody to follow a guideline when designing a
Web page. This is the reason why WAI has also created an evaluation and repairing
working group, which is in charge of compiling information about software tools that
evaluate, repair or transform Web pages [4]. This helps not only Web designers, but
also users, to build a more accessible Web page.
These tools can be classified into two main groups. The first one contains tools
oriented to Web designers. Among this group of tools, there are:
266 M. Macías, J. González, and F. Sánchez
• Evaluation tools: They usually perform a static analysis of a Web page by evaluat-
ing its accessibility and returning a report. One example is Bobby [5], which guar-
antees that a page complies with all the standards of accessibility recommended by
the section WAI of W3C.
• Repairing tools: Once the errors of accessibility on a Web page have been identi-
fied, repairing tools help the designer to correct them. One example is Tidy [6]. It
detects and cleans the syntactic errors found in HTML and XML files, warns of
accessibility problems and replaces all the presentation tags with style rules in
CSS.
The second group contains tools oriented to the users, allowing their participation
in the final presentation style of a Web page. Users are able to enhance their browsers
by personalising certain characteristics of presentations. Some of these tools are inte-
grated in the user’s browser even though the majority work on the server. Among this
group of tools there are:
• Filtering tools: their task consists in eliminating certain elements that could be
inaccessible while giving the user the possibility of deciding what these elements
are. One example is Muffin [7], which offers the user various configurable filters
to eliminate cookies, GIF animations, advertisements, Java applications,
JavaScript, and add/eliminate/modify some HTML tags.
• Transformation tools: it is not enough to eliminate the elements that seem to be
inaccessible at first glance. Occasionally, some of them can be recovered if they
are transformed correctly [8,9,10].
The tool proposed in this article classifies, evaluates, filters, repairs, restructures,
transforms and presents the content of a Web page to the user. Moreover, it does all
this in a personalised way. It is a software tool oriented to both, the user and the de-
signer. In next section an overview of KAI is given.
3 KAI
3.1 A General Overview
KAI is a personalised environment for surfing the net. Users with a visual handicap
decide how to access the content of a Web page. They can choose both the contents
they are interested in, and the way they are presented. KAI classifies, filters and re-
structures the contents of any existing Web page in order to eventually present them
to the user as he/she desires.
Beforehand, it is necessary to identify all the components of a given page. Nowa-
days, HTML hinders this identification since it is a permissive language. In other
words, it allows one to obtain the same visual presentation of one component of a
page though several language elements. For example, the component table should be
constructed with the HTML element <TABLE>, but the designer could also scan the
table and place it in the Web page with the HTML element <IMG>, so the table
would be treated as an image. Designers could also use certain combinations of

ASCII characters to obtain a similar effect, but only visually. And in many authoring
tools, the element <TABLE> is used to distribute the different contents in a certain
visual way, but not to build a real table. A screen reader is able to read the content of
an element <TABLE> but it can not read the content of an image. And the way in
which it reads a sequence of ASCII characters is no longer valid.
Faced with these difficulties, we have developed a new language to construct ac-
cessible Web pages, called BML, Blind Markup Language, upon which KAI relies.
KAI also offers an accessible authoring tool to simplify the process of making acces-
sible Web pages. The language BML has been developed following the specifications
of the meta-language XML. It is very simple. Its fundamental function is to code the
real components of a Web page, like the table mentioned above, but not those HTML
elements used to achieve them.
The creation of an audio/touch platform (WebTouch), which can represent the
structure of a Web page, has also been proposed. This platform permits users with
visual handicaps to access the contents selectively, rather than sequentially.
Figure 1 shows the outline of the architecture supporting KAI.
Accesibility ORIGINAL
metrics measure W eb
Web browser page
Synthesizer
IMPROVED BML Detection &
Touch panel W eb Transformation
classification
page
Authoring tool
Fig. 1. A general overview of KAI
Traditionally, a client accesses a Web page through a connection with a server.

The server sends to the user the code of the page he/she wishes to visit but the user’s
browser is responsible for the interpretation of this code and its final presentation. We
suggest that the browser could transform the original code, improving it as an added
service to their users.
Our transforming tool performs several tasks:
• It transforms the source code of the web page from HTML or XML to the inter-
mediate language BML. This transformation is achieved by classifying the content
and the structures of the page. In order to do this, techniques of pattern recognition
and source code analysis are used. These techniques are necessary to correctly
identify the real components of a web page, such as the table mentioned above,
which can not be coded as a table but as an image.
• Once the contents and structures are classified and coded in BML, they can be
filtered, repaired, restructured and presented according to the user’s preferences or
needs which have been previously identified.
• If the user's browser does not support XML, KAI transforms the BML code to
HTML again. Then the Web page is ready to be sent. We would like to emphasise
that the final visual aspect of the Web page could be identical to the original. The
difference is that the components are now conveniently tagged. For example, a
structure table would be marked correctly, with an element <TABLE>, so that the
screen reader could read it in the appropriate way. Of course, the final browser
may not be a traditional one, for example WebTouch. Then, BML is translated to
the appropriate structures supported by this platform.
During this transformation process, different accessibility metrics based on the

standard ISO9126-1 [11] and the model discussed in [12,13] are applied. These met-
rics allow us to establish a ranking of web pages according to their accessibility. This
feature is very useful for search engines. After transforming the original page into an
improved one, the user can compare the accessibility level between the original page
and the new one. Finally, when the page is presented in one specific browser, the
quality in use can be measured to know if the result is satisfactory for the end-user.
Next we give several details of BML, WebTouch and the metrics used.
3.2 The BML Language
BML has been developed following the XML specifications. It has been designed to
build Web pages independently from the presentation device and the technology
used. Besides, isolating the contents of a page from the aspects of presentation per-
mits the extraction of the structures that support them. This separation of aspects:
contents, presentation and structures, is fundamental in order to achieve easily trans-
formable pages. Thus, all the users obtain the same information, independently from
the senses they use: sight, hearing and/or touch.
This language is very simple and its simplicity plays an important role in the cor-
rect coding of the different components of an existing Web page. Besides, BML gives
KAI total independence from the original language of the Web page (HTML or
other), and from the language used when sending a page to the user (HTML,
XHTML, XML,…, with CSS or XSL).
In addition, this language offers:
• An authoring tool that guarantees that the Web page constructed is accessible.
Simply, it only constructs accessible Web pages. At first, Web designers con-
structed their pages in HTML language directly, and for this reason, there were
very few designers. Now, there are many authoring tools to help designers con-
struct sophisticated Web pages. Although the majority of tools build pages based
on HTML language, the designer can use them without knowing this language.
Anybody can be a designer, if he/she is able to see. Generally, once the page is fin-
ished and the author wants to verify whether it is accessible or not, he/she must use
another specific tool. Many elements might result inaccessible and the majority of
authoring tools can not prevent this situation. In general, the designer has to repair
the inaccessible elements in HTML by hand, thus making it difficult for many de-
signers.
• An authoring tool which is accessible itself. This means that anybody can be a
Web designer. Even a visually handicapped person will be able to design a Web
page in a very attractive way.
• BML serves as a reference to restructure any Web page on the net, increasing its
degree of accessibility.
• The possibility of configuring specific style sheets for different output devices,
technologies or versions. Therefore, users are able to obtain the contents of any
Web page based on their previous preferences, which can be changed dynamically.
3.2.1 The Elements of BML.

The definition of BML includes the existence of a group of components that consti-
tute the basis of any Web page. Two criteria have been followed to consider a possi-
ble element of BML:
• Identifiable: a component is identifiable when any user, regardless of the presenta-
tion device used, is able to recognise it. The visual (images, bright signs) and audi-
tory perspectives have both been kept in mind. Sight and hearing are senses that
are regularly used in the perception of Web pages. Nevertheless, BML language
considers an element identifiable from haptic or tactile points of view too (Braille,
raised symbols, movements, or vibrations).
• Functionality: a well-defined function must also be considered. This means that
each component can provide additional information enhancing its own content. A
component has a well-defined functionality when all users are able to perceive the
same function.
For example, a link would be a component of a Web page. It is identifiable by any
user both in a visual and auditory way. Its function is to connect the current page with
another point on the network. Therefore, the component link constitutes a clear ele-
ment of the language.
Each component plays a role within a Web page. The elements of the language can
be classified according to this role:
• Constructive elements: when combined correctly, they can compose any Web
page. There are two different types of constructive elements:
o Elemental constructors: these elements represent the simple components within
a Web page. They are not the result of any combination. They include: Text,
Image, Sound, Email Address, Link, Button, Control or Touch. For example, the
last one collects components from the Web page that provides information of
interest to the user from the tactile point of view. Although this tactile aspect is
not yet commonly used in Web pages, some examples can be seen in [14].
o Structural constructors: they combine elemental constructors and add some
complementary information regarding such association. The following elements
are included in this group: List, Table, Form, Sub-window or Group.
• Organizational elements: this type of element helps the user to surf a Web page by
orienting him/her. For example, the element Abstract of a Web page provides in-
formation such as the summary, the authors, the languages available, the data of
creation or publication, etc. Another organizational element is Directory, which
guides the user through the different contents of the Web page in order to facilitate
surfing.
• Semantic elements: in this case, they are embedded in constructive elements, giv-
ing them added meaning and functionality, such as Language, Date or Telephone.
This last element allows the user to navigate among the different telephone num-
bers in a Web page, for example.
Besides its role, each component has a different relationship with the user. With
respect to this connection, different structures or elements of the BML language could
be classified in the following way:
• Interactive structures: These structures permit the user to actively participate in the
information process. For example, Email would be a clear interactive element,
since it allows the user to activate it and contact its owner.
• Navigational structures: They allow the user to surf through the interior or exterior
of a page that is being visited. For example, the element Directory offers a table of
contents to guide the user.
• Information structures. In this case, structures give the users the content they are
seeking. For example, Text uses to be the main container of information.
• Meta-information structures. They offer additional information about the content
that is being visited. For example, Abstract gives the user additional information
about the languages available or the date of creation of the Web page.
Table 1 classifies some of the elements of the BML language according to the two
criteria explained above: the role carried out in the interior of a Web page and its
relation with the user.
Table 1. Classification of some elements of the BML language
ELEMENTS
CONSTRUCTIVES
Elemental Structural ORGANIZAT. SEMANTIC
INFORMATION Text Form
Image Sub-window
Sound
Touch
META- Group Section Language
INFORMATION List Extract Date
Table
INTERACTION Button Phone
Control
E-mail address
SURFIN Link Directory
3.3 The Audio/Touch Browser
KAI provides an audio/touch browser (WebTouch) in which the structure of a Web

page is represented. This browser includes several buttons so users can interact with
the different elements of a web page. Thus, a blind user can obtain information and
surf the net just like other users who are not visually impaired. Visually-handicapped
users can access the contents of a Web page more selectively, thus complementing
their usual tools that already offer certain ability to select information like the voice
browser HomePage Reader [15] or the popular screen reader Jaws [16].
Nowadays, we rely on both graphic and auditory platform simulators. A touch
simulator is being developed with the help of a touch mouse, which produces differ-
ent textures when it passes over distinct representations of Web page’s components.
Consequently, the user is provided with general information about the content’s struc-
ture. The first version of WebTouch is being finishing at the time of writing this pa-
per.
3.4 Accessibility Metrics Measure
One of the main goals of KAI is to ensure that the final page is better than the original
one in terms of accessibility. In ISO 9126, the quality of a product is high-levelled
defined into six main characteristics: functionality, reliability, usability, efficiency,
maintainability, and portability. In our case, we are interested in functionality and
usability. Moreover, we introduce the concept of quality in use that can be defined as
the capability of the software to enable specified users to achieve specified goals with
effectiveness, productivity, safety and satisfaction in specified context of use. We
focus on the issue of choosing metrics to measure these characteristics, especially in
the case of visually impaired users.
For example, and regarding to images, several measures to take into account when
fixing the accessibility level of a given page are:
• Image count: it helps to measure the amount of provided visual information.
• Different image count: counts the non-repeated images.
• Percentage of image redundancy: the relation between the amount of different
images and the image count. An image repetition may be interpreted as the level of
redundancy of visual information.
 DifferentImageCount  (1)
PercentageImageRedundancy = 100 * 1 − 
 ImageCount 
But each image can have associated an alternative text, which is used to give extra
information or simply when it is not possible to display or see the image. We can also
define several measures regarding the alternative text:
• Number of images with alternative text.
• Percentage of presence of alternative text. This metric is calculated as follows:
#ImageswithAlternativeText (2)
PercentagePresenceAlternativeText = 100*
ImageCount
• Number of images with descriptive alternative text.

• Percentage of presence of descriptive alternative text.
• Percentage of descriptive alternative text.
Of course, there are other measures that are taken into account but it is impossible
to mention all of them here.
Throughout this paper, we have presented several tools that help to solve problems
that visually handicapped users experience when accessing information published on
the Internet. Some of these problems can be analysed from the Web designer’s posi-
tion and others from the user’s perspective. Then, KAI has been presented, a proposal
that face the problem from both sides.
Designers can use the language BML and its accessible authoring tool to create ac-
cessible Web pages. At the same time, users can personalise the environment for
surfing by deciding what information they want and how they want to receive it.
Thus, KAI classifies the contents of a Web page in relation to the elements of the
language BML and the user’s preferences, enabling him/her to increase the degree of
accessibility of any Web page. The audio/touch browser (WebTouch) facilitates Web
surfing and the user’s ability to perceive the content’s global aspect and its distribu-
tion within a Web page.
We are currently developing personalised metrics that permit us to determine the
degree of accessibility and the degree of transformability that a Web page has, in
relation to the needs of a user.
The metrics could be integrated in a search engine, so that besides the data that
these tools already offer, searching could be filtered in relation to a user’s predefined
threshold of accessibility or transformability.
References
1. Chisholm, W., Vanderheiden, G. and Jacobs, I: Web Content Accessibility

Guidelines 1.0., W3C Recommendation,
http://www.w3.org/TR/WAI_WEBCONTENT/, 5/5/1999.
2. Treviranus, J., McCathieNevile, C., Jacobs, I., Richards, J.: Authoring Tool Ac-
cessibility Guidelines 1.0., W3C Recommendation,
http://www.w3.org/TR/ATAG10/, 3/2/2000.
3. Jacobs, I., Gunderson, J., Hansen,E.: User Agent Accessibility Guidelines 1.0.,
W3C Working Draft, http://www.w3.org/TR/UAAG10/, 22/6/2001.
4. Web Accessibility Initiative. Evaluation, Repair, and Transformation Tools for
Web Content Accessibility. W3C. http://www.w3.org/WAI/ER/existingtools.html
5. CAST, Welcome to Bobby 3.2., CAST, http://www.cast.org/bobby
6. Raggett, D.: Clean up your Web pages with HTML TIDY, W3C,
http://www.w3.org/People/Raggett/tidy, 4/8/2000.
7. Muffin, World Wide Web Filtering System 0.9., http://muffin.doit.org/
8. Vorburger, M.: ALTifier Web Accessibility Enhancement Tool,
http://www.vorburger.ch/projects/alt, 1998-1999.
9. Gunderson, J.: Web Accessible. http://slappy.cs.uiuc.edu/fall99/team5, 1999-00.
10. Cascado et al: Web adaptor for the visually handicapped, Grupo de robótica y
Tecnología de computadores aplicado a la Rehabilitación. Facultad de
Informática. Universidad de Sevilla. Proceedings of Iberdiscap2000.
11. ISO/IEC 9126-1: 2001 (E), International Standard Software Engineering - Prod-
uct Quality- Part 1: Quality model.
12. Olsina, L.: Web Engineering: A Quantitative Methodology for Quality Evalua-
tion and Comparison of Web Applications, Doctoral Thesis (in Spanish), Cien-
cias Exactas School, UNLP, La Plata, Argentina. 2000, April.
13. Olsina L., Gonzalez-Rodriguez, J., Lafuente G., Pastor, O.: Toward Automated
Web Metrics. Proceedings of WQS2001, Anais do VIII Workshop de Qualidade
de Software, October, RJ, BR, pp.74-86.
14. Logitech, IFeel Mouseman, http://www.logitech.com.
15. IBM, IBM Home Page Reader. http://www-3.ibm.com/able/hpr.html.
16. Henter Joyce, Jaws for Windows, http://www.hj.com/JAWS/JAWS.html.
A Framework
for Filtering and Packaging Hypermedia Documents
Lucimar C. Martins, Tatiana A.S. Coelho, Simone D.J. Barbosa,

Marco A. Casanova, and Carlos J.P. de Lucena
Informatics Department, Pontifical Catholic University of Rio de Janeiro

Marquês de São Vicente, 225 – 22453-900 – Rio de Janeiro – Brazil
{lucimar, tati, sim, casanova, lucena}@inf.puc-rio.br
Abstract. Despite great effort in attempting to develop systems that personalize

both content and presentation, there are still some important challenges related
to information filtering, packaging and formatting that adapt to user’s goals,
interests and presentation preferences. This paper addresses these issues by
proposing a three-level framework that achieves a high degree of separation of
concerns. The framework dissociates the packaging process from the filtering
and formatting processes, and thus facilitates the implementation, user testing
and fine-tuning of the system representations and algorithms.
1 Introduction
The feeling of being “lost in hyperspace” is familiar to most Web users. The
excessive amount of information confuses us to the extent of not knowing where we
are and forgetting what we were looking for when we started browsing.
Many applications related to e-learning, e-commerce and information retrieval
have been designed so as to adapt both content and navigation access to the users’
supposed knowledge [7,8], to their preferences and goals [1,2,3,9,11,13], to their
tasks and receptivity [4,5,15]. Many of these references illustrate the increase in
efficiency that certain adaptation techniques bring about [12,17].
Despite great effort in attempting to develop systems that personalize both content
and presentation, there are still some important challenges, some of which are related
to:
− acquisition and representation of relevant document and user information;
− information filtering that takes into consideration such information;
− information packaging and formatting that adapt to user’s presentation preferences,
his current browsing device and network conditions.
In this context, the primary contribution of this paper is to propose a three-level
framework that achieves a high degree of separation of concerns, by dissociating the
packaging process from the filtering and formatting processes. We also illustrate a

A Framework for Filtering and Packaging Hypermedia Documents 275
possible use of our framework through an instantiation called MyNews, which is a

personalized electronic newspaper.
A high degree of flexibility, provided by the framework, is necessary because the
success of adaptive systems relies heavily on the choice of representations and
algorithms adequate to the underlying domain and application. The evaluation of this
success must be done empirically, by means of user testing, which may require
further fine-tuning. The final goal is to achieve a high level of user acceptance and
satisfaction, with minimal redesigning and code rewriting.
This paper is organized as follows. Section 2 presents an overview of the proposed
framework. Section 3 describes the user, document and packaging models. Section 4
discusses the process of information filtering. Section 5 describes information
packaging. Finally, Section 6 presents the conclusions and suggests directions for
future research.
2 Framework
The idea of separation of concerns [18] provides a major motivation for the definition
of a framework for personalized filtering and packaging of hypermedia documents.
Our goal is be able to analyze issues related to document and user modeling,
information filtering and information packaging, as independently as possible.
This approach makes it easier to build a variety of applications in diverse domains,
by deriving multiple instantiations of the framework using different algorithms and
configurations. This characteristic is essential to adaptive applications, since user
satisfaction can only be empirically verified, and a great amount of tweaking and
fine-tuning may thus be necessary, according to the successes or limitations of each
configuration.
In order to illustrate the proposed framework and personalization processes, we
will use an instantiation of the framework, called MyNews, for the domain of
electronic newspapers. MyNews creates a personalized newspaper that selects
adequate content and presentation structures to be delivered to its users.
In the following subsections, we will present the architecture of the proposed
framework and a brief description of its hotspots.
2.1 Architecture
The overall goal of the architecture presented in this section is to outline the
adaptation process the available content goes through, from being requested by the
user to finally being presented to him. This process takes into account document,
user, packaging and formatting models.
Figure 1 presents a functional architecture of a document personalization system.
A user starts by requesting a personalized view of the set of documents contained in
the repository. This request is processed by the Interface Subsystem and dispatched to
the Filtering Subsystem, which is responsible for creating an ordered set of
276 L.C. Martins et al.
documents, based upon the user and document models. These models will be
discussed in Sect. 3.1 and 3.2, respectively.
The Packaging Subsystem is responsible for regrouping, reordering and
restructuring the filtered set of documents, based on the packaging model described in
Sect. 3.3.
The Formatting Subsystem creates the final layout and visual design of each
document, according to the formatting model. This task relies heavily on the expertise
of graphics designers and will not be further discussed in this paper.
Document
Interface Model
User
Document Filtering Model
Repository
Packaging
Packaging Model
Formatting
Formatting Model
Interface
Fig. 1. Functional architecture.
2.2 Hotspots
Our framework contains several hotspots to achieve a high degree of separation of

concerns. Whereas the basic architecture is invariant, the models and algorithms may
differ from one domain or application to another.
The framework hotspots are as follows:
− document model: defines the documents’ structure, including metadata that help
classify the documents in various ways.
− user model: allows the definition of constraints related to each metadata, i.e., acts
as a definition of views upon the document model.
− filtering algorithm: constructs a query by combining the document and user
models, and returns the ordered set of documents computed as most relevant to the
user.
− packaging model: defines which portions of the document’s structure should be
presented, and in which arrangements or presentation structures.
− packaging algorithm: uses the packaging model to regroup and rearrange the
filtered set of documents, as well as to restructure the documents.
− formatting model: defines how each document should actually look like when
presented to the user, including platform dependencies.
− formatting algorithm: uses the formatting model to create the final presentation
layout of the documents.
3 Models
3.1 Document Model
The document model is any hypermedia model that defines the documents’ structure,
including metadata attributes that reflect characteristics of the documents. The
definition of each metadata domain must include comparison and metrics operators
that permit defining precise filtering algorithms.
In the context of MyNews, documents are structured into: title, subtitle, authors,
summary, image and whole text. The metadata domains are:
− the semantic metadata domain is represented by a labeled directed graph, in which
a node label represents a theme in the domain, and an arc label represents the
strength of the connection between two nodes. The distance between two nodes A
and B is the weighed sum of the arcs in the shortest path between A and B.
Figure 2 presents a sample semantic graph;
− the importance metadata domain is the integer interval [1,10] and indicates the
importance of a document, where 1 means the highest importance. This metadata
determines the relation between a document and its corresponding theme;
− the temporal metadata domain is a set of timestamps and allows the computation
of a document’s obsolescence.
A B
1
1 1
1 D E F
C 2
1 1
G 1 H
Fig. 2. Part of the semantic metadata domain with the themes of an electronic newspaper.
3.2 User Model
The user model must contain information about the preferences and goals of a user
when using the system. It depends on the document model. The user model may be
provided directly by the user through a questionnaire, or implicitly inferred by the
system, which then needs to monitor his interaction behavior.
In the MyNews example, the user model consists of a triple (S,I,T), where:
− S is a list of themes that interest the user, taken from the themes listed in the
semantic metadata domain;
− I indicates the minimum importance of the documents to be retrieved.
− T is a set of constraints on the obsolescence of the documents.
Suppose that the user is interested in theme A and that he wishes to view all news
pieces (any importance value between 1 and 10) that were inserted in the last 3 days.
The user model could then be represented by the triple ({A},10,{t <3}).
3.3 Packaging Model
The packaging model specifies:

− the groups into which the filtered set of documents will be partitioned;
− the order of the groups and the order of the documents within each group;
− the document views, which define the components that will be retained and passed
to the formatting process, in addition to references or links to other document
views.
A document view specifies a different structure for a document, that may reorder or
even omit its components.
In MyNews, the document views are represented as in Fig. 3.
dv1 = {title, subtitle, link}

dv2 = {title, subtitle, summary, link}
dv3 = {title, subtitle, summary, image, link}
dv4 = {title, subtitle, image, whole text}
Fig. 3. Document views used in MyNews.
In the example, we define a very simplistic packaging model as follows. The

model contains the following groups:
− the front page contains news pieces with rank less than 6 (see Sect. 4 for the
definition of rank);
− for each theme T in the user model, there is group, called thematic group T,
containing all filtered news pieces whose theme is T.
The front page is always the first group and the thematic groups are listed by
alphabetical order of their theme. The news pieces within each group are listed by
rank order, as computed by the filtering algorithm.
For the first three documents in the front page, the packaging model retains the
components defined in view dv3 of Fig. 3; for all other documents in the front page, it
uses view dv1. The thematic groups use view dv4 for their documents. In addition,
for the documents in the front page, the hyperlinks in views dv1 and dv3 point to the
corresponding news pieces in the thematic groups.
An alternative packaging model for MyNews would, for example, use only view
dv1 for all documents and hyperlink each view to the document’s full version in the
document repository. A second and more sophisticated model would take into
account limitations of the platform and indicate to use views dv1 and dv2, instead of
views dv3 and dv4, respectively, if the platform has a small presentation area.
4 Filtering
The filtering algorithm creates an ordered subset of the set of available documents,
based on a ranking function that maps each document into its rank. according to their
metadata values and the goals and interests represented in the user model. Intuitively,
the filtering algorithm correlates the documents’ metadata with the user model,
creating an ordered subset of documents that supposedly interest the user the most.
The definition of the ranking function is an interesting issue. In order to verify that
a certain document is indeed relevant to the user, a great deal of user testing must be
done, and the ranking function will probably need to be fine-tuned for best results.
Returning to our running example, recall from Sect. 3 that the user model is the
triple ({A},10,{t <3}). Consider a ranking function that combines the semantic and
importance metadata and is defined as P[U ]( x) = ( D s [U ]( x)) 2 + ( D I [U ]( x)) 2 , where
Ds [U ]( x) corresponds to the shortest distance between the theme labeling document
x and the themes of interest to the user, and the value D I [U ]( x) represents the
importance of document x.
Table 1 shows a possible result of the filtering process, where the rank column
contains the document rank,computed by the above function, and all other columns
correspond to the document metadata and document components.
Table 1. A sample set of documents resulting from the filtering process.

Document rank theme importance obsolescence title
Doc2 2.24 C 1 3-Dec-2001 Wall Street...
Doc1 3.16 A 3 5-Dec-2001 Dolly...
Doc4 3.16 A 3 3-Dec-2001 Crisis in...
Doc53 5.10 A 5 3-Dec-2001 Research...
Doc18 7.21 D 6 4-Dec-2001 Christmas sales...
Doc7 10.05 A 10 1-Dec-2001 New clones...
5 Packaging
Packaging comprises three major processes, grouping, ordering and document

restructuring, and is driven by the packaging model.
The first step partitions the set of documents returned by the filtering process into
several groups, according to the criteria defined by the packaging model. The second
step reorders the groups, and the documents within each group. It may directly follow
the documents’ rank order or it may be recomputed as a function of both the rank and
some of the documents’ attributes. For instance, the documents may be presented in
reverse chronological order, indicating the recency of events, independently of their
relative importance. The final step selects which components of each document will
be passed to the formatting process.
Thus, the packaging process results in a sequence of tuples containing: a group

descriptor, a sequence of documents, the corresponding document view and the
“target” document views, in case the selected view contains links.
In the MyNews example, using the packaging model of Sect. 3.3, the documents
shown in Table 1 will be rearranged as follows:
− front page: contains documents doc2, doc1, doc4 and doc53, in this order;
− Thematic Group A: contains document doc1, doc4, doc53 and doc7, in this order;
− Thematic Group C: contains document doc2;
− Thematic Group D: contains document doc18.
The documents in the front page retain the following components:
− the first three documents - docs2, doc1 and doc4 – follow view dv3 in Fig. 3, that
is, they retain the title, subtitle, summary and image, if any;
− the last document, d53, follows view dv1, that is, it retains the title and subtitle.
The documents in Thematic Groups A, C and D follow view dv4, that is, they retain
the title, subtitle and the whole text.
This result may be represented as the following sequence:

[
(front pagea, [doc2, doc1, doc4], dv3, dv4),
(front pageb, [doc53], dv1, dv4),
(thematic group A, [doc7], dv4),
(thematic group D, [doc18], dv4)
]
Figure 4 schematically shows the final result of the packaging process, including
the hyperlinks between the filtered documents. Note, however, that the document
views used in Fig. 4 do not imply any formatting scheme. They are used just for
illustrative purposes.
6 Conclusions
In this paper we have defined a framework that helps design personalization systems.
Our major contribution was to organize the framework into three levels that achieve a
high degree of separation of concerns, by dissociating the packaging process from the
filtering and formatting processes.
The framework facilitates the implementation, user testing and fine-tuning of the
system representations and algorithms. The filtering process, for example, has three
hotspots - the filtering algorithm, the document model and the user model – that can
be instantiated in different ways. Likewise, the packaging process has two hotspots –
the packaging model and the packaging algorithm.
FRONT PAGE
Doc2 Doc1 Doc4 Doc53

Wall Street Dolly Crisis... Research
IIIzzz
IIIz New...
Down... Aging... Argentina
IIIxxxI
IIIIIII mmmmm ddddddd cccccccc
Thematic Group A
Doc1 Doc4 Doc53 Doc7
Dolly Crisis Research New Clon
Aging Argentina New... Other ...

dddddddd ccccccccc rrrrrrrrrrrrr nnnnnnnn
ddddd nnnnnnn
Thematic Group C Thematic Group D
Doc2 Doc18
Wall Str Christmas
Down Sales up
wwwwww ccccccccc
ccccccc
Fig. 4. Packaging of the documents in Table 1.
If we consider the filtering process, our approach is similar to other approaches in

the literature. It aims to select the documents (topics and news) that are most relevant
to each user.
In our framework, we introduce the packaging level, which allows us to define the
abstract organization of the selected information, by selecting and organizing relevant
elements of the filtered documents.
Our framework may be used as a resource for analyzing existing approaches to
personalized systems. Ossenbruggen’s work on Cuypers [16] decomposes what we
call formatting level according to their algorithms for automatic generation of the
final multimedia presentation. If we can design algorithms to generate an adequate
semantic structure based on communicative intentions, our packaging process would
be able to provide the input for their prototype. Our packaging process would be thus
positioned before the Cuypers Engine.
Our work is somewhat similar to the IMMPS proposal [6], which defines a
Standard Reference Model for Intelligent Multimedia Presentation Systems. Their
design layer, however, encapsulates our packaging level and, to some extent, our
formatting level as well.
The approach described in [3] proposes an electronic newspaper that allows the
personalization of the content and presentation detail of the news item based on
receptivity (a dimension used for estimating the amount of information that a user
might read). In place of our packaging process, their approach provides only two
different types of pages for organizing information: index pages and news pages. Our
instantiation, MyNews, maintains the distinction between the packaging and
formatting processes, allowing editors and designers to experiment more easily with
alternative solutions.
We are currently implementing the framework described in Sect. 2, with the
MyNews instantiation [15]. We plan to investigate an alternative framework where the
packaging process drives filtering. Also, we plan to extend the framework to address
the problem of processing specific user requests, such as keyword searches.
Acknowledgments
Lucimar Martins would like to thank CAPES for supporting her work. Tatiana
Coelho and Simone Barbosa thank CNPq for supporting their research.
References
1. Ardissono, L., and A. Goy. Tailoring the Interaction with Users in Electronic Shops. In J.
Kay, ed.: UM99 User Modeling: Proceedings of the Seventh International Conference.
Wien New York: Springer-Verlag, 35-44, 1999.
2. Ardissono, L., Console, L., and I. Torre. On the application of personalization techniques
to news servers on the WWW. In: Lecture Notes in Artificial Intelligence N. 1792. Berlin:
Springer Verlag, pp. 261—272, 1999.
3. Ardissono, L., Console, L., and I. Torre. Strategies for personalizing the access to news
servers. Working Notes of the Adaptive User Interfaces. Spring Symposium of AAAI
(Technical Report SS-00-01), pp. 7-12, Stanford, CA, 2000, AAAI Press.
4. Billsus, D. and M. Pazzani. User Modeling for Adaptive News Access. User Modeling and
User-Adapted Interaction 10(2-3), 147-180, 2000.
5. Billsus, D., Pazzani, J., and J. Chen. A Learning Agent for Wireless News Access.
Proceedings of the 2000 International Conference on Intelligent User Interfaces , 2000,
Pages 33 - 36.
6. Bordegoni, M., Faconti, G., Maybury, M.T., Rist, T., Ruggieri, S., Trahanias, P., and M.
Wilson. A Standard Reference Model for Intelligent Multimedia Presentation Systems.
Computer Standards & Interfaces, 18(6-7):477-496, December, 1997.
7. Bradley, K., Rafter, R. and B. Smyth. Case-Based User Profiling for Content
Personalization. Book: AH. 2000.
8. Brusilovsky, P. and D. W. Cooper. ADAPTS: Adaptive hypermedia for Web-based

nd
performance support system. Proceedings of the 2 Workshop on Adaptive Systems and
User Modeling on the WWW. 1999.
9. Cingil, I., Dogac, A., and A. Azgin. A broader approach to personalization.
Communications of the ACM 43(8):136--141, August 2000.
10. Fayad, M., Schmidt, D., and Johnson, R. Building Application Frameworks, Wiley
Computer Publishing, 1999.
11. Fink, J., Kobsa, A., and J. Schreck. Personalized Hypermedia Information Provision
through Adaptive and Adaptable System Features: User Modeling, Privacy and Security
Issues. Proc. of the Workshop Adaptive Systems and User Modeling on the World Wide
Web of the 6th Int. Conf. on User Modeling, Chia Laguna, Sardinia, June 1997.
12. Hof, R. D., Green, H., and Himmelstein, L. Now it’s YOU WEB. Business Week, pages
68-74, October 5, 1998.
13. Kamba, T., Bharat, K., and M. C. Albers. The Krakatoa Chronicle – An Interactive,
Personalized, Newspaper on the Web. In Proc. of the Fourth International World Wide
Web Conference, pp. 159-170, Nov 1995.
14. Khan, L., and D. McLeod. Audio Structuring and Personalized Retrieval Using
Ontologies. In Proceedings of IEEE Advances in Digital Libraries, Library of Congress,
Washington, DC,May 2000.
15. Martins, L.C. Personalização de Visões sobre Documentos Hipermídia. Technical report
in preparation. Informatics Department, Pontifical Catholic University of Rio de Janeiro,
2002.
16. Ossenbruggen, J.V., Geurts, J., Cornelissen, F., Rutledge, L., and Lynda Hardman.
Towards Second and Third Generation Web-Based Multimedia. In The Tenth
International World Wide Web Conference, pages 479-488, Hong Kong, May 1-5, 2001
17. Parsaye, K. PQ: The Personalization Quotient of a Website. Published by
personalization.com. NovuWeb, Inc. 2000.
18. Tarr, P., Ossher, H., Harrison, W., and S. Sutton. N Degrees of Separation: Multi-
Dimensional Separation of Concerns. Proceedings 21st International Conference on
Software Engineering (ICSE'99), May 1999.
Adaptation in an Evolutionary Hypermedia System:
Using Semantic and Petri Nets1
1 2 1
Nuria Medina-Medina , Lina García-Cabrera , Mª José Rodríguez-Fortiz ,
1
and José Parets-Llorca
1
Depto. L.S.I. E.T.S.I. Informática. Universidad de Granada. SPAIN
{nmedina, mjfortiz, jparets}@ugr.es
2
Depto. Informática. Universidad de Jaén. SPAIN
[email protected]
Abstract. In this paper a classification of adaptive hypermedia systems is

presented. Advantages and disadvantages of these systems are also discussed.
As a consequence, the need of evolution is argued. An adaptive and evolving
hypermedia system is presented and outlined, where the user model is
formalized by means of Petri Nets. Finally, an example shows how in our
proposal the user model is initialized and updated during the browsing and how
adaptation is carried out over the conceptual structure of navigation.
Keywords. Hypermedia System, Adaptation, Petri Net, Evolution, SEM-HP

model
1 Introduction: A Taxonomy of Adaptive Hypermedia Systems
Adaptive hypermedia systems (AHS) appeared with the purpose of improving the
usability of traditional hypermedia systems. Most of them facilitate the activity of the
user, because they adapt the system to some user features. The design of an adaptive
hypermedia system suggests four interrogations concerning adaptation: What?
Whom? How? When?. Following De Bra [8,25] three elements are implicit or
explicitly presented in most adaptive hypermedia systems: domain model, user model
and adaptation model. Depending on the answer to the previous questions, the author
will design these elements. Table 1 summarizes the approaches followed in literature
from different points of view and relates each criterion to the most important solved
interrogation.
On the basis of the revision and analysis realized over the current situation of
adaptive hypermedia systems, we observe a lot of benefits due to adaptation features,
which convert these systems in very powerful tools. Also, we notice some problems;
for instance the entire life cycle (design, construction and maintenance) of adaptive
hypermedia systems are not enough considered and the authoring-tools do not
incorporate mechanisms that facilitate changes in the system during and after its
construction. Considering these two problems of AHS and the fact that hypermedia
systems should change continuously, we propose an evolving approach to the
1 This research is supported by MEIGAS, a project financed by the Spanish MCT (TIC2000-
1673-C06-04) which is a subproject of the DOLMEN project (TIC2000-1673-C06).

Adaptation in an Evolutionary Hypermedia System: Using Semantic and Petri Nets 285
Table 1. A taxonomy of adaptive hypermedia systems
Criterion Types
General adaptive hypermedia systems: The documents include very different
subjects
Application Specific adaptive hypermedia systems: All the documents explain concepts of an
domain uniform information domain. ADAPTS [5] is an adaptive diagnostics and
personalized technical support project, ELM-ART [7] is a Lisp course, PUSH
What? project [9] reorganize a manual on an object-oriented software development
method
Representatio Pairs (attribute/value). AHA [24,25]
User
n of the user Bayesian model. KBS Hyperbook [14]
model Probabilistic and episodic approach. ATS [21]
Adaptation to Petri Net [Our proposal]
Personalized recommendations. Approach of Johan Bollen[4]
Whom? User group
Recommendations to user groups. INTRIGUE [2]
Others. For instance, adaptation to different kinds of devices in mobile
technology [1]
Guidance Local or Global
Orientation support Local or Global
Adaptive Adaptive navigation
Personalized views
methods Additional explanations
[6] Adaptive presentation
Prerequisite explanations
Comparative explanations
What? How? Explanation variants
Sorting
Kind of Pedagogical prerequisites: State relationships between concepts related to
learning
prerequisites [16] Prerequisites as ordering mechanism: Try to establish a partial order between
How? pages
Integrating of Open adaptive hypermedia systems: These systems can integrate information
information from resources located anywhere in the WWW. KBS Hyperbook [14]
different origin
What? Closed adaptive hypermedia systems
Adaptable hypermedia systems: User model is only updated after an explicit
solicitation of the user
Interactivity with Adaptive hypermedia systems: The
The user can have some control over the
the adaptivity user model is automatically updated
adaptive behavior of the system. PUSH [9]
as the user browses and the content
When? How? and the link structure are adapted to The user can’t have any control over
it adaptation
Adaptable/Adaptive hypermedia systems. Approach Intensional Hypertext [23]
Creating Dynamic AHS: Documents are dynamically created under user demand
No dynamic AHS: Documents exist before of their use, though their presentation
hypermedia is adapted to each user
documents Mediating between adaptive and dynamic hypermedia. Macronodo Approach
When? How? [18]
Explicit history representation. A kind of navigation rule uses path history [15]
Navigation
history The representation of user browsing path is not explicit
How?
User context: Role of the user in a group, physical localization, etc.
Contextual Textual context: Surrounding phrase, paragraph or document
information Spatial context: User browsing path through the hyperspace before arriving at
Whom? present page. Context-based approach [3]
286 N. Medina-Medina et al.
construction and maintenance of adaptive hypermedia systems. We have two main

reasons in doing this. The first reason is that AHS are software systems and the
quality of the obtained product depends on the quality of the development process.
Consequently, we propose a software engineering approach in the development of
adaptive hypermedia systems. This process is divided in four phases that we consider
inherent to the design of a conceptual and navigational system. We will further
explain this process.
The second reason in taking one such approach stems from the fact that the
preparation of hyperdocuments includes a lot of changes, additions and updates
carried out by the authors; this implies that the author needs change mechanisms.
Therefore, we consider that user adaptation in current AHS should be completed with
mechanisms of evolution [22]. Adaptation carried out by AHS is adaptation to
different users, however evolution implies the possibility of changes in the system,
which are produced by the developer in order to improve the system. Transformations
happened in the system context would determine which are the changes that the
author (developer) must carry out in the system.
In this paper we present the main features of our approach, called SEM-HP and the
integration of adaptation features using Petri Nets. In Sect. 2 the main characteristics
of the model are briefly outlined including adaptive features. Section 3 presents an
example of adaptation, and finally the related works, conclusions and further work are
considered.
2 SEM-HP: Evolutionary and Adaptive Hypermedia System
SEM-HP [13] is a systemic, evolutionary and semantic model for the development of
adaptive hypermedia systems. It is systemic because it conceives a hypermedia
system as a set of interacting subsystems. It is semantic because it offers a flexible
semantic approach which allows to author characterizes his information domains. An
explicit semantic representation increases the possibilities of adaptation and
evolution.
Joining the two reasons discussed in the previous section, we propose a
evolutionary and iterative development process for adaptive hypermedia systems,
where each phase integrates changes of the system structure produced by the
developer in a flexible and consistent way:
• Conceptual phase. The author constructs conceptual and information worlds.
• Presentation phase. The author selects different presentations of a concrete
conceptual and information world.
• Navigation phase. The author states how the reader can browse the offered
information.
• Learning phase. The author resolves the aspects of adaptation. Here, the author
answers the four essential questions: What? Whom? How? When?, i.e., he must
decide representation, initialization and updating of the user model and adaptive
techniques.
2.1 Architecture of SEM-HP
In SEM-HP an adaptive hypermedia system is conceived as composed of four

interrelated and interacting systems, each one corresponding to a development phase
[12]:
Memorization System. It is in charge of the storage, structuring and maintenance of
the knowledge that the AHS offers. The main element of this system is the
Conceptual Structure (CS). The CS is a semantic net, in which, nodes and links are
labeled with semantic meanings. The CS has two types of nodes: concepts and items.
Concepts are labeled ideas and items contain pieces of information, which are related
to the concepts. Links in the CS are relationships among concepts or among concepts
and items. The author can change the CS after or during its construction, in order to
do it, evolutionary actions are provided which allows the creation, deletion or
modification of concepts, items or relationships.
Presentation System. This system allows a selection of subsets of concepts, items
and relationships included in the conceptual structure.
Navigation System. It uses as basis the CS of the presentation system and allows
the addition of special order restrictions in the graph. These restrictions are expressed
in temporal logic and they state a partial order between the pieces of information
offered by the system, i.e., they provide a partial order of the items. Considering the
CS and these order restrictions a Petri Net [19] is automatically constructed, which
allows an operational navigation. Evolution is carried out deleting or modifying
restrictions.
Learning System. Adaptation of the hypermedia system according to an user model
is carried out by this system. Also in this system, evolution is needed because the
author can modify, add or delete adaptation rules or user model updating rules.
Furthermore, when the author modifies the CS in the knowledge system or he changes
order relationships in the navigation system, these changes can imply changes in some
elements of the learning system. In SEM-HP these last changes are carried out
automatically by the system, facilitating the work of the author and guarantying the
system coherence. For example, if the author adds new concepts in the CS, these
concepts will be included into the user model, with the objective of keeping track of
the knowledge level of the user about these concepts.
The three first systems are fully described and formalized in [13]. The user
adaptation performed by the learning system is described in paragraph 2.3 and
exemplified in Sect. 3.
2.2 Evolution in SEM-HP
Because adaptive hypermedia systems offer knowledge captured by their authors,

they change frequently. In order to carry out these changes, a set of evolutionary
actions is provided by SEM-HP model [13]. The author uses the different
evolutionary action depending on the modifications that he needs. In the SEM-HP
model, a set of restriction that the system must satisfy at any moment exists. In order
to keep the consistency, an evolutionary action only will be executed if it fulfils the
set of restrictions. In addition, new restrictions can be added by the author to this set,
it will be performed through a special kind of evolutionary actions which restrictions
are named meta-restrictions. The model also takes into account that the modifications
of an element into a system (memorization, presentation, navigation or learning) can

imply some modifications in other elements of the system itself or of others systems.
These two situations are named respectively internal propagation of changes and
external propagation of changes.
2.3 User Adaptation in SEM-HP
Obviously, the learning system supports the greatest charge of adaptation. It initializes
and maintains the user model, and performs adaptation. Both navigation adaptation
and presentation adaptation is supported by SEM-HP. In this paper, we focus on the
navigation adaptation. In SEM-HP, the main browsing mechanism is the conceptual
structure. Therefore, it will be the recipient of the navigation adaptation.
Table 2 shows the information stored in the user model. The third column of the
table indicates what user features are taken into account in SEM-HP. We think that
the user models considered in literature include two types of features: 1) Variable
features which change frequently during navigation -rows 1 to 5- and 2) Stable
features whose changes are long term -rows 6 to 9-. Variable features are represented
by a Petri Net, except the goal which is not considered at present. Stable features are
stored apart from the Petri Net.
Table 2. User model in SEM-HP

Information Description SEM-HP?
Goal Information that the user wants to know NO
Knowledge User knowledge value about each item. 0 <= N:Integer <= 100 OK
Read Items read by the user OK
Read number Readings number of each item OK
Ready to read Items that can be read by the user OK
Subject experience User knowledge about the conceptual domain of the system OK
Navigation experience Practice of the user in the use of hypermedia systems OK
Preferences Predilections and tastes of the user OK
Personal data Data as name, age, sex, profession, etc. OK
Navigation adaptation techniques are applied over the conceptual structure of

navigation, in order to tailor it to user features:
• Links annotation. In the current conceptual structure (CS), each item previously
visited is annotated with both the number of visits and a violet color. In addition,
concepts and items are annotated indicating the user knowledge level about them.
• Links hiding and disabling. An item only can be reached if the user knows its
prerequisite items. When it does not occur the item is hidden and disabled.
• Personalized CS. In the presentation phase, the author creates different views of
the same CS. For each user the view that better tailors his knowledge and
preferences is selected.
3 Adaptation in SEM-HP: An Example of Use
The following example describes how both adaptation of conceptual structure and
updating of the user model is performed in SEM-HP, using a Petri Net.
Step one. A conceptual structure (CS) is created by the author. Concepts are
represented as circles. Items -documents offered to the reader- are represented as
squares. Arrows are relationships among concepts and lines are relationships among
concepts and items. Conceptual relationships means prerequisites between concepts.
In Fig. 1, the arrow from C3 to C4 declares that if the user knows the concept C3 can
read the concept C4.
I1 I1 C3 C4 I1
C1 C2
I2
I1 I2
Fig. 1. Example of conceptual structure. Circles and squares are labeled
Step two. The system generates automatically order restrictions among items, these
order restrictions are written in propositional temporal logic. We will call them order
rules. Because relation concept-item has cardinality N-N, an item is identified using
the name of the concept and its name. Table 3 shows the four rules automatically
generated from CS of Fig. 1. The first rule states that the order prerequisite for the
item C2.I1 is the previous lecture of item C1.I1 or item C1.I2.
Table 3. Order rules: Order restrictions among items of the CS

1) C1.I1 or C1.I2 ! C2.I1 3) C2.I1 or C2.I2 ! C4.I1
2) C1.I1 or C1.I2 ! C2.I2 4) C3.I1 ! C4.I1
Step three. Order rules can be modified by the author, if desired. An order rule can be
deleted, its logical operators can be altered or it can be extended by the author. For
instance, the author modifies the second rule (substitutes or for and), he extends the
third rule and he deletes the fourth rule of the table 3.
Table 4. Modifications performed by the author, using a graphic interface

Modifications Final set of order rules
2) C1.I1 and C1.I2 ! C2.I2 1.a C1.I1 or C1.I2 ! C2.I1
3) C1.I1 and C3.I1 and (C2.I1 or C2.I2) ! C4.I1 2.a C1.I1 and C1.I2 ! C2.I2
4) C3.I1 ! C4.I1 3.a C1.I1 and C3.I1 and (C2.I1 or C2.I2)!C4.I1
Step four. The influence of an item over knowledge of its associated concept is
defined by the author for each item in the CS. In the example, the author has decided
that both item C1.I1 and item C1.I2 have the same weight upon the knowledge of the
concept C1, however C2.I1 has more weight than C2.I2 upon concept C2.
100%
I1 C3 C4 100% I1
I1 50%
C1 C2
60% 40%
I2 50%
I1 I2
Fig. 2. Weights defined by the author
Based on these weights, the system generates automatically a set of weight rules in
order to calculate user knowledge about each concept at every moment. Knowledge
about a concept is based on knowledge about its associated items. Table 5 displays the
generated weight rules. The function K returns the knowledge value of the argument
(item or concept).
Table 5. Weight rules generated from Fig. 2

K(C1) = 50%K(C1.I1) + 50%K(C1.I2) K(C3) = K(C3.I1)
K(C2) = 60%K(C2.I1) + 40%K(C2.I2) K(C4) = K(C4.I1)
Step five. Knowledge prerequisites are defined by the author for each order rule
generating the knowledge rules. In order to do it, the author establishes knowledge
thresholds for each prerequisite item into an order rule. The knowledge threshold of
an item says what is the lowest knowledge value required for it in the current rule.
Figure 3 shows the definition of knowledge thresholds for the rule 3.a of table 4. The
required items (left part of the rule) are identified by means of prohibition signs. The
head of the rule is identified by means of a smiling face. The author writes knowledge
threshold next to each prohibition sign. In the example, the author has decided that the
user will be ready to read C4.I1 if the user knowledge about C1.I1 is equal to 100 and
C3.I1 is higher than 40 and C2.I1 higher than 50 or C2.I2 higher than 30.
>40 I1 C3 C4 I1
=100 I1
C1 C2
I2 I1 I2
or
>50 >30
Fig. 3. Knowledge rule 3.b generated from the order rule 3.a
The following table shows the set of knowledge rules defined from rules of table 4.
Table 6. Knowledge rules

1.b. K(C1.I1)>40 or K(C1.I2)> 60 ! C2.I1 2.b. K(C1.I1)>75 and K(C1.I2)>50 ! C2.I2
3.b. K(C1.I1)=100 and K(C3.I1)>40 and (K(C2.I1)>50 or K(C2.I2)>30) ! C4.I1
Step six. The author states how the user knowledge about items increases during
navigation. In each rule a smiling face marks the current item and a ring marks those
items whose knowledge value will be updated due to a visit to the current item.
Symbol + represents a relative update (an increase). Symbol = represents an absolute
update (fixed value). Symbol % represents a percentage on the knowledge value of
the smiling item. Figure 4 implies: each time that item C1.I1 is visited its knowledge
value is increased on 50 and knowledge value of item C3.I1 is increased on 50% of
the current knowledge value of C1.I1. In this example, the full knowledge of C1.I1
requires two visits.
+ %50 I1 C3 C4 I1
+ 50 I1
C1 C2
I2
I1 I2
Fig. 4. Graphic interface for the update rule: Visit(C1.I1)! InK(C1.I1, 50),
InK(C3.I1, 50%K(C1.I1))
The rules generated are called update rules and are represented internally using
logic. The predicate Visit(Cc, Ii) will be true when the item Ii associated to the
concept Cc will be visited by the reader. The function FixK(Cc.Ii, N) establishes the
knowledge value of item Cc.Ii to value N (0<=N<=100). The function InK(Cc.Ii, N)
increases the knowledge value of the item Cc.Ii on value N (the result is normalized
between 0 and 100). The visited item is always updated at first place.
Table 7. Internal representation of update rules defined by the author

Up11 Visit(C1.I1)!InK(C1.I1, 50), InK(C3.I1, 50%K(C1.I1)) Up22 Visit(C2.I2)!InK(C2.I2, 25)
Up12 Visit(C1.I2)!FixK(C1.I2, 100), InK(C3.I1,
Up31 Visit(C3.I1)!FixK(C3.I1, 100)
10%K(C1.I2))
Up21 Visit(C2.I1)!FixK(C2.I1, 100), InK(C3.I1,
Up41 Visit(C4.I1)!FixK(C4.I1, 100)
10%K(C2.I1))
Step seven. Based on previous information (from step 1 to step 6) a Petri Net (PN) is
generated automatically by the system. This PN represents and updates the user model
described in section 2.3. The PN is obtained applying an adaptation of the algorithm,
which transforms temporal logic formulas in a Petri Net, explained and demonstrated
in [20]. The PN integrates the user model, knowledge rules and update rules (Fig. 5).
In SEM-HP, the user navigates by selecting items in the CS. The selection of an
item is represented by means of a transition in the PN. Therefore, knowledge rules
are split as conditions of the transitions, depending on logical operators and following
[20]. Consequently, an item can’t be selected if its prerequisites are not satisfied. An
update rule is associated to the output arc of the transition that leads up to the selected
item. So when the reader visits an item this update rule is executed. There are four
types of places in the PN:
• User-selection place. There is a sole place of this type in the PN. This place
determines what the user selection is, that is to say, what is the fired transition in
the PN. It has always one token. There is an input arc to each transition. In order to
conserve the token there is an output arc from each transition.
• Item places. Each item Cc.Ii in the conceptual structure has associated an item
place in the PN. It represents the item and stores the user knowledge value of this
item. In Fig. 5: C1.I1, C1.I2, C2.I1, C2.I2, C3.I1and C4.I1 are item places. Initially
each one of them has one token, but each visit to the item increases its number of
tokens in one. Therefore, the number of visits is equal to the number of tokens
minus one. There is an output arc to each transition whose condition evaluates the
knowledge of this item. There is an input arc, which has associated the update rule,
that must be executed in each visit to this item.
• Starting places. A starting place for each item Cc.Ii without knowledge
prerequisites exists. C1.I1 Start, C1.I2 Start, C3.I1 Start in Fig. 5. It has always one
token. There is an input arc to the transition, which leads up to the item place Cc.Ii.
• Assistant places. This type of place represents a condition, which is part of a
complex knowledge rule. In Fig. 5: X. Initially these places have no tokens, it
means that the represented condition is not satisfied.
Figure 5 shows the PN generated by the system. For clarity reasons only the
interaction of the user-selection place with one transition is shown. An arc with a
double arrow represents one input and one output arcs. Upci is the update rule of the
table 7, which is executed when item Cc.Ii is visited.
C1.I1Start User-selection C1.I2Start
Up11 Up12
C1.I1 C1.I2
K(C1.I1)>75
K(C1.I1)>40 K(C1.I2)>60 and K(C1.I2)>50
Up21 Up21 Up22
C2.I1 C2.I2 C3.I1Start
K(C2.I2)>30
X Up31
C3.I1
K(C1.I1)=100 and K(C3.I1)>40

Up41
C4.I1
Fig. 5. Petri Net automatically generated
Step eight. Firing the Petri Net both adaptation and user model updating is carried out
while the user navigates. Figure 6.a displays the initial CS. In the PN, only the item
places C1.I1, C1.I2 and C3.I1 can be visited firing the transitions from the user
selection place. Therefore, items C2.I1, C2.I2 and C4.I1 are hidden and disabled in
the CS of navigation to avoid that the reader selects them. Initially, items and
concepts are annotated with knowledge value 0.
0% 0% 0% 25% 25% 0%
I1 C3 C4 I1 I1 C3 C4 I1
0% 50%
I1
0% C2 I1 25% C2
C1 0% C1 0%
1
I1 I2 I1 0% I2
I2 0% I2 0%
Fig. 6.a Initial CS disabling the forbidden items Fig. 6.b CS after item C1.I1 is visited
The rule Up11 is executed when the user selects the item C1.I1. This rule increases
the knowledge of both C1.I1(in 50) and C3.I1(in 50%50). Consequently, the
knowledge on concepts C1 and C3 has been increased. In the PN, the condition which
enables the transition to C2.I1 will be true. Then in the CS, the item C2.I1 will be
visible and the item C1.I1 is annotated indicating one visit (Fig. 6.b).
4 Related Work, Conclusions and Further Work
Hypermedia models that use finite state machines different from Petri Nets have been
proposed in the scientific literature. For instance, HMBS [10] is a hypermedia model
based on statecharts. We use Petri Nets because have an inherent mechanism of
execution, whereas the classical formalism of statecharts is not executable. Statecharts
are more focused on specification of the structure and static browsing semantic of
hypermedia systems, for this purpose we use the conceptual structure and the logic
rules.
The caT model [17] is a context-aware hypertext model, which also uses Petri
Nets. CaT extends Trellis hypertext model [11], using a high-level PN in order to
authoring, browsing and analysing dynamic documents. It supports user adaptation to
changes in environmental information (localization, time, etc.). However, the caT
model follows an approach very different from our proposal, because it builds directly
the PN-based document structure, without constructing a conceptual structure, neither
performing conceptual and presentation phases. Moreover, caT does not support
adaptation to user knowledge, due to that it does not establish knowledge rules and
does not update the user knowledge during navigation. Another difference is that in
our case the PN represents the variable features of the user model while in caT the
user model is a globally visible file used in the PN, which contains the less dynamic
data of the user such as work organization. In addition, caT and the works referenced
in table 1 [1-9, 14-16, 18, 21, 23-25] do not offer mechanisms of evolution.
Following the taxonomy of table 1, the SEM-HP model can be characterized as a
model that allows the creation of open systems with a generic application domain, in
these systems both evolution and user adaptation is performed. Evolutionary
modifications can be carried out on the created system by the author in a flexible way,
due to the evolution mechanisms supported in the model. In addition the user model is
represented by means of a Petri Net, both adaptive navigation and adaptive
presentation is supported, the prerequisites between items are used as order
mechanism and navigation history is implicitly represented by means of the Petri Net.
However, the hypermedia systems designed using the model are not dynamic because
the information items are a priori established.
Our future work will be centered around three main problems related to the
dynamic construction of presented information: a) How can the system find out the
user goals? b) How can we represent and update the user subjective experience and
item difficulty over the Petri Net? c) How can the system generate items in a dynamic
way depending on the user goals and user experience? At present the last two
problems are partially considered in the model by means of certain item properties
which characterize the information items and dynamic composition of items, but this
approach requires further refinements.
References
rd
1. Alatalo T. Peräaho J. A Modelling Method for Designing Adaptive Hypermedia. 3
Workshop on Adaptive Hypertext and Hypermedia. UM2001. Sonthofen, Germany. July
13-17, 2001.
2. Ardissono L. Goy A. Petrone G. Segnan M. Torasso P. Tailoring the Recommendation of
rd
Tourist Information to Heterogeneous User Groups. 3 Workshop on Adaptive Hypertext
and Hypermedia. Hypertext’01. Aarhus, Denmark. August 14-18, 2001, LNCS 2266, pp.
280-295.
3. Bailey C. El-Beltagy S. Hall W. Link Augmentation: A Context-Based Approach to
rd
Support Adaptive Hypermedia. 3 Workshop on Adaptive Hypertext and Hypermedia.
Hypertext’01. Aarhus, Denmark, August 14-18, 2001.
4. Bollen J. Group User Models for Personalized Hyperlink Recommendations. LNCS 1892–
Int’l Conference on Adaptive Hypermedia and Adaptive Web-based Systems.,pp. 39-50.
Trento. 2000.
5. Brusilovsky P. Cooper D. ADAPTS: Adaptive Hypermedia for a Web-based Performance
nd
Support System. 2 Workshop on Adaptive Systems and User Modeling on the WWW.
Canada, 1999.
6. Brusilovsky P. Methods and Techniques of Adaptive Hypermedia. User Modeling and
User-Adapted Interaction, 6: 87-129. Kluwer Academic Publishers. 1996.
7. Brusilovsky P. Weber G.
http://www.psychologie.uni-trier.de:8000/projects/ELM/elm.html
8. De Bra P. Houben G. Wu H. AHAM, A Dexter-based Reference Model for Adaptive
Hypermedia. ACM Conference on Hypertext and Hypermedia, pp. 147-156. 1999.
9. Espinoza F. Höök K. A WWW Interface to an Adaptive Hypermedia System. Conference
on Practical Application of Agent Methodology (PAAM’96). London. April, 1996.
10. Ferreira-De Oliveira C. Santos-Turine M. Masiero P. A Statechart-Based Model for
Modeling Hypermedia Applications. ACM Transactions on Information Systems. April,
2001.
11. Furuta R. Stotts D. Trellis: A Formally-defined Hypertextual Basis for Integrating Task
and Information. Olson G.M. et al. editors. Coordination Theory and Collaboration
Technology, 2001.
12. García-Cabrera L. Parets-Llorca J. A Cognitive Model for Adaptive Hypermedia Systems.
st
1 Int’l Conf. on WISE. Workshop on WWW Semantics, pp. 29-33. Hong-Kong, China.
June, 2000.
13. García-Cabrera L. SEM-HP: A Systemic, Evolutionary, Semantic Model for Hypermedia
System Development (in Spanish). Ph Thesis. November, 2001.
14. Henze N. Nejdl W. Bayesian Modeling for Adaptive Hypermedia Systems. ABIS99, 7.GI
- Workshop Adaptivität und Benutzermodellierung in Interaktiven Softwaresystemen
29./30.9. 1999. Otto-von-Guericke-Universität Magdeburg.
15. Hijikata Y. Yoshida T. Nishida S. Adaptive Hypermedia System for Supporting
rd
Information Providers in Directing Users through Hyperspace. 3 Workshop on Adaptive
Hypertext and Hypermedia. Hypertext’01. Aarhus, Denmark. August 14-18, 2001, LNCS
2266, pp. 322-326.
16. Hübscher R. What’s in a Prerequisite. Int’l Conference on Advanced Learning Technology
(ICALT 2001). Madison, WI, USA. 2001.
17. Na J. Furuta R. Dynamic Documents: Authoring, Browsing, and Analysis Using a High-
Level Petri Net-Based Hypermedia System. ACM Symposium on Document
Engineering’2001, pp.38-47. Atlanta, Georgia. November, 2001.
18. Not E. Zancanaro M. The MacroNode Approach: Mediating Between Adaptive and
Dynamic Hypermedia. Int’l Conference on Adaptive Hypermedia and Adaptive Web-
based Systems (AH’2000). Trento. August, 2000, LNCS 1892, pp. 167-178.
19. Peterson J. Petri Net Theory and the Modeling of Systems. Prentice-Hall. Englewood
Cliffs, 1981.
20. Rodríguez-Fortiz M. Software Evolution: A Formalisation Based in Predicate Temporal
Logic and Coloured Petri Nets (in Spanish). Ph Thesis. October, 2000.
21. Specht M. Oppermann R. ATS – Adaptive Teaching System a WWW-based ITS. U.Timm
(Eds.). Workshop Adaptivität und Benutzermodellierung in Interaktiven
Softwaresystemen: ABIS 98.
22. Torres-Carbonell J. Partets-Llorca J. A Formalization of the Evolution of Software
Systems. Computed Aided Systems Theory, EUROCAST’99. Pp. 269-272. Vienna.
September, 1999.
23. Wadge B. Schraefel M. A Complementary Approach for Adaptive and Adaptable
rd
Hypermedia: Intensional Hypertext. 3 Workshop on Adaptive Hypertext and
Hypermedia. Hypertext’01. Aarhus, Denmark. August 14-18, 2001, LNCS 2266, pp. 327-
333.
24. Wu H. De Bra P. Aerts A. Houben G. Adaptation Control in Adaptive Hypermedia
Systems. Adaptive Hypermedia Conference (AH2000), LNCS 1892, pp. 250-259. 2000..
25. Wu H. Houben G. De Bra P. Supporting User Adaptation in Adaptive Hypermedia
Applications. On-line Conference and Informatiewetenschap 2000 (InfWet2000). De
Doelen, Rotterdam. 2000.
Evaluating the Effects
of Open Student Models on Learning
Antonija Mitrovic and Brent Martin
Intelligent Computer Tutoring Group

Computer Science Department, University of Canterbury
Private Bag 4800, Christchurch, New Zealand
{tanja,bim20}@cosc.canterbury.ac.nz
Abstract. In previous work [10], we reported on an experiment performed in

the context of SQL-Tutor, in which we analysed students’ self-assessment
skills. This preliminary study revealed that more able students were better in
assessing their knowledge. Here we report on a new study performed on the
same system. This time, we analysed the effect of an open student model on
students’ learning and self-assessment skills. Although we have not seen any
significant difference in the post-test scores of the control and the experimental
group, the less able students from the experimental group have scored
significantly higher than the less able students from the control group. The
more able students who had access to their models abandoned significantly less
problems the control group. These are encouraging results for a very simple
open model used in the study, and we believe that a more elaborate model
would be more effective.
1 Introduction
Self-assessment is one of the meta-cognitive skills necessary for effective learning.

Students need to be able to critically assess their knowledge in order to decide what
they need to study. The same skill is also important for students to assess the
difficulty of the problem they are working on, and to decide whether to abandon to
problem or keep working on it. Intelligent educational systems must support the
acquisition of meta-cognitive skills in order to support deep learning, and therefore
must also support the acquisition of self-assessment skills.
In previous work [10], we hypothesized that more able students were better when
assessing their own knowledge than their less able peers. This hypothesis is in
accordance with findings from other studies [1]. In order to evaluate the hypothesis,
we analysed students’ self-assessment skills in the context of the SQL-Tutor system.
The system was modified slightly to allow for data collection. We focused on
situations when students abandoned the problem they were working on, and asked for
a new problem. In such situations, students were asked two questions. Firstly, we
asked the student to specify the reason for abandoning the current problem. Three
possible replies were offered: the student may think that the current problem was too
easy or too difficult, or may simply want to work on a problem of a different nature.
The student was then asked to specify what kind of problem they would like to work
on next. For this purpose, problems were characterized by the clause, so seven options

Evaluating the Effects of Open Student Models on Learning 297
were available, one for each clause of the SELECT statement, plus the any clause
option. The results of the data analyses performed supported our hypothesis.
In this paper, we report on a new experiment conducted in the context of the same
system. We were interested in using open student models as a way to support self-
assessment skills. We begin by overviewing related work in Sect. 2, and then
introduce SQL-Tutor briefly in Sect. 3. SQL-Tutor provides a facility for students to
select problems on their own, which requires students to be able to assess their own
knowledge. Section 4 describes the experiment performed, and is followed by a
description of the findings in Sect. 5. The conclusions are presented in the final
section.
2 Related Work
Metacognition includes the processes and activities involved with awareness of,
reasoning and reflecting about, and controlling one’s cognitive skills and processes. A
number of studies showed that better metacognitive skills result in improved problem
solving and better learning [2, 6, 7, 15, 16], and that such skills can be taught [3, 6].
White et al. [16] define metacognition as consisting of three fundamental
components: knowledge about knowledge (including cognitive and social expertise),
regulatory skills, needed to monitor and control knowledge and development
expertise, which allows students to reflect and improve knowledge. Their hypothesis
is that students need to be conscious of theories for socio-cognitive processes for
learning. In [16] they present SCI-WISE, a system that supports inquiry learning. In
this type of learning, students cycle through several stages, starting by formulating a
question, followed by the generation of hypotheses, and experimentation. After that,
collected data is analysed, and students develop a model on the basis of results of
analysis, and finally evaluate the model by applying it in a novel situation. The
system supports inquiry learning by asking the student to explicitly state the purpose
of each action and provide a justification for it. The system provides a simulation
environment in which to perform experiments, and incorporates several agents, each
of which specializes in a particular metacognitive process, such as inventing,
collaborating and analysing, and advises the student in relevant situations. The student
develops his/her own theory of scientific inquiry by modifying the agents so that they
reflect student’s beliefs. In order to test his/her theory, the student then engages in a
physics project, during which agents provide advice. The student evaluates the agents’
behaviour, which may result in the modification of the theory.
Most of the metacognitive research focuses on self-explanation, which is a skill of
“generating explanations and justifications to oneself to clarify an example solution”
[2, 7]. Self-explanation enables students to focus on general principles by examining
specific examples [6]. Aleven and Koedinger [1] evaluate students’ abilities to
identify situations when help is needed and to ask for appropriate help. They show
that not all students possess this skill, and recommend several ways in which the
system may support students in acquiring it.
Reflection is a metacognitive skill in which students examine their own knowledge.
This metacognitive skill is encouraged by allowing the student to inspect and, in some
cases, to modify the student model. Several projects focused on the effects of opening
298 A. Mitrovic and B. Martin
the student model on students’ learning [4, 5, 8, 9]. The efforts range from simply
visualizing the model, to actively involving the student in the modeling process
through negotiation or collaborative construction of the model. In these systems, the
student model is not just a source of knowledge about the student of value to the
system, but becomes an important learning resource in its own right.
The representations used to visualize the student model range from simple
skillometers to very complex ones. Skillometers allow for an easy to understand,
high-level overview of the student model, and have been used in ELM-ART [4] and
cognitive tutors [1]. In addition to just showing the student model at a high level,
some systems allow the student to challenge and negotiate the content of the model.
This process is referred to as open interactive [8], collaborative [5], cooperative [9] or
participative [12] student modelling. Such approaches use more complex
representations of the student model based on conceptual graphs [8], Bayesian
networks [17], tree structures [9], tables [5] and Prolog clauses [14]. These projects
differ in the content of the student model, representation chosen for visualizing the
model, the type of interactions (inspection or negotiation) and the mechanism for
dealing with conflicts. The area is still just emerging, and proper ways of evaluating
developed systems is yet to be identified. In all projects only very limited evaluation
has been done, the results of which are not irrefutable.
3 SQL-Tutor
SQL-Tutor is an intelligent educational system aimed at university-level students

learning SQL. For a detailed discussion of the system, see [11]; here we present only
some of its features. SQL-Tutor consists of an interface, a pedagogical module, which
determines the timing and content of pedagogical actions, and a student modeller,
which analyses student answers. The system contains definitions of several databases,
and a set of problems and the ideal solutions to them. SQL-Tutor contains no problem
solver. To check the correctness of the student’s solution, SQL-Tutor compares it to
the correct solution, using domain knowledge represented in the form of more than
500 constraints. It uses Constraint-Based Modeling [13] to model knowledge of its
students. Students may work their way through a series of problems for each database,
or ask the system to select a problem on the basis of his/her student model.
4 The Study
The purpose of this study was to determine the effect of a simple open student model
on students’ learning and self-assessment skills. Let us first describe the way we
visualize the student model. The student model in SQL-Tutor is implemented as an
overlay on top of the constraint base. There are currently more than 500 constraints in
the system, and therefore it is not possible to visualize information about each
constraint. Instead, we have decided to compress the student model into a simple
structure that resembles the structure of the SELECT statement. The student is shown
six skillometers, which show the student model in terms of the six clauses of the
SELECT statement. For each clause, we find all the relevant constraints, and compute
the coverage (the percentage of constraints that the student has used) and correctness
(i.e. the percentage of all relevant constraints that the student has used correctly).
These two percentages are visualized as shown in Fig. 1.
We focus on situations when students abandon the current problem and ask for a
new one. In such cases, the students were asked the same two question as in the
preliminary study: to specify the reason for abandoning the current problem, and to
specify the type of the problem they would like to work on next. The available
options were the six clauses of the SELECT statement (SELECT, FROM, WHERE,
GROUP BY, HAVING, and ORDER BY), plus the any clause option. For the
experimental group, if the student’s selection of the problem type was different from
the one suggested by the system, the student was shown the open model, and asked
whether they wanted to continue with the type of problem they specified, or whether
they would prefer to switch to the system’s suggestion.
The participants were enrolled in an introductory database course at the University
of Canterbury, New Zealand, in the second half of 2001. Prior to the experiment, all
students listened to four lectures on SQL and had two labs on the Oracle RDBMS.
During the experiment, there were 4 additional lectures on SQL, and a series of four
Fig. 1. The visualization of the student model
more labs. The experiment required the students to sit a pre-test, which was
administered in a lecture on 10 September 2001. The pre-test consisted of three
multichoice questions. All three questions contained the text of the query, and some
solutions. The students were asked to classify the solutions as correct or incorrect.
The maximum mark for the pre-tests was 9.
The students who sat the pre-test were given user accounts to use in SQL-Tutor
from September 12, and were randomly allocated to one of the three possible versions
of the system. The experiment was designed to combine two evaluations into one: the
evaluation of an enhanced constraint and problem set, and the evaluation of the open
student model. The control group served as the control group for both studies. The
problem selection group is not of importance for this paper, and the experimental
group is the group that had access to the open student model.
The interaction with the system was voluntary. The course involved a test on SQL
a month after the system was introduced to the class. The post-test consisted of three
questions of similar nature and complexity as the questions in the pre-test, and was
administered in a lecture, on 9 October 2001.
5 Results
This section presents the results of the analyses performed on the data collected in the
experiment. Section 5.1 presents the general findings about how students learnt with
SQL-Tutor. Data analyses relevant to our hypotheses are discussed in Sect. 5.2.
5.1 Learning with SQL-Tutor
Out of 159 students enrolled in the course, 100 sat the pre-test. Table 1 gives the
number of students in each group, their pre-test scores, and some additional
information about their logs. The mean score for the pre-test for the whole class was 5
(out of 9, SD=1.36). The t-test reveals there are no significant differences between the
pre-test scores, which means that the three groups are comparable. The problem
selection group is of no importance for this paper, so we report on the control and
experimental groups only.
As the usage of the system was voluntary, 80 students actually logged on to SQL-
Tutor. Table 1 gives the number of students in each of the groups who sat the pre-test,
and also the number of students who actually used the system. However, some of
these students have only briefly looked at the system. We excluded the logs of 9
students who attempted no problems, and the number of valid logs is given in the
table.
Table 1. The three groups

Group Students Pre-test Accounts Valid logs Post-test
mean (SD) used
Control 34 4.82 29 24 12
(1.44)
Experimental 33 5.12 (1.41) 23 21 12
Problem selection 33 5.06 (1.25) 28 26 14
There were 101 students who sat the post-test (mean=5.99, SD=1.60). In the post-
test, the students were asked whether they have used SQL-Tutor and, if they have, to
specify their account. The mean score on the post-test for students who have used
SQL-Tutor was 6.42 (SD=1.50), which is higher than the overall mean. The mean
score of the students who have not used SQL-Tutor was 5.67 (SD=1.61). However,
we cannot claim that SQL-Tutor was responsible for this higher mean, as the
participation in the study was voluntary, and the students who participate are usually
more motivated.
Although we asked students to specify the account if they used the system, some
students have not done that. The last column in Table 1 contains the number of
students in each group who specified their account names. These are the only students
whose pre- and post-test results we were able to match. Table 2 contains the statistics
for the pre- and post-test results for those students only. It can be seen that the scores
for the experimental group are slightly higher, but not significantly.
Table 2. Pre- and post-test results for the students who can be matched
Group Matched tests Pre-test mean (SD) Post-test mean (SD)
Control 12 4.42 (1.24) 6.42 (1.38)
Experimental 12 4.50 (1.17) 6.67 (1.56)
Table 31 gives some simple statistics gathered from the valid logs. The number
of sessions ranged from 1 to 13, and the lengths of individual sessions ranged from 1
minute to almost four hours. The total interaction time ranged from 5 minutes to more
than ten hours. The Total solved column gives the mean number of problems the
students have solved during all sessions, which ranges from 1 to 70. Problem/session
gives the average number of problems students have seen in a session (ranges from 1
to 40). Solved/session gives the averages for the number of solved problems (the
minimum for solved problems per session is 0, the maximum is 35), while the
percentage of problem completion is given in the next column. On the average, the
control group needed slightly more time per solved problem than the experimental
group. None of the reported numbers are significantly different.
Table 3. Statistics about the sessions

Session Total Total Problem Solved % prob. Mins
Group Sessions length time solved /session /session solved per
(min) (min) /session solved
Contr 4.7 39 (41) 183 20.5 5.3 (3.1) 4.4 (3.3) 77.6 8.6
(24) (3.28) (189) (15.2) (21.7) (5.2)
Exper. 4 36 (32) 144 22.8 6.7 (3.8) 5.6 (3.7) 77.1 7.2
(21) (2.66) (125) (18.8) (18.2) (3.7)
5.2 Analyzing the Self-assessment Skills
We also analysed the effects of the open student models with respect to students’
abilities. Each group was split into two subgroups, with less and more able students,
depending on their scores on the pre-test. Students who scored above the average (5
or more marks) on the pre-test were put into the more able group, while the students
who scored 0 to 4 marks were put into the less able group. Table 4 contains statistics
about the subgroups, which are of similar sizes. There is no significant difference on
1 Standard deviations are given in parentheses.

the pre-test scores between the less able and more able students in the control and the
experimental groups. The results on the post-test for the more able students are not
significantly different. However, the scores on the post-test for the less able students
in the experimental group are significantly better than the results of the less able part
of the control group (t=1.43, p<0.1). This result suggests that the less able students
benefit more from open student models.
The logs also contain the data relevant to our hypothesis. Out of 45 logs we
analysed, 12 students did not abandon any problems. These students interacted with
the system for a shorter time (105 minutes on average) and solved 16 problems.
However, they solved all the problems they attempted. The remaining 33 students
abandoned some problems. Whenever a student asked for a new problem before
completing the current one, the system asked the two questions. Each student was
asked these questions at least once, and at most 40 times, and the means for the four
subgroups are given in Table 4 in the Questions row. Out of the total of 2422
abandoned problems, 93 (38.4%) were from the more able students, and 149 were
from the less able students. Therefore, less able students are much more likely to
abandon a problem. The more able students from the experimental group have
abandoned significantly less problems than the more able students from the control
group. However, there is no significant difference between the two groups of less able
students. This might suggest that the more able students were encouraged to complete
the problems by being exposed to their student models.
Table 4. Statistics for the groups of students with different prior knowledge
Control Experimental
Less able More able Less able More able
Pre-test 3.6 (0.5) 5.6 (0.9) 3.7 (0.5) 5.6 (0.9)
Post-test 5.9 (1.5) 7.2 (0.8) 7 (1.5) 6.2 (1.6)
Time 164 (135) 321 (229) 169 (142) 115 (107)
Problems solved 19.8 (10.4) 28.8 (17.2) 27.4 (26.2) 19.4 (14.4)
Questions 7.1 (4.1) 10.87 (13.2) 8.5 (5) 3.9 (3)
0 attempts 5.1 (2.9) 7.6 (11.2) 4.7 (2.8) 2.1 (1.6)
0/new problem 1.6 (1.1) 2.2 (3.6) 1.1 (0.7) 1.5 (1.5)
The number of attempts before abandoning the problem ranged from 0 (the total
of 165 cases) to 13 (an average of 1.43). The 0 attempts row in Table 4 gives the
mean number of cases when students abandon the current problem without attempting
it. The numbers reported here are a bit misleading, as they include all problems, not
just the new problems. In other words, the student might attempt a problem, and then
abandon it, only to receive the same problem again, and then abandon it for the
second time. Such a problems is counted twice. The following row (0/new problem)
counts each problem only once in such situations. It can be seen that the more able
students in the experimental group attempt to solve problems more often than the
2 Although we use the word “problem” here, we refer to situations when the student abandons
the current problem, and get the two questions. It may be the case that the student abandons
the same problem several times, so the total number reported here does not equal to the
number of distinct problems abandoned.
more able students from the control group; however, the differences are not
significant.
The distribution of answers to the first question is given in Fig. 2. The control
group students thought that the problem is too easy more often than experimental
group students, especially the more able ones, although the inspection of the logs very
often contradicts the reason they specified.
Figure 3 shows
the distribution of
0.5
answers to the
0.4 second question.
0.3 As we hypothesiz-
ed, less able
0.2
students are not
0.1 good at identifying
0 the kind of
Too easy Too hard Diff type No answer problem to work
next, and therefore
Control - less Control - more Exper. - less Exper. - more
they specify any
Fig. 2. The distribution of answers for question 1 clause most often
(in 54.69% of the
cases in the control
group, and 40.98% in the experimental group).
After answering the second question, the students in the experimental group were
shown the summary of their student models if their selection of a clause to work on
next differed from the system’s selection. The number of such cases for the two
subgroups are given in the No of cases column in Table 5. In those situations, the
students were asked to specify whether they wanted to go on with their selection, or
adopt the system’s suggestion. The following column gives the percentages of the
total number of questions when the student’s selection differed from the system’s one.
Although the more able students have opinions about what they should be doing that
more often varies from the system’s suggestion, compared to the less able students,
the difference is insignificant. The percentages of cases when the student accepts the
system’s suggestion (last column in Table 5) is almost identical for the more and less
able students. When comparing the experimental group with the control group, there
is no significant difference in the percentages of cases when more able student’s
selections differ from system’s selections.
Table 5. The statistics for the two subgroups of the experimental group
Group No of cases % of questions Student agrees % agree
Exper. - more able 2 (2.69) 57.25 (42.57) 0.75 (0.70) 51.56 (51.94)
Exper. - less able 3.70(3.40) 35.78 (31.70) 1.80 (1.64) 57.5 (42.94)
Control – more able 8.25 (11.37) 42.20 (28.20) N/A N/A
Control – less able 5.44 (4.44) 41.20(20.62) N/A N/A
35
30
25
20
15
10
5
0
SELECT FROM WHERE GROUP HAVING ORDER ANY No answer
BY BY
Control - less Control - more Exper - less Exper - more
Fig. 3. The distribution of answers for Question 2
6 Conclusions
We presented a study of the effects of open student models on student self-assessment

skills. Although we have not seen any significant difference in the post-test scores of
the control and the experimental group, the less able students from the experimental
group have scored significantly higher than the less able students from the control
group. The more able students who had access to their models abandoned
significantly less problems than the more able students from the control group, and
had stronger opinions on what they should work on next, which often varied from the
system’s suggestions.
Overall, these results suggest that the open model may have improved the
performance of the less able students, and that it may have boosted the self-
confidence of the more able students, such that they abandoned fewer problems and
judged their own abilities more readily. Alternatively, more able students in the
experimental group may have specified a clause more often in order to view the
model, which would suggest that they judged it to be beneficial. Such results are
encouraging given the simplicity of the open student model developed for this study.
Our students only received a high-level overview of their knowledge, and they could
not challenge it or make any modifications. We believe that a more sophisticated open
student model would be more effective, and plan to conduct a new study along these
lines in the future.
Acknowledgements
The work presented here was supported by the University of Canterbury grant U6430.
References
1. Aleven, V., Koedinger, K. (2000) Limitations of Student Control: Do Students Know

When They Need Help? Proc. ITS’2000, Springer-Verlag, pp. 292-303.
2. Aleven, V., Koedinger, K., Cross, K. (1999) Tutoring Answer Explanation Fosters
Learning with Understanding. In: Lajoie, S.P., Vivet, M. (eds): Proc. AIED’99, pp. 199-
206.
3. Bielaczyc, K., Pirolli, P., Brown, A.L.: Training in Self-Explanation and Self-Regulation
Strategies: Investigating the Effects of Knowledge Acquisition Activities on Problem-
solving. Cognition and Instruction, 13(2) (1993) 221-252.
4. Brusilovsky, P., Schwarz, E., Weber, G. (1996) ELM-ART: an Intelligent Tutoring
System on World Wide Web. Proc. ITS’96, pp. 261-269.
5. Bull, S., Brna, P. (1997) What does Susan Know that Paul Doesn't? (and vice versa):
Contributing to each other's student model. Proc. AIED 97, pp. 568-570.
6. Chi, M. T. H., Bassok, M., Lewis, M.W., Reinmann, P., Glaser, R. (1989) Self-
Explanations: How Students Study and Use Examples in Learning to Solve Problems.
Cognitive Science, 13, 145-182.
7. Conati, C., VanLehn, K.: Further Results from the Evaluation of an Intelligent Computer
Tutor to Coach Self-Explanation. Proc. ITS’2000, Springer-Verlag, (2000) 304-313.
8. Dimitrova, V., Self, J., Brna, P. (2001) Applying Interactive Open Learner Models to
Learning Technical Terminology. Proc. UM 2001, Springer, pp. 148-157.
9. Kay, J. (1995) The UM toolkit for cooperative user modelling. User Modelling and User-
Adapted Interaction, 4, 149-196.
10. Mitrovic, A. Investigating students’ self-assessment skills. Proc. UM-2001, Sonthofen,
July 2001, Springer-Verlag LNAI 2109, pp. 247-250.
11. Mitrovic, A., Ohlsson, S.: Evaluation of a Constraint-based Tutor for a Database
Language. Int. J. on Artificial Intelligence in Education, 10(3-4), (1999) 238-256.
12. Morales, R., Pain, H., Conlon, T. (2001) Effects of Inspecting Learner Models on
Learners’ Abilities. Proc. AIED 2001, pp. 434-445.
13. Ohlsson, S.: Constraint-based student modeling. In: Greer, J.E., McCalla, G (eds): Student
modeling: the key to individualized knowledge-based instruction, (1994) 167-189.
14. Paiva, A., Self, J. (1995) TAGUS – a user and learner modelling workbench. User
Modeling and User-Adapted Interaction, 4, 197-226.
15. Swanson, H.L.: Influence of Metacognitive Knowledge and Aptitude on Problem Solving.
J. Educational Psychology, 82 (1990) 306-314.
16. White, B.Y., Shimoda, T. A, Frederiksen, J.R. (1999) Enabling Students to Construct
Theories of Collaborative Inquiry and Reflective Learning: Computer Support for
Metacognitive Development. Int. Journal on AI in Education, vol 10, 151-182.
17. Zapata-Rivera, J.D., Greer, J.E. (2000) Inspecting and Visualizing Distributed Bayesian
Student Models. Proc. ITS'2000, Springer, pp. 544-553.
Ephemeral and Persistent Personalization in Adaptive
Information Access to Scholarly Publications on the Web
Stefano Mizzaro and Carlo Tasso
Artificial Intelligence Laboratory

Department of Mathematics and Computer Science
University of Udine
{mizzaro, tasso}@dimi.uniud.it
Abstract. We show how personalization techniques can be exploited to

implement more adaptive and effective information access systems in electronic
publishing. We distinguish persistent (or long term) and ephemeral (or short
term) personalization, and we describe how both of them can be profitably
applied in information filtering and retrieval systems used, via a specialized
Web portal, by physicists in their daily job. By means of several experimental
results, we demonstrate that persistent personalization is needed and useful for
information filtering systems, and ephemeral personalization leads to more
effective and usable information retrieval systems.
1 Introduction
Oversupply of information constitutes a well known phenomenon that is

progressively becoming worse and that threatens Web usefulness. Other related issues
are information waste (documents published on the Web do not always reach the
appropriate readers, or reach them too late), and low information quality (the amount
of available information is increasing, but its quality is decreasing). These problems
are very general, and affect all kinds of Web contents, i.e., information to be accessed,
products to be purchased, and services to be exploited. Personalization allows a more
effective information access: end users can be delivered personalized information,
tailored to their individual needs, and, more generally, it enables a more effective and
efficient transfer of the published information from the authors to the most
appropriate readers.
At the University of Udine, we have been investigating the issue of personalization
in information access for several years [1, 8, 9, 10, 20, 22]. In this paper, we present
the most recent results concerned with the application of adaptive and personalized
information access to the electronic publishing field, and more specifically in
scholarly publishing. We claim that personalization is needed and useful in
information access, and especially in scholarly publishing, where users (i.e.,
researchers) are interested in it for two important reasons: (i) detecting newly
published information which is relevant to their interests and preferences, and (ii)
accessing stored information for satisfying specific information needs. However, this
twofold situation requires a novel approach, in which two distinct and complementary

Ephemeral and Persistent Personalization in Adaptive Information Access 307
personalization techniques (i.e., ephemeral and persistent personalization) are applied

together to meet user’s requirements.
This paper is structured as follows. Section 2 briefly describes the world of
scholarly publishing, and the heavy changes introduced in it by the Web. In Sect. 3
we present a short overview of information access approaches, and discuss how
personalization techniques can be useful in this field. In Sects 4 and 5 we show the
application of personalization techniques to information access to scholarly
publications. Section 6 closes the paper and sketches some future developments.
2 Scholarly Publishing and the Web

The communication mechanism adopted by science today arose in the 17th Century,
with the publication of the first scientific journals. Since about 1930, the
dissemination of scholarly information is based on peer review, that usually assures a
high quality of the published papers. Internet has changed, and is changing, this
situation [3]. Now a peer reviewed journal can be distributed by electronic means, and
the peer reviewing can take place completely electronically, drastically reducing time
and money for publishing (see, e.g., JHEP at jhep.sissa.it or Earth Interactions at
EarthInteractions.org). Many publishers now allow their subscribers to electronically
access the full text of the papers published on standard journals. Beyond modifying
the standard scholarly journals and proceedings, the Web has also introduced a new
way of disseminating scholarly knowledge: e-prints, i.e., open online repositories of
scholarly papers (see, e.g., arXiv.org, mainly about physics, or cogprints.soton.ac.uk,
about disciplines concerning cognition). The repositories usually contain preprints,
i.e., electronic versions of submitted papers made publicly available before review,
acceptance, and, possibly, publication.
The exploitation of the Web has also highlighted another essential characteristic of
scholarly publications, i.e., their hypermedia nature. A rich hypertext structure is
provided by both the citations across publications and the references (to chapters,
sections, figures, and so on) within each publication. Multimediality is also important
since it leads to a more effective communication, and even though still limited today,
it will increase in the next years. Another aspect that further extends the richness of
the hypermedia structure is the storage (easy obtainable on electronic media) of a
publication as a multilayered document (dlp.cs.berkeley.edu), that includes the
various versions of a document, the slides and presentations resulting from it, the
referees’ and readers’ comments, and any other remark about the document. This
provides a richer information on the topic at hand, and adds more hypermedia
information as links among the various layers.
All these new means available to authors allow an ever growing rate of production
of scholarly articles (see, e.g., the arXiv usage statistics: arXiv.org/show_stats). The
new Web-based approach guarantees easier access, more powerful and richer means
of information seeking, and better timeliness, but it features also some drawbacks
(e.g., the quality of preprints is not assessed by a peer review process) and poses some
new problems (copyright problems, social and legal acceptance, and so on). However,
after a slow start [15], the impact of scholarly publishing is steadily increasing [11].
As a result, the scholar is nowadays overloaded by a large amount of highly
structured hypermedia information, in the form of scholarly publications, online
repositories, commentaries, and so on. In this scenario, it is important to allow the
308 S. Mizzaro and C. Tasso
scholar: (i) to stay up-to-date, being notified when new information on some topics of
interest is published, and (ii) to quickly and easily find, on demand, information on
specific topics. Both goals can be approached by advanced personalization
techniques, as shown in the next section. Personalization plays indeed a fundamental
role not only for the highly subjective nature of the information seeking process, but
also because the job of a researcher is highly innovative, it does not conform to any
standard behavior, and it is therefore quite different for each researcher.
3 Personalized Information Access

Information access is the process exploited by a seeker who wishes to find and
retrieve some data/information which satisfies an information need. It is common to
distinguish between two kinds of information access:
• Information retrieval (IR) [2] is characterized by a static database of documents, a
short term information need, and a query made up by a few (usually less than two)
terms. Web search engines are the most known instance of IR systems. It is well
known that IR is a difficult task [4], since users have to specify in a query some-
thing that they do not know (if they knew it, they wouldn’t be searching for it).
• Information filtering (IF) [14], on the other side, is characterized by a dynamic
database (actually an incoming flow of documents) and a long term and rather
static information need. Users of IF systems are more motivated to express their
information needs as more accurate and complete descriptions that will last for
longer time. These descriptions are usually called user profiles (or models), and are
made up by a lot of data: concepts, relationships among them, weights, etc.
Web personalization is the process of selecting, preparing and delivering Web
contents for a given user, by taking into account his specific needs and preferences
[23]. Personalization means delivering to the user the most relevant contents, in the
most adequate way, and at the most appropriate time. A personalization system is
based on three main functions, which all can be performed in a personalized way:
selection, visualization, and delivery. In order to be personalized, all the three
functions have to be supported by specific information about the user, which is
included in a user profile and has to be available when the personalization process
takes place. In this paper we deal only with the first of the three functions, i.e.,
selection of the most appropriate content.
Personalization techniques are very numerous and are ranging from simple user-
controlled customization of Web content, to autonomous system-controlled adaptation
[17, Reader’s Guide, p.6]. We distinguish two types of personalization [23]: persistent
(or long term), i.e., based on a user profile which lasts over time and is stored in a
persistent information structure; and ephemeral (or short term), which is not based on
a persistent user profile. The main differences are the temporal features of the process
aimed at building and managing the user profile. In persistent personalization, the
user profile is incrementally developed over time and at the end of each session it is
stored in order to be used later on in subsequent sessions. Usually, but not necessarily,
the information exploited for building the profile comes from various sources, it
concerns different aspects of the user, and it is often extended by means of (possibly
sophisticated) reasoning or learning processes. In ephemeral personalization, the
information used to build the user profile is gathered during the current session only,
and is immediately exploited for executing some adaptive process aimed at
personalizing the current interaction. At the end of each session, the user profile is
lost, and no information about the user is stored in a persistent way for later use.
Information access systems should and can exploit both kinds of personalization
[4, 12]. We propose here a twofold approach. On the one side, personalization in IF
means capturing the long term information interests and preferences of the user, in
order to tailor the selection process to the specific personal characteristics. On the
other side, in IR persistent personalization is not feasible, since in that context infor-
mation needs have a short term nature and are different, for the same user, in the dif-
ferent sessions. However, ephemeral personalization can be used in an effective way,
with the goal of modeling the search session, rather than the information need, for
immediately providing personalized support during the searching session. The idea of
long and short term modeling in information access is not new (see, e.g., [7]), how-
ever it has been considered from the IF perspective only, i.e., it consists in building
user profiles across a shorter or longer period of time (a limited number of sessions or
very many sessions), and the profiles, in both cases, model only the topics interesting
for the user. Our approach is innovative for two reasons: (i) short term modeling is
performed through ephemeral personalization, restricting the scope of observation to
the current session only, and (ii) we do not build a model of the information need
(difficult, if not impossible, during just one session), but rather a session model. This
novelty allows to provide adaptive support to the user, as it will be shown in Sect. 5.2.
We have experimented this twofold approach in scholarly publishing portals for
physics. We chose that community since the physics (especially high energy physics)
field seems well ahead in exploiting the full potential of web publishing (no surprise,
since the Web was born at CERN, one of the major physics institutions worldwide):
the above cited arXiv repository (formerly known as xxx) is already a used, valid, and
widely accepted media for physics and astronomy fields [11], and the SPIRES
(www.slac.stanford.edu/spires/hep) citation index is almost three times more
complete than the ISI well known database. In the next sections we present an
application of persistent and ephemeral personalization within the Torii vertical portal
(torii.sissa.it) on physics, which has been developed in the 5th FP IST project TIPS
(Tools for Innovative Publishing in Science), see tips.sissa.it.
4 Persistent Personalization in Information Filtering

In previous work, we have developed and evaluated several content-based filters [18]
for persistent personalization. Among them [1, 20, 22], the most effective has been
the information agent ifT (information filtering Tool) [20], which is based on the user
modeling shell UMT (User Modeling Tool) [10]. The work presented here concerns
the exploitation of ifT in the Torii portal.
4.1 Content-Based Filtering through ifT

ifT exploits lightweight natural language processing and co-occurence-based semantic
networks for building long term user profiles and for evaluating the relevance of text
documents with respect to a profile. The main mechanism for building user profiles
exploits explicit relevance feedback provided by the user on both positive and nega-
tive examples. The learning capabilities of this mechanism have been evaluated by
means of several laboratory experiments [1]. In one of them, four subjects received
2000 documents (20 each day, for 100 days) on various computer science topics. Each
subject was interested in some specific area(s) of computer science, and ifT was fil-
tering and ranking the incoming documents according to their relevance. Initially the
user profile was empty, and the user was allowed to ‘explain’ his interests through
relevance feedback only. Throughout the experiment, standard precision and recall
were measured. Fig. 1 shows the evolution of precision over time (100 sessions): dots
represent the observed data, the irregular line represents the moving-average of order
5, and the regular line is an interpolation curve. The results show good learning
capabilities (a precision of 80% is reached after 8 sessions), as well as a very high
final precision value which saturates at about 92% in the interpolation model.
Another significant application of ifT has been developed within the ifWeb
system for filtering Web documents [22]. The system includes the information agent
ifSpider, aimed at the autonomous navigation of the Web for searching documents
relevant to a specific user profile. The navigation performed through hyperlinks is
opportunistic: only the paths including documents which feature relevance scores
above a given threshold are considered.
ifWeb has been evaluated in several laboratory experiments. In one of them,
devoted to assess its ranking capabilities, each subject was initially defining a
profile through relevance feedback given on 4-6 documents, and then he was
performing a series of nine sessions with ifWeb. After each session, the subject was
requested to provide the correct ranking of the documents given by ifWeb, and
human and system rankings were compared. Fig. 2 shows precision (continuous
line) and the ndpm measure [24], which evaluates the difference between the two
rankings (good performance is indicated by decreasing values). After the first
sessions, as precision reaches good values, the ndpm starts to decrease, indicating
the capability to produce a better ranking.
Fig. 1. Precision of ifT over 100 sessions. Fig. 2. Precision and ndpm measures of
ifWeb over 9 sessions.
4.2 Exploiting Content-Based Filtering in Electronic Publishing
Given the performance reached by ifT, we decided to adopt ifT as the filtering engine
of the Torii portal. More specifically, the problem approached with persistent person-
alization has been the high (and currently increasing) rate of incoming documents:
about 100-200 new e-prints are submitted every day and included in arXiv, which is
available in Torii. Normal users (researchers in high energy physics) are used to start
the working day by browsing the long list of new e-prints. By adding a personalized
filtering engine to Torii, each user can define one or more profiles related to his inter-
ests, and all the new incoming information is automatically filtered. In this way, Torii
displays (in the first positions) only the documents which best match user’s interests.
Information overload is then reduced, as well as the cognitive load of analyzing many
documents every day. Fig. 3 shows a snapshot of the Torii portal.
Fig. 3. Torii portal: documents ranked by the ifT filtering engine.
The relevance measure produced by ifT and exploited in Torii is a unique figure
(see the bars shown in Fig. 3) which combines the document topicality value, i.e., a
measure of how much concepts relevant for the user are present in the considered
document, and the conceptual coverage value, i.e., a measure of how many of the
concepts relevant for the user are present in the considered document.
Torii has undergone a first validation phase through field testing in July 2001.
Twenty users were using the system for 29 days. All their sessions have been
monitored and tracking logs of all actions acquired. Final interviews were also
delivered. Cognitive filtering was working well and judged well by the users, who
proposed to extend the system with the possibility to rank any set of documents
(possibly coming as the result of a search in one of the available collections) by
means of ifT. SISSA (the managing institution of Torii), has decided to use the
filtering engine ifT as a standard tool available to all users of their portal.
5 Ephemeral Personalization in Information Retrieval
5.1 Supporting Users of IR Systems
The interactive nature of IR is advocated since years [16] and is now widely accepted:
between the user and the IR system a dialogue takes place [6], during which the user
should receive adequate support [4]. The help should be provided proactively by the
system and suggestions should be given “on the background”, with the user retaining
the control of the interaction [5]. A basic kind of support is terminological help,
which identifies and suggests to the user terms that improve the query [13, 19].
Another kind of support is strategic help, which provides to the user useful hints on
how to improve the strategy adopted for organizing the searching process (see a
survey of this issue in [9]).
We propose to use ephemeral personalization techniques to provide both strategic
and terminological support to IR users. We have been doing research on this issue for
several years. We implemented the FIRE prototype [8] that, by means of thesauri, is
capable of suggesting to the users of a boolean IR system alternative terms to better
(re)formulate their information needs. After a carefully designed laboratory
experiment involving 45 participants, we had evidence that terminological help alone
is useful, but needs to be complemented by strategic help.
On the basis of these results, we added to FIRE a Strategic Aid Module (SAM)
capable of providing to the users suggestions on which strategies are more likely to be
effective in a certain situation [9]. SAM is based on a collaborative (between the user
and the system) view of the session: users know their needs, judge the relevance of
the retrieved documents, select the terms to be added to the query, and retain the
control of the session; the system monitors users’ actions and provides contextual
suggestions, proposing alternative routes, emphasizing mistakes (e.g., term spelling),
and so on. SAM is based on a detailed conceptual model of the session, made up by
representing user actions, the current situation of the session, and the set of feasible
and more appropriate suggestions. By exploiting a knowledge base, the current
situation of the session is inferred from the actions made by the user, and personalized
suggestions are selected on the basis of the current situation. We performed two
laboratory experiments (one in which we simulated the activity of the users of the
previous FIRE experiment, and one that involved six new participants), both of which
showed that strategic support is useful, well accepted, and it allows users to learn the
best strategies.
5.2 Supporting Users of IR Systems on the Web
Following the positive evaluations of the two prototypes mentioned above, we de-
cided to apply ephemeral personalization to an IR system deployed in a real setting:
we implemented the Information Retrieval Assistant (IRA), a system providing vari-
ous kinds of suggestions to users that are searching the paper and e-print database
available in the Torii portal. IRA exhibits some innovative features with respect to the
previous two prototypes. It fully integrates terminological and strategic suggestions.
The underlying IR system is a probabilistic one (Okapi, see web.soi.city.ac.uk/re-
search/cisr/okapi/okapi.html) in place of a boolean one, and it works on an underlying
full text database, containing almost 200,000 scholarly documents about physics (as
opposed to the bibliographic, and smaller, database used in the previous experiments).
IRA is designed to be deployed in a real life environment, and used by physicists in
their daily job. IRA can also be easily tailored to be used with other IR systems.
However, the most important innovative features in IRA are on the conceptual
side, and concern the new models on which ephemeral personalization, i.e., both
terminological and strategic suggestions, is based. The sorted term lists suggested in
terminological help are obtained by a new spreading activation algorithm capable of
browsing heterogeneous, dynamically generated, and integrated thesauri, starting
either from the last inserted search term, or from the set of all the search terms used
by the user so far. This new version of terminological help has shown, by means of an
experimental evaluation involving six participants, significant improvements with
respect to the terminological help previously used in FIRE: more terms are suggested
(since more term sources are used), they are more adequate to the current context and
ranked in a better way (mainly for two reasons: the synergy among the different term
sources and the new spreading activation algorithm).
The enhanced reasoning process for suggestion generation is represented in Fig. 4.
Each user action (i.e., any operation performed by the user, such as term
insertion/removal/modification, search in the database, document reading, relevance
judgment, etc.) on the Okapi user interface is notified to IRA by Okapi. IRA monitors
these time-stamped actions and builds a model of the session history, that is made up
by a sequence of interleaved actions and states. A state is a set of parameters
describing the current state of the system, like number of terms in the query, number
of retrieved, read, and judged (as relevant or not relevant) documents, etc. At each
state, i.e., after each action, a new set of situations is inferred. A situation is a history
pattern, or an abstract description of the session history. Situations can be very
simple, like ‘insertion of a zero posting count term in the query’ (a term that is not
contained in any document), or they can concern a longer time interval, like ‘two
consecutive searches with no changes to the query’. Moreover, they can be more
Fig. 4. IRA reasoning process.

abstract and difficult to infer certainly, like ‘user not reading the content of the
retrieved documents’. Situation derivation is triggered by the last user action, but
takes into account the whole session history.
From each situation, a set of suggestions is derived. One of the most important
suggestions is terminological help, but IRA suggestions also include simple hints, that
merely make aware the user of alternative actions (like reminding the user to have a
look at the full text of the documents, or to judge, by clicking on the appropriate
button, the relevance of the read documents), and more complex advices, that are
carried out collaboratively by the user and IRA (like author search, that suggests to
look for documents written by the same author as the documents already judged
relevant by the user). IRA suggestions are always contextual and are provided in two
kinds of situations: critical (i.e., the user is experiencing some problem, as repeatedly
retrieving no documents, or not making progress) and enhanceable (i.e., when the
user could follow other – possibly more – appropriate alternative routes). Finally, IRA
suggestions are ranked and proposed to the user as two types of textual descriptions, a
short one and a longer one displayed on demand. They are shown in IRA own
window on the background, thus allowing the user to maintain the control of the
interaction with the IR system. The user can either accept the suggestions received
(e.g., he can insert into the query some new relevant term provided by terminological
help), or can ignore them. IRA knowledge bases by now contain 33 actions, 28
situations, and 20 suggestions, and are still being extended.
We performed a first laboratory evaluation that highlighted some positive
qualitative results: the sample users that used IRA are satisfied with the adequacy,
timeliness, comprehensibility, and usefulness of the suggestions. Moreover, as
foreseen, terminological help is especially appreciated. IRA is now being deployed
and used by real end-users, and another more extensive evaluation of it will take place
in the next months.
In this paper we have shown how persistent and ephemeral personalization techniques
can be exploited to implement more adaptive and effective information access sys-
tems. More specifically, the research presented here approaches two problems of the
user of a scholarly publishing system: the need to be timely and accurately updated
about new relevant information and the request for adequate, effective and easy-to-
use support during search of archive information. Several experimental results show
that persistent personalization is useful for information filtering systems, and ephem-
eral personalization leads to more effective and usable information retrieval systems.
So far, we have kept separated the two approaches, but they naturally complement
each other. Therefore, we plan to integrate them in various ways: the long term user
profile can be used in IR, e.g., to rank the retrieved documents in a more personalized
way; vice versa, the suggestions can be useful during the initial construction of the
profile, or during feedback iterations. We also believe that these personalization
techniques can be fruitfully applied also outside the scholar community, for instance
in the more general context of electronic publishing, where various media such as
newspapers, magazines, news agencies, and so on are continuously fed with new
information. Finally, the quality of information is another important issue, that we
have not considered in this paper, and that we are approaching with a collaborative
work approach [21].
References
1. F.A. Asnicar, M. Di Fant, C. Tasso, User Model-Based Information Filtering, in M.

Lenzerini ed. AI*IA 97: Advances in Artificial Intelligence – Proc. of the 5th Congress of
AI*IA, LNAI 1321, Springer, Berlin, D, 1997, 242-253.
2. R. Baeza-Yates, B. Ribeiro-Neto, Modern Information Retrieval, Addison-Wesley, New
York, NY, USA, 1999.
3. C.W. Bailey, Jr., Scholarly electronic publishing bibliography - Version 40: 12/12/01,
http://info.lib.uh.edu/sepb/sepb.html, visited 5/1/02.
4. N.J. Belkin, Helping People Find What They Don’t Know, Comm. of the ACM 43(8),
2000, 59-61.
5. N. Belkin, C. Cool, D. Kelly, S.-J. Lin, S.Y. Park, J. Perez-Carballo, C. Sikora, Iterative
exploration, design and evaluation of support for query reformulation in interactive
information retrieval, Information Processing and Management 37(3), 2001, 403-434.
6. N. Belkin, C. Cool, A. Stein, U. Thiel, Cases, scripts, and information-seeking strategies:
On the design of interactive information retrieval systems, Expert Systems with
Applications 9(3), 1995, 379-395.
7. D. Billsus, M.J. Pazzani, User Modeling for Adaptive News Access, User Modeling and
User-Adapted Interaction Journal 10(2-3), 2000, 147-180.
8. G. Brajnik, S. Mizzaro, C. Tasso, Evaluating User Interfaces to Information Retrieval
Systems: a Case Study on User Support, Proc. of the 19th Annual International ACM
SIGIR Conference, Zurich, CH, 1996, 128-136.
9. G. Brajnik, S. Mizzaro, C. Tasso, F. Venuti. Strategic help in user interfaces for
information retrieval, J. of the Am. Soc. for Information Science and Technology, 2002, in
press.
10. G. Brajnik, C. Tasso, A shell for developing non-monotonic user modeling systems,
International Journal Human-Computer Studies 40, 1994, 31-62.
11. C. Brown, The E-volution of Preprints in the Scholarly Communication of Physicists and
Astronomers, J. of the Am. Soc. for Information Science and Technology 52(3), 2001, 187-
200.
12. W.B. Croft, S. Cronen-Townsend, V. Lavrenko, Relevance Feedback and Personalization:
A Language Modeling Perspective, DELOS Workshop: Personalisation and
Recommender Systems in Digital Libraries, 2001, www.ercim.org/publication/ws-
proceedings/DelNoe02/.
13. E.N. Efthimiadis, Query expansion, Annual Review of Information Science and
Technology (ARIST), M. E. Williams ed., vol. 31, 1996, 121-187.
14. U. Hanani, B. Shapira, P. Shoval, Information Filtering: Overview of Issues, Research and
Systems, User Modeling and User-Adapted Interaction 11(3), 2001, 203-259.
15. S.P. Harter, Scholarly communication and electronic journals: An impact study. J. of the
Am. Soc. for Information Science 1998, 49(6), 507-516.
16. P. Ingwersen, Information Retrieval Interaction, Taylor Graham, London, UK, 1992.
17. A. Jameson, C. Paris, C. Tasso eds., User Modeling – Proc. of the 6th Intl. Conference
UM97, Springer-Verlag, Wien New York, 1997.
18. T. Malone, K. Grant, F. Turbak, S. Brobst, M. Cohen, Intelligent information sharing
systems, Comm. of the ACM 43(8), 1987, 390-402.
19. R. Mandala, T. Tokunaga, H. Tanaka, Query expansion using heterogeneous thesauri,
Information Processing & Management 36, 2000, 361-378.
20. M. Minio, C. Tasso, User Modeling for Information Filtering on Internet Services:
Exploiting an Extended Version of the UMT Shell, UM96 Workshop on User Modeling
for Information Filtering on the World Wide WEB, Kailua-Kona, Hawaii, USA, January
1996.
21. S. Mizzaro & P. Zandegiacomo Riziò. An automatically refereed scholarly electronic
journal: Formal specifications. Informatica - An International Journal of Computing and
Informatics 24(4), 2000, 431-438.
22. C. Tasso, M. Armellini, Exploiting User Modeling Techniques in Integrated Information
Services: The TECHFINDER System, in E. Lamma and P. Mello eds., Proc. of the 6th
Congress of the Italian Association for Artificial Intelligence, Pitagora Editrice, Bologna,
I, 1999, 519-522.
23. C. Tasso, P. Omero, La personalizzazione dei contenuti Web: e-commerce, i-access, e-
government, Franco Angeli, Milano, I, 2002.
24. Y.Y. Yao, Measuring retrieval effectiveness based on user preference of documents, J. of
the Am. Soc. for Information Science 46(2), 1995, 133-145.
Fuzzy Linguistic Summaries
in Rule-Based Adaptive Hypermedia Systems
1 1 1 2
Miguel-Ángel Sicilia , Paloma Díaz , Ignacio Aedo , and Elena García
1
Laboratorio DEI, Computer Science Department, Universidad Carlos III de Madrid
Av. Universidad 30, 28911 Leganés (Madrid), Spain
{msicilia, pdp}@inf.uc3m.es, [email protected]
http://www.dei.inf.uc3m.es
2
Computer Science Department, Universidad de Alcalá de Henares
Ctra. Barcelona km. 33.600, 28871 Alcalá de Henares (Madrid), Spain
[email protected]
Abstract. Rule-based adaptive hypermedia systems personalize the structure of

the hypermedia space using an inference mechanism that operates on a specific
knowledge representation about its users. Approximate quantifiers are very fre-
quently used in human language expressions that entail the summarization of a
large number of facts. We describe how quantified expressions can be used in
adaptation rules to specify common adaptation behaviors, enhancing rule’s ex-
pressive power for the human expert. Those quantified expressions can be im-
plemented through fuzzy quantification mechanisms operating on fuzzy lin-
guistic labels and relations, and can be integrated as extensions in general-
purpose rule-based adaptive hypermedia systems.
1 Introduction
Adaptive hypermedia systems (AHS) personalize the information, links and naviga-
tion features of the hypermedia space by using knowledge about its users, represented
in a user model. Several adaptation techniques can be used for this goal, which are in
turn abstracted in adaptation methods at a conceptual level (as defined in [2] and
extended in [3]). In addition, a particular technique can adapt different aspects of the
hypermedia structure (aspects that are called adaptation technologies). More specifi-
cally, rule-based AHS use an inference mechanism to implement adaptation and/or
user modeling behaviors, resting on some kind of knowledge representation model.
These systems can be considered as general-purpose ones when they are defined on a
conceptual AH architecture as in [22]. In many cases, adaptation rules are defined by
domain experts (e.g. marketing directors in Web recommendation systems or teachers
in educational AHS), in a process of knowledge acquisition that results in a reusable
and easily modifiable knowledge base about the intended adaptation behavior of the
system. Indeed this approach is taken in several Web personalization engines [6]).

318 M.-Á. Sicilia et al.
In this paper, we describe a general-purpose syntax and execution semantics for using
fuzzy quantifiers in rule-based AHS. Although different theories of uncertainty repre-
sentation have been applied in user modeling [11, 17], quantified expressions have
not been applied as a general-purpose rule-modeling construct, even though their
closeness to human language expression suggests that they could significantly en-
hance the rule definition process. The use of quantifiers in adaptation rules can be
helpful for different adaptation technologies, and can be used also for the task of user
modeling. In [19], an application of fuzzy linguistic quantified expressions is de-
scribed for the specific user-modeling task of classifying users of a Web application
in vague categories (fuzzy stereotypes), based on their navigation history. In this
work, we focus on the description of adaptation behaviors using quantified expres-
sions.
The rest of this paper is structured as follows. Section 2 describes fuzzy quantified
expressions in rules and how they can be applied to extended-for-fuzziness AH mod-
els. Examples of adaptation methods that can be implemented using them are de-
scribed in Sect. 3, along with the most relevant implementation details. Finally, con-
clusions and future research directions are provided in Sect. 4.
2 Fuzzy Quantified Rules in Adaptive Hypermedia
2.1 Quantified Expressions and Rule Formulation
Approximate quantifiers like ‘almost all’ or ‘many’ are very frequently used in hu-
man language, and serve the important purpose of abstracting from details, and sum-
marizing a large number of evidences into a global view [12]. A quantified linguistic
expression can be specified as an expression in the form “Q X” in the case of absolute
quantification (for example, ‘quite a few visits’) or in the form “Q X are Y” in the
case of relative – or proportional – quantification (for example, ‘most visits are
short’). Both X and Y are natural language nouns or phrases, which, for our purposes,
must refer to user or domain model information that can be crisp or vague, and that
must be interpreted in the context of a specific AH schema. Those expressions can be
integrated in the antecedent of a rule, in the general form “if Q X then <<action>>”,
where action stands for the activation of a specific adaptation technology, or the addi-
tion of an inferred fact to the user model.
Fuzzy quantified expressions can be added to existing adaptation rule syntax to en-
rich the expressiveness of the rule language. An experiment was carried out to gather
some evidence about the appropriateness of including quantifiers in the syntax of the
adaptation rules. We focused on a specific and basic adaptation technology, namely a
direct guidance, a kind of adaptive navigation support [2]. Both computer specialists
and non-technical experts were included in the population (although experience in
using the hypermedia technology we’re analyzing – the Web – was considered man-
datory). In consequence, we partitioned the sample according to two different user
profiles: technicians and non-technicians. The objective of the study was to find how
frequently natural language quantifiers were used in the formulation of simple adapta-
tion rules. A scenario was developed in which the main page of a research group’s
Web site was described. The site had a navigation bar in a left frame that guided the
user to different sections of the site. Participants were asked to write down ‘if/then’
rules in natural language to perform direct guidance to sections to which the user
might be interested in, based exclusively in previous navigation history. Results
showed that eighty three percent of the thirty participants used a quantifier (as ‘most’,
‘many’ and ‘often’) to describe the rule (either referring to frequency of visits, ses-
sions or interactions in general), with no significant difference between the two user
profiles. Although this is only a small experiment, it reveals the importance of quanti-
fiers in human expression – a review of available evidence is in [16].
2.2 Fuzzy Linguistic Quantifiers
Zadeh’s sigma-count operator [24] is the earliest operational definition of the concept
of linguistic quantifier in terms of the theory of fuzzy sets and possibility theory.
Although this and other fuzzy linguistic definitions of quantifiers do not achieve
perfect linguistic adequacy (see, for example, [7]), a number of application exists that
prove their usefulness in modeling natural language constructs. The approach pre-
sented here uses sigma-count and OWA operators, but other quantification mecha-
nisms can be used either, with no impact in the conceptual semantics of the adaptation
methods.
Most existing approaches define fuzzy linguistic quantifiers as fuzzy subsets of the
non-negative real numbers (absolute quantifiers) or of the unit interval (proportional
or relative quantifiers), and therefore can be considered as fuzzy numbers. In conse-
quence, a quantifier Q is represented either as a mapping µQ:R →I or µQ:I→I respec-
+
tively, where I is the unit interval. From an information modeling viewpoint, an es-
sential difference exists between absolute and relative quantifiers, since the former
are entity-dependent, that is, their definition is tied to the subject on which they’re
applied (we can consider that four are definitely many cars for a buyer but four is not
many if we’re expressing a football team’s supporters number). In addition, it can be
instance-dependant in the following sense: if we’re measuring the user’s visits in a
Web site, many visits depend on the particular site we’re considering. As a conse-
quence, we can use the notation µ Qx to denote quantifier Q applied to subject x.
Example definitions as S-functions of the absolute quantifier ‘many’ (applied to
visits) and the relative one ‘most’ are provided in Fig. 1, defined as Zadeh’s S-
functions S(x;10;55;100) and S(x;0.6;0.75;0.9) respectively.
Given a domain D of elements (in our case, elements in the user or domain model,
for example, the set of users or nodes in our AH system), a unary quantification
mechanism in the form:
M : X ∈℘( D) → R +
where ℘(D) is the fuzzy power set of D, provides a way of representing a summary
of the cardinality of a fuzzy subset of elements in that domain through a real number.
This numeric summary can then be matched to a quantifier Q to obtain the degree of
conformance of the quantified expression, thus achieving the transformation:
Q : X ∈℘( D ) → I
The definition can be extended to n-ary quantifiers, but we’re only concerned with
unary (absolute) and binary (relative) quantifiers. Zadeh’s sigma-count is one of those
quantification mechanisms, with the following absolute and relative formulations:
sigma − count ( X ) = ∑
e∈ D
µ X (e)
sigma − count ( X 2 / X 1 ) = sigma − count ( X 2 ∩ X 1 ) sigma − count ( X 1 )
In the relative formulation, X1 is the subset on which the proportion is computed.

Additional details about the computation of linguistic summaries can be found, for
example, in [18].
Fig. 1. Example quantifiers: many and most
The other quantification mechanism we have used is based in Yager’s proposal

based on the Ordered Weighted Averaging (OWA) operator [23], only for propor-
tional ones (we do not describe it here in detail, since it does not introduce changes in
rule formulation).
2.3 Fuzzy Quantified Rules in an AHS
Adaptation techniques are characterized by a specific kind of knowledge representa-

tion and a specific adaptation algorithm. In this subsection, we describe a minimal
abstract model that supports fuzzy linguistic labels and common fuzzy inference
techniques, which can be implemented with a generalized fuzzy relation model
framework (see, for example, [15]). This partial model could be combined with those
from generic user models [13], or general purpose AHS like [22].
We’ll assume a common division of our AH model in three components: the User
Model (UM), the Domain Model (DM) and the Adaptation Model (AM). We also
assume that an object model exists defining the entities in the three components in
terms of classes, associations and other common object-oriented modeling constructs.
The three model’s components can be described with attributes in a relation as de-
scribed in [22], but we describe here a more abstract rule syntax, that could easily be
mapped onto these more specific schemas. In addition, some specific fuzzy types are
introduced as extensions in the model, namely:
• Independent terms, for example, frequent, that is defined by an associated
membership function µfrequent. Each term is defined on a domain that, in our
case, could be groups of users or some entity belonging to the domain
model.
• Linguistic labels picked from predefined label sets. For example (bargain,
cheap, somewhat cheap, medium-priced, somewhat expensive, expensive,
unaffordable) as price description. This second model construct allow us to
store only references to the specific label in a label set.
Label sets fulfil some properties: (1) they must form totally ordered sets with odd
cardinality Tg+ 1, in the form LSg = {si}, i ∈ {0, ..., Tg} such that the set is ordered: si
>= sj if i >= j; (2) it exists a negation operator: Neg (si) = sj , such that j = Tg – i, and
(3) it exists maximization and minimization operators: Max (si, sj ) = si if si >= sj, and
Min (si, sj ) = si if si <= sj. Label sets are specially well suited to model and aggregate
user’s preferences, since they’ve been extensively used in processes of group decision
making (see for example [9]), and can be also used in content rating [5].
The set of users of our UM is denoted by U={ui}. Each user has a set of attributes
each of them with an associated domain (including label sets), and can be included in
one or more sets that represent associated information about the user. In the formula-
tion of adaptive rules, with the general syntax C→A, the set of antecedents (C) ex-
press conditions on the user model and the context in which the adaptation is per-
formed, and actions (A) are updates on the UM (user modeling rules) or dynamic
adaptation behaviors (adaptation rules). We describe here only the syntax of antece-
dents, in which the quantifiers are inserted. An antecedent can have the following
formulations:
instance.attribute op expression
instance op expression
Where valid instances are the user currently connected (ui), the current navigation
context (ci), an entity in the DM, or one of the groups defined by linguistic independ-
ent terms gterm. The context is a placeholder for navigation specific information, in-
cluding session information and characteristics of the DM. Attributes are defined on
entities on a domain, and operators (op) and expressions are defined on that domains
(e.g. character string, number comparison and the like). Association relationships
between entities in the model are considered as a special kind of attribute, that de-
notes the collection of instances associated to the instance. An example of ‘crisp’
antecedent is ui.age>20, and an example of a simple fuzzy expression is ui is
loyal. We’ll focus here only on fuzzy attributes and its applications and on fuzzy
antecedents that are used in standard fuzzy rules like Mamdani min implication [21]
(antecedents in the same rule are implicitly connected by and operators). Expressions
regarding fuzziness include the following:
• Membership of the user in a group: ui includedIn gterm, e.g. ‘includedIn
frequent’ or ‘includedIn advanced’.
• Fuzzy label sets domains, with two forms: ui.att op label, e.g.
‘likesLongNews is high’, and ui.rel aj op label, where rel is a
(fuzzy) relationship between the user and some other entity, from which aj is
an instance v.g. ‘interestIn music is low’.
Quantification can operate on that expressions; Table 1 describes the main alterna-
tive syntaxes and examples (fuzzy subsets are in italics), which can be considered as
extensions of existing data design notations in hypermedia development methods [4]
with imprecise (or vague) information (according to Smets [20]).
Table 1. Example of quantified expressions that can be used in rules

Syntax Example Scenario
Q gi Many positive_answers Suppose the system is asking its users
for opinion about a new feature, and
positive answer is defined as a fuzzy
set on a ten-point scale valuation
obtained through a form.
Q gi are gj Most frequent_users are Proportional quantification.
beginners.
Q instance.att Few [user’s] visits Queries about the absolute amount of
Q instance.rel [in- Few [user’s] visits to current user’s visits, and the relative
stance] amount of visits to an specific section.
channel A.
Q instance.att are gk About half the [user’s] Query about the typical session of the
Q gh instance.att are gk sessions are short. current user, and query about the
Most excellent [user’s] correlation of assignment scoring and
assignment are short. length for a user.
The first and second syntaxes are related to subpopulations or groups of entities
(characterize user’s groups or global domain entities descriptions). The rest are about
countable evidence related to a particular entity (i.e. a specific user or hypermedia
node).
3 Adaptation Methods and Techniques Using Fuzzy Quantifiers
A quantifier is specified as a predefined function and its parameters (if other func-
tions need to be specified, MATHML1 content markup could be used instead). For
example:
<fuzzy:quantifier>
<fuzzy:quantName>most</fuzzy:quantName>
<fuzzy:quantKind>relative</fuzzy:quantKind>
<fuzzy:sFunction initRange=0 endRange=1 >
<fuzzy:par1>0.6</fuzzy:par1>
1 <http://www.w3.org/TR/MathML2/>
</fuzzy:sFunction>
</fuzzy:quantifier>
Linguistic labels can be defined in a similar way, but they need to be defined on at-
tributes of ‘the user’ or an entity in the DM. For example, ‘loyal’ users can be defined
by a left linear function on attribute numberOfPurchasesYear as follows:
<fuzzy:label>
<fuzzy:labelName>loyal</fuzzy: labelName>
<fuzzy:entity> user </fuzzy:entity>
<fuzzy:attribute>numberOfPurchasesYear</fuzzy:attribute>
<fuzzy:definition>
<fuzzy:lFunction range=numberOfPurchasesYear >
<fuzzy:par1>2</fuzzy:par1>
<fuzzy:par2>10</fuzzy:par2>
</fuzzy:lFunction>
</fuzzy:definition>
</fuzzy:label>
Note that the domain of the fuzzy set is the user entity. Information about the schema
(attributes and associations) of UM and DM models are defined in the persistence
mechanism, so we only need to specify the entity.
Label sets can be defined in a similar way with a <fuzzy :labelSet> tag that
includes several label definition, all of them on the same attribute or association.
We have used the Fuzzy Java Toolkit2 for the implementation of rules (inference is
performed in two phases, one for user modeling and the second for adaptation),
mixed with our SHADOWS framework for handling fuzzy types in relational databases.
Independent terms are stored as fuzzy relations, which are implemented in a straight-
forward manner in a relational database by adding attributes to the users table or by
including an additional table for each term (both representations are allowed in our
implementation). Linguistic label sets are implemented as metadata and values are
stored as pairs of keys (label_set, label). In what follows some examples of rules are
given; we have specified adaptation rules in a XML syntax that we have tried to keep
close to the ongoing research effort of RuleML3 [8] (although that effort considers
fuzzy logic as one of its future target semantics, it does not support currently fuzzi-
ness).
Preferences in Information Browsing Contexts

Ad-hoc fuzzy quantified expressions have been used to personalize presentation pref-
erences (number, text size and reading complexity of recommendations) in browsing
environments with potentially large information spaces [14]. In this case, the user
provides indirect feedback about his/her preferences by rating the items he visited so
that evidence can be found over time with common-sense rules in the form “if most
2 <http://www.iit.nrc.ca/IR_public/fuzzy/fuzzyJToolkit.html>
3 <http://www.dfki.uni-kl.de/ruleml>
long news are rated poorly [by the user] then tend to select shorter ones”. Adaptive
technologies implemented this way can act as a second filter for a basic information
retrieval function. The rule antecedent can be codified as follows:
<rule>
<_head>...</head>
<_body>
<and><atom>
<_opr> <rel>most</rel> </_opr>
<var>large(user.newsVisited)</var>
<var>poor(user.newsVisited@)</var>
</atom></and>
</_body>
</rule>
Where ‘most’ refers to the previously defined quantifier, ‘long’ is a single label de-
fined on each news’ text length and ‘poor’ is one of the labels in a fuzzy label set
‘ratings’ (note that they’re extensions to RuleML syntax). As news ratings are link
attributes of the association between User and News classes, the ‘at sign’ is used
according to the ONN notation described in [1].
Examples on Adaptive Navigation Support

Adaptive hypermedia technologies can be performed with quantified rules as a tech-
nique when exploiting quantitative information about user’s browsing history or hab-
its. For example, a simple direct guidance [2] behavior can be implemented for the
homepage of a university research group that has several sections (e.g. teaching, re-
search, people and resources) with a rule in the form ‘if most user sessions browse
teaching nodes then <<directly guide the user to that section>>’:
<_body><and><atom>
<var>user.sessions</var>
<var>teaching-related(user.sessions)</var>
</atom></and></_body>
As a second example, adaptive annotation techniques can augment a link with some
kind of information about the node behind the link. A rule like the following can
activate a longer description or visual differentiation for links that are not frequently
visited or are shortly visited by the user:
<_body><and><atom>
<_opr> <rel>few</rel> </_opr>
<var>user.visits</var>
</atom>
<atom>
<_opr> <rel>few</rel> </_opr>
<var>long(user.Visits)</var>
</atom>
</and></_body>
Note that in this latter case, the first quantification is operating on a crisp set, but both
atoms yield a fuzzy membership grade due to the application of the quantifier.
Adaptation based on User’s Knowledge

Quantifiers can help in specifying vague or compensatory criteria in adaptive text
presentation, for example, in the case of adaptive educational systems, a specific
hypermedia node n explaining knowledge item k is showed to user u only if a set of
knowledge prerequisites are fulfilled. Then, we have a (fuzzy) relationship between
users and knowledge items (i.e. a user ‘knows’ an item with a specified grade), and
another (possibly fuzzy) relation between knowledge items. If we define prerequisites
as a term defining the fuzzy subset of the knowledge items that are k’s prerequisites
and currentKnowledge as the fuzzy subset of items user u knows, we can relax the
presentation criteria with expressions in the form ‘if most prerequisites are cur-
rentKnowledge then <<show the item>>’. The rule antecedent can be expressed as
follows:
<_body><and>
<atom>
<var>prerrequisite(context.item)</var>
<var>user.currentKnowledge</var>
</atom>
</and></_body>
Note that the context is used to refer to the item that is connected to the node that the
system is deciding to show or not. The fuzziness of the association currentKnowl-
edge implicitly defines a fuzzy subset, and therefore this rule has the semantics of the
last example in Table 1.
Quantified expressions enhance the expressive power of rule languages due to its
closeness to human language expressions. In the formulation of adaptation rules in
AHS, quantified expressions can be used as adaptation methods to specify a number
of adaptation technologies. Finally, fuzzy quantification mechanisms can be used to
implement adaptation techniques based on quantification, operating on a fuzzy
knowledge representation that includes linguistic label sets and fuzzy relations.
Current markup interchange languages for rules need to be complemented to be able
to express adaptation rules that operate on complex user and/or domain models. We
have defined our own semiformal language combining RuleML, ONN and extensions
to the first, and currently a formal definition is in progress.
Future research will address the formulation of complex quantified expressions,
applied to general nominal phrases that can involve implicit queries in the user model.
In addition, the linguistic adequacy of quantifier implementations should be
experimentally validated to ensure they have an appropriate behavior in diverse
mentally validated to ensure they have an appropriate behavior in diverse adaptation

contexts, using usability testing techniques (as is common in the field [10]).
Acknowledgements
This work is supported by “Dirección General de Investigación del Ministerio de

Ciencia y Tecnología”, project number TIC2000-0402.
References
1. Blaha, M. and Premerlani, W.: Object-Oriented Modeling and Design for Database
Applications. Prentice Hall, Upper Saddle River, New Jersey (1998)
2. Brusilovsky, P.: Methods and techniques of adaptive hypermedia. User Modeling and
User- Adapted Interaction, 6(2/3), Kluwer Academic Publ., The Netherlands (1996)
87–129
3. Brusilovsky, P.: Adaptive Hypermedia. User Modeling and User-Adapted Interaction
11(1/2), Kluwer Academic Publ., The Netherlands (2001) 87–110
4. Díaz, P., Aedo, I., Montero, S.: Ariadne, a Development Method for Hypermedia. In:
th
Proceedings of the 12 International Conference on Database and Expert Systems Ap-
plications (DEXA’01) (2001) 764–774
5. Dodero, J. M., Sicilia, M.A., García, E.: A Fuzzy Aggregation-Based Reputation
Model for e-Learning Exploitation of Public Domain Resources. In: Proceedings of
th
the 4 International ICSC Symposium on Soft Computing and Intelligent Systems for
Industry (SOCO/ISFI'2001), Paisley, Scotland, UK (2001)
6. Fink, J. and Kobsa, A.: A Review and Analysis of Commercial User Modeling Servers
for Personalization on the World Wide Web. User Modeling and User-Adapted Inter-
action 10(3/4), Kluwer Academic Publ., The Netherlands (2000) 209–249
7. Glöckner, I. and Knoll, A.: A Framework for Evaluating Fusion Operators Based on
the Theory of Generalized Quantifiers. In: Proceedings of the 1999 IEEE International
Conference on Multisensor Fusion and Integration for Intelligent Systems (MFI '99),
Taipei, Taiwan (1999)
8. Grosof, B.N.: Standarizing XML Rules. In: Proceedings of the Workshop on E-
Business & the Intelligent Web (IJCAI 2001), Seattle, USA (2001)
9. Herrera, F., Herrera-Viedma, E., Verdegay, J. L.: Aggregating Linguistic Preferences:
th
Properties of LOWA Operator. In: Procceedings of the 5 IFSA World Congress
1995, Sao Paulo, Brasil (1995) 153–156
10. Höök, K.: Evaluating the Utility and Usability of an Adaptive Hypermedia System.
Journal of Knowledge-Based Systems 10(5) (1998)
11. Jameson, A.: Numerical Uncertainty Management in User and Student Modeling: An
Overview of Systems and Issues. User Modeling and User-Adapted Interaction, 5,
Kluwer Academic Publ., The Netherlands (1996) 193–251
12. Kacprzyk, J., Zadrozny S., Yager R.: A fuzzy logic based approach to linguistic sum-
maries of databases. Intl. Journal of Applied mathematics and Computer Science 10(4)
(2000)
13. Kobsa, A.: Generic User Modeling Systems. User Modeling and User-Adapted Inter-
action 11(1/2), Kluwer Academic Publ., The Netherlands (2001) 49–63
14. López, L., Sicilia, M.A., Martínez, J.J., y García, E.: Personalization of Web Interface
rd
Structural Elements: A Learning-Scenario Case Study. In: Proceedings of the 3
Mexican International Conference on Computer Science (ENC’01), Aguascalientes,
Mexico (2001)
15. Medina, J. M., Pons, O. and Vila, M.A.: GEFRED: A Generalized Model of Fuzzy
Relational Databases, Information Sciences, 76(1/2) (1994) 87–109
16. Moxey, L.M. and Sanford, A.J.: Communicating Quantities: A review of psycholin-
guistic evidence of the control of perspective by quantifying expressions. Applied
Cognitive Psychology, 14 (2000) 237–255
17. Presser, G. Personalization of Newsletters Using Multistage Fuzzy Inference. In: Pro-
ceedings of the Fuzzy Days Conference, Lecture Notes in Computer Science 2206,
Springer-Verlag (2001) 629–636
18. Rasmussen, D., Yager, R.R.: SummarySQL: A Fuzzy Tool For Data Mining. Intelli-
gent Data Analysis, Elsevier (1997)
19. Sicilia, M.A., Dodero, J.M.: User Stereotype Modeling for Approximate Reasoning-
th
Based Adaptive Enterprise Portals. In: Proceedings of the 10 European-Japanese
Conference on Information Modeling and Knowledge Bases, Säariselka, Finland
(2000) 177–184
20. Smets, P.: Imperfect information: Imprecision-Uncertainty. In: Motro, A., Smets, P.
(eds.) Uncertainty Management in Information Systems: From Needs to Solutions,
Kluwer Academic Publishers (1997) 225–254
21. Tsoukalas, L.H., Uhrig, R.E.: Fuzzy and Neural Approaches in Engineering. John
Wiley & Sons, New York (1996)
22. Wu, H., De Kort, E., De Bra, P.: Design Issues for General-Purpose Adaptive Hyper-
media Systems. In: Proceedings of the ACM Conference on Hypertext and Hyperme-
dia, Aarhus, Denmark (2001) 141–150
23. Yager, R.R.: Connectives and quantifiers in fuzzy sets. Fuzzy Sets and Systems, 40
(1991) 39–75
24. Zadeh, L.A.: A Computational Approach to Fuzzy Quantifiers in Natural Language.
Computing and Mathematics with Applications, 9(1) (1983) 149–184
The Plight of the Navigator:
Solving the Navigation Problem for Wireless Portals
Barry Smyth1 and Paul Cotter2
1 ChangingWorlds Ltd.
Trintech Building
South County Business Park
Leopardstown, Dublin 18, Ireland
[email protected]
2 Smart Media Institute
University College Dublin
Belfield, Dublin 4 , Ireland
[email protected]
Abstract. The wireless Internet, as epitomized by the first generation of WAP-

enabled phones and portals, has failed to meet user expectations, resulting in
limited take-up, and poor revenue growth for mobile operators. A combination
of factors has been responsible: unreliable early handsets; poor content; slow
connections; and portals that were difficult to use and navigate. Today, the first
3 of these issues have been solved (or are about to be) by improved handsets,
high-quality content and high-speed infrastructure such as GPRS. However,
portal usability remains a key problem limiting the ease with which users can
locate and benefit from wireless content services. In this paper we describe how
personalized navigation techniques can greatly enhance the usability of
information services in general, and WAP portals in particular, by personalizing
the navigation structure of a portal to the learned preferences of individual
users, and how this has resulted in increased WAP usage in live user trials.
1 Introduction
The wireless Internet, as represented by the current generation of WAP portals (see
www.wapforum.org) and WAP-enabled mobile phones, has failed to live up to the
marketing hype. Poor usability is a key problem, due mainly to poor portal design,
and exacerbated by limited device functionality, bandwidth, and content. However,
the bandwidth and content issues have largely been, or are being, resolved, and the
current generation of phones offer users improvements over the original models.
However, poor portal design remains an issue. The core problem is that the menu-
driven nature of WAP portals, whereby users access content services by navigating
through a series of hierarchical menus, means that users are spending a significant
amount of their time on-line navigating to content. That this frustrates users and limits
the efficiency of mobile information access should be clear. But the fact that most

The Plight of the Navigator: Solving the Navigation Problem for Wireless Portals 329
mobile operators charge users for their navigation time (as well as their content time)
simply adds insult to injury. The result: WAP offers poor value-for-money.
Recent studies highlight the scale of this usability problem and the gross mismatch
between user expectations and WAP realities. For example, the Neilson-Norman
WAP Usability Report [6] indicates that while the average user expects to be able to
access relevant content within 30 seconds, the reality is often closer to 150 seconds.
In this paper we present an automated solution to this navigation problem (used in
TM
the ClixSmart Navigator system developed by ChangingWorlds Ltd.1) that allows
navigation effort to be reduced by more than 50% by personalizing a WAP portal to
the needs and preferences of users, based on their historical usage. We show that for
every 1 second of navigation time saved, users invest an additional 3 seconds
accessing content, which provides strong practical evidence for improved user
experiences, in addition to the obvious incremental revenue benefits for operators.
2 Click-Distance as a Model of Navigation Effort
The length of time that it takes a user to access a content item can be viewed as a
good independent measure of navigation effort. Moreover, we suggest that the
navigation effort associated with an item of content depends critically on the location
of that item within the portal structure, and specifically on the number of navigation
steps that are required in order to locate and access this item from a given starting
position within the portal (typically the portal home page).
16 35
Click-Distance 30
Navigation Time (secs)
12 Navigation Time
25
Click-Distance
20
8
15
10
4
5
0 0
1 3 5 7 9 11 13 15 17 19 21 23 25 27
Day
Fig. 1. A plot of click-distance versus navigation time shows a clear correlation between both
measures of navigation effort.
With the current generation of WAP phones, there are two basic types of
navigation steps. The first is the menu select: the user clicks to select a specific menu
1 The authors of this paper would like to thank the Navigator team in ChangingWorlds,
Elizabeth McKenna, Nigel Hanby, James Reilly, Paul McClave, Michael Kerrigan, and
Creina Mahon, for their hard work and dedication to this product.
330 B. Smyth and P. Cotter
option. The second is a menu scroll: the user clicks to scroll up or down through a
series of options. Accordingly, an item of content within a WAP portal can be
uniquely positioned by the sequence of selects and scrolls needed to access it, and the
navigation effort, associated with this item can be simply modeled as click-distance,
the number of these selects and scrolls (see Eq. 1).
For content item, I, Click − Dis tan ce(i ) = Selects(i ) + Scrolls (i ) (1)
Although this simple model of navigation effort assumes equal weights for scrolls
and selects, when we evaluate click-distance in comparison to navigation time by
analyzing the behaviour of live-users on commercial WAP portals, we find a near-
perfect correlation. Fig. 1 presents the results of a recent evaluation (6 weeks of WAP
usage for 100 users - see Sect. 4) that results in a correlation coefficient of 0.92
between click-distance and navigation time. Thus, the click-distance of a content item
is a strong predictor of the navigation time associated with its access.
Our strategy for decreasing navigation effort is to reduce the click-distance of the
content items that a given user is likely to be interested in by promoting these items
(or the links that lead to them) to higher positions within the portal menu structure.
3 Personalized Navigation
The basic idea behind our personalized navigation technique is to use a probabilistic
(Markov) model of user navigation to predict the probability that a given menu option
o will be selected by a given user u given that they are currently in menu m, and based
on their past navigation history; that is, we wish to compute P(o|m) (the access
probability of o given m) for all options o accessible from m (either directly or
indirectly, through descendant menus). Put simply, when a user arrives at menu page
m, we do not necessarily return the default options, o1, …, on, instead we compute the
options, o’1,…,o’k, that are most likely to be accessed by the user given that they have
navigated to m; that is the k menu options accessible from m that have the highest
access probabilities. This can mean promoting certain menu options, which by default
belong to descendant menus of m, up to m. The size of the final personalized menu is
constrained by some maximum number of options, k, and the constituent options of m
are ordered according to their access probabilities.
3.1 Profiling Navigation Histories
To construct an accurate picture of a user’s navigation history it is necessary to track

and store the user’s accesses to individual menu options. This is efficiently achieved
through the use of a so-called hit table; see Fig. 2(a) for an example of a partial menu
tree and corresponding hit table entries. This hash table is keyed according to menu id
and stores a list of accesses made by that user to options within that particular menu.
For example, Fig. 2(a) indicates that the particular user in question has accessed
option B of menu A 10 times and option C 90 times.
In fact, we need two tables: a global static hit table that is initialized with respect to
the default portal structure (Fig. 2(b)); and a user hit table that records each user’s
particular history on the portal. The static table makes it possible to deliver the
standard menu structure early on but this will eventually be over-ridden by the
personalized menu once the access probabilities build. Moreover, the default hit
values that are set in the static hit table make it possible to control the personalization
latency – low values mean that personalization takes effect very quickly, while large
values make the system less sensitive to user activity.
(a) (b)
A (100) Hit Table (User) A (40)
A (B 10)(C 90)
B (10) C (90) B (20) C (20)
B (D 5) (E 5)
(5) (5) (80) (10) (10) (10) (10) (10)
D E F G C (F 80) (G 10) D E F G
Fig. 2. (a) A partial menu tree and corresponding user hit table; (b) Represents a static menu
tree feeding a static hit table.
3.2 Constructing a Personalized Menu
To construct menu m we use a Markov model to identify the k most probable options
from m; that is the k options with the highest P(o|m). We take account of the hit
values listed for each option in both the static and user hit tables. For the data shown
in Fig. 2 the following access probabilities are determined:
P(B|A) = (20+10)/(40+100) = 0.214

P(C|A) = (20+90)/(40+100) = 0.786
P(D|A) = P(B|A)P(D|B) = (30/140)(10+5)/(20+10) = 0.107
P(E|A) = P(B|A)P(E|B) = (30/140)(10+5)/(20+10) = 0.107
P(F|A) = P(C|A)P(F|C) = (110/140)(10+80/20+90) = 0.642
P(G|A) = P(C|A)P(G|C) = (110/140)(10+10)/(20+90) = 0.142
So in descending order of access probability we have C, F, B, G, D, and E. And for

k=3, C, F, and B are selected, in order, for menu A.
3.3 Promotion and Demotion
This personalized navigation method supports two basic types of menu adaptation,
which we call promotion. A menu option may be promoted within the context of its
default menu; that is, its relative position may be changed. A promotion between
menus occurs when a menu option is promoted into an ancestral menu. Promotions
(and conversely demotions) are side effects of the probability calculations. For
instance, in the above example link F is promoted to A’s menu – and in theory links
can be promoted from deeper levels once their probabilities build sufficiently.
Of course if F is subsequently selected from A, it is added to A’s entry in the user’s
hit table. So the next time that menu A is created, and P(F|A) needs to be computed
we must account for the new data on F (see Fig. 3(a) for example). Specifically,
assuming a single access to F as an option in A, we get:
P(F|A) = 1/101 + (110/141)(10+80/20+90) = 0.009 + 0.638 = 0.647
3.4 An Efficient Algorithm for Computing Promotable Options
The complexity of the proposed method depends on the complexity of the process that
identifies the k most probable options for the menu, m. This can mean examining not
just the options of m, but also all the options contained in menus that are descendents
of m. Fortunately, a more efficient algorithm is possible once we recognize that,, by
definition P(o|m) is always greater than or equal to P(o’|m) where o’ is an option of a
menu, m’, which is itself a descendent of m through o. This means that we can find
the k most probable nodes for menu m by performing a depth-limited breadth-first
search over the menu tree rooted at m. Moreover, we only need to expand the search
through an option o’ if P(o’|m) is greater than the kth best probability so far found.
(a) (100) (b)

A A
B (10) F (1) C (90) B .214 F .009 C .786
D (5) E (5) F (80) G (10) D E F .642 G .142
Fig. 3. (a) Option F has been promoted to menu A from menu C and accessed once; (b)
Cutting-off search at B and G during the personalization of A..
For example, as Fig. 3(b) indicates, during the calculation of the access probabilities
for A’s descendants with k=2, search can be initially cut-off at option B, since B’s
children cannot have access probabilities greater than 0.214, which is the probability
th
of the k best option found so far (B itself). Similarly, after computing the access
probabilities for C’s options (F and G), search can be cut-off at G since its probability
th
is less the 0.642, the new k best option. In practice this technique can result in
significant reductions in search effort allowing probabilities to be computed on-the-
fly without a noticeable impact on the personalization time.
4 Experimental Evaluation
The following evaluation is based on live-user field trials on European WAP portals.
The standard trial consisted of a 2-week profiling period in which no personalization
took place but the behaviour of the users was monitored in order to profile their
navigation patterns. The remaining 4 weeks are divided into two 2-week
personalization periods. During this time profiling continued but in addition,
personalization was switched on so that users experienced a new portal structure that
was adapted to their navigation preferences. The trial consists of approximately 100
trialists from a variety of backgrounds and with a range of mobile usage habits and
handsets.
4.1 Click-Distance
Figure 4 shows how click-distance changed during the trial. It presents the average
user click-distance from the portal home page to each of the user’s top 3 content sites,
during each trial period. The results show that the default click-distance of 13.88 for
the static portal (experienced during the profiling period) dropped significantly by
over 50% to 6.84 during the first personalization period and by a further 2% for the
final two weeks of personalization. These results show two things: first that
significant click-distance reductions are possible; and secondly, that these reductions
are realized very rapidly, in this case within the first two weeks of personalization.
60%
16 52.45%
50.72%
14 13.88 50.%
Relative Decrease
Click-Distance
12
40.%
10
30.%
8
6.84 6.6
6 20.%
4
10.%
2
0%
0 0.%
Static (Wk 1-2) Pers. (Wk 3-4) Pers. (Wk 5-6)
Fig. 4. Comparative click-distance results for the profiling and personalization periods.
4.2 Navigation vs Content Airtime
Figure 5 shows the reduction in average daily navigation time for the trialists. We find
that over the 4-week personalization period (weeks 3-6) average daily navigation time
has reduced by 36%. During the initial static (profiling) period users were spending
an average of 56.42 seconds navigating to content each day, and this fell to only 35.99
seconds for the 4 weeks of personalization. Indeed if we look at the results for the
final two weeks of personalization (weeks 5-6) in comparison to the first two weeks
of personalization (weeks 3-4) we see that the benefits of personalization are
incremental, with navigation time reducing from an average of 36.55 seconds (weeks
3-4) to 35.43 seconds (weeks 5-6).
When we look at the airtime spent by users accessing content pages (see Fig. 6) we
find that there is a significant increase as a result of personalization. Over the 4-week
personalization period (weeks 3–6) average daily user content time increases by
nearly 17%. During the static period the average daily content time per trialist was
312.46 seconds compared to 364.55 for the 4-week personalization period. Moreover,
if we again look at the content time for the final two trial weeks (as opposed to the
final 4 weeks) we find a relative increase of more than 22% (content time of 382.62
seconds). In other words, the relative change in content time for the final two weeks
of the trial (22.45%) has more than doubled in comparison to the first two weeks of
personalization (10.89%); as personalization proceeds so too do the benefits increase.
60 56.42 10%
0%
50 0%
Relative Increase
Time (secs)
40 36.55 -10%
35.99 35.43
30 -20%
20 -35% -30%
-36% -37%
10 -40%
0 -50%
Wk 1-2 Wk 3-4 Wk 3-6 Wk 5-6
Fig. 5. Comparative navigation airtime results for the profiling and personalization periods.
25%
420 22%
400 20%
Relative Increase
16% 382.62
Time (secs)
380
15%
10% 364.55
360
10%
340 346.4
5%
320 312.46
0%
300 0%
Wk 1-2 Wk 3-4 Wk 3-6 Wk 5-6
Fig. 6. Comparative content airtime results for the profiling and personalization periods.
These results show that users are willing to trade savings in navigation time for
increases in content time. For every second of navigation time saved the average user
increases their content time by 3 seconds – by the final two weeks of personalization
the average user has saved 22.99 seconds of navigation time but increased their
content time by 70.16 seconds. There are obvious benefits here for the mobile
operator from a revenue point of view, not only in terms of existing airtime-based
charging models but also as operators move to content-based charging models.
4.3 Page Impressions
Figure 7 illustrates that, as a result of personalization, the average number of daily

page impressions per user increases by 35% by the end of the trial (the 29.4 average
daily page impressions per user in the static period grows to 39.65 during the final 2
weeks of personalization).
45 45%
39.65
40 37.53 40%
35.42 34%
Page Impressions
Relative Increase
35 35%
29.40 27% 30%
30
25 25%
20%
20 20%
15 15%
10 10%
5 5%
0%
0 0%
Wk 1-2 Wk 3-4 Wk 3-6 Wk 5-6
Fig. 7. Comparative page impression results for the profiling and personalization periods.
Once again, as with the airtime results, we have separately reported the average
results for the final 2 weeks of personalization in addition to the normal 4-week
period in order to emphasise that the page impression count is continuing to increase
as personalization proceeds. For instance during the first 2 weeks of personalization
the page impression count increases by 20% compared to the static period. Thus, there
is an additional 15% rise associated with the final 2 weeks of personalization.
4.4 Discussion
The results presented in this paper represent only a small subset of the personalization
benefits learned from the trial, and for reasons of space it has been necessary to omit
many other important results. In addition similar usage increases have been found for
other important statistics including the average number and length of user sessions
and the number of unique accesses to content sites. Very briefly: the average number
of daily sessions per user increased by 31%; average session length increased by 25%;
and the average number of user accesses to content sites increased by 30%.
Finally, it is worth pointing out that while the above results are based on a 100-user
trial, qualitatively and quantitatively similar results are being found from a much
broader 100,000-user trial that is underway at the time of writing.
5 Related Work
In general, with the advent of the Internet and the reality of the information overload
problem, there has been significant attention given to how personalization techniques
can be used to filter information more precisely according to the learned preferences
of individual users. Many different application domains have been investigated from
news articles [2] to TV programs [9], and a range of techniques has been developed
including content-filtering, collaborative filtering, and probabilistic methods [7]. For
the most part, research has focused on content personalization, that is, the
prioritization of content items within particular content domains. This is in contrast to
work described in this paper where the focus is on navigation (see also [8]).
In the navigation area the work of Perkowitz [4] has addressed the related so-called
shortcut link problem in traditional web sites, and attempts to make shortcut
recommendations from a page P to another page Q based on how often Q is viewed
by users after P in some navigation trail. An alternative approach to shortcut
recommendation is taken in the SurfLen [3] and PageGatherer [5] systems. This time
page links are suggested based on page request co-occurrences with past sessions.
Briefly, the basic idea is to suggest the top k pages that are most likely to co-occur
with pages that the user has visited in their current session. These techniques have
been used to good effect in adaptive Web sites as a means of generating automatic
index pages or for inserting shortcut links into existing pages and we believe they
could be equally well-applied to personalizing navigation on wireless devices. More
recently, the wireless navigation problem has also been considered by Anderson et al.
[1] who have investigated Naïve Bayesian and Markov modeling techniques to make
navigation link recommendations as a way of automatically repurposing Web pages
for wireless devices. The particularly interesting feature of this work is the explicit
use of a recommendation metric that not only accounts for the relevance of the
navigation link in question but also the level of navigation effort that can be saved by
recommending this link, a feature that could also be incorporated into the technique
presented in this paper.
The above represents an abbreviated list of related research in the area of
personalization and for reasons of space it has not been possible to provide a more
complete analysis. The above techniques share some motivations, objectives and
technical features with our approach. However it is also worth pointing out that the
research in this paper is currently deployed as part of an enterprise-level solution for
mobile operators capable of handling millions of users and real-time personalization.
The ClixSmart NavigatorTM product by ChangingWorlds Ltd. also includes a wide
range of sophisticated administration and reporting tools for the mobile operator in
order to control the level and type of personalization offered. It is also worth pointing
out that the probabilistic personalization technique describe here is just one of a
number of personalization methods currently supported by the Navigator system.
Additional approaches include, for example, collaborative recommendation
techniques that allow navigation links to be recommended based on the navigation
patterns of related users (see www.changingworlds.com for further information).
6 Conclusions
The current incarnation of the mobile Internet, which is largely based on WAP
(wireless application protocol) phones and portals, has met with only limited success.
One of the most significant problems facing mobile users is the time it normally takes
to locate relevant content. Many mobile portals have average click-distances in excess
of 15 and recent studies have shown that many users are taking an average of 120
seconds to perform many common content access tasks, while the users themselves
are expecting access times of about 30 seconds. This has resulted in frustrated users
and limited WAP up-take and usage.
In this paper we have described a personalized navigation technique, developed by
TM
ChangingWorlds Ltd. as part of its ClixSmart Navigator solution for mobile
operators, that can actively adapt the structure of a portal to match the preferences and
needs of individual users in order to radically reduce portal click-distance on a user-
by-user basis. Specifically we have shown how click-distance can be reduced by over
50% and how the corresponding navigation time can be reduced by nearly 40%.
From an end-user point of view, the result is a WAP portal that is easier to use and
that delivers superior value-for-money. In turn, this leads to increased usage by users
and the results reported here highlight content time increases of over 20% and page
impression increases of 35%. Furthermore, for each second of navigation time that is
saved, the average user is willing to invest an additional 3 seconds in content time to
deliver significant increases in overall airtime. Crucially, this means improved
incremental revenue opportunities for mobile operators alongside improved user
loyalty and reduced churn.
As mobile operators attempt to shift from pure airtime-based charging models
(where navigation time is charged at standard rates) to content-based charging models
(where navigation time not charged for, or charged at a reduced rate), ClixSmart
Navigator ensures that revenue levels are maximized by increasing content time in
favour of reduced navigation times. At the same time, network resources that would
have been utilized in the service of navigation are being more profitably re-deployed
in the service of content.
References
1. Anderson, C., Domingos, P., and Weld, D. Adaptive Web Navigation for Wireless Devices.
th
In: Proceedings of the 17 International Joint Conference on Artificial Intelligence. Seattle,
WA (2001).
2. Billsus, D. ,Pazzani, M.J. and Chen, J. A learning agent for wireless news access.
Intelligent User Interfaces (2000) 33-36
3. Fu, X., Budzik, J., and Hammond, K. Mining Navigation History for Recommendation. In:
Proceedings of the Conference on Intelligent User Interfaces (2000).
4. Perkowitz, M. Adaptive Web Sites: Cluster Mining and Conceptual Clustering for Index
Page Synthesis. PhD Thesis, Department of Computer Science and Engineering. University
of Washington. (2001).
5. Perkowitz, M. and Etzioni. Towards Adaptive Web Sites: Conceptual Framework and Case
Study. Journal of Artificial Intelligence. 118(1-2) (2000).
6. Ramsey, M. and Nielsen, J. The WAP Usability Report. Neilsen Norman Group (2000).
7. Reiken, D. Special Issue on Personalization, Communications of the ACM. 43(8) (2000)
8. Rucker, J. and Polanco, M.J. Personalized Navigation for the Web. Communications of the
ACM. 40(3) (1997) 73-75.
9. Smyth & Cotter. Wapping the Web: A Case-Study in Content Personalization for WAP-
st
enabled Devices. In: Proceedings of the 1 International Conference on Adaptive
Hypermedia (AH2000). Trento, Italy. (2000)
Towards an Adaptive Web Training Environment Based
on Cognitive Style of Learning: An Empirical Approach
Maria A. M. Souto1, Regina Verdin2, Ricardo Wainer2, Milton Madeira2,

Mariusa Warpechowski1, Karine Beschoren3, Renata Zanella3,
Juarez Sagebin Correa3, Rosa M. Vicari1, and José Palazzo M. de Oliveira1
1 PPGC/UFRGS, Instituto de Informática, Caixa Postal 15.064, CEP 91501-970 Porto Alegre,
RS, Brasil, Fax: +55(51)3316.7308
{Souto, Mariusa, Rosa, Palazzo}@inf.ufrgs.br
2 UNISINOS, Faculdade de Psicologia, Caixa Postal 275, CEP 93022-000, São Leopoldo, RS,
Brasil, Fax: +55(51)590-8268
{Regina, Wainer, Madeira}@cirrus.unisinos.br
3 CRT Brasil Telecom, CEP 91430-001, Porto Alegre, RS, Brasil, Fax: +55(51)3378-4828
{Renata, Karine, Jsagebin}@crt.net.br
Abstract. The paper presents an investigation about the learner’s cognitive

profiles in the context of Tapejara Project1. In this investigation, we have
assumed that the learner’s Cognitive Style of Learning (CSL) actually
influences his/her cognitive behavior while performing a long distance Web
course. The paper describes the empirical procedures used to generate the
learners’ cognitive profiles and his/her associated learning behavior that must
be considered in the adaptive Web training environment. The training target
population refers to the Telecommunications Company’s employees. The
learner’s CSL assessment was obtained from the statistical analysis of the Ross
Test data [14]. The CSL learning behavior was obtained from the statistical
analysis of the navigational log data in an experimental training module
designed for this purpose. The statistical results have shown five CSL groups,
their style dimensions and the correlation between each CSL group and their
navigational behavior in the Web training module.
1 Introduction
Nowadays, long distance corporate training via the Web is an increasingly necessary
issue to be considered. From a corporation’s viewpoint this means to minimize
training expenses and not taking much time of its employees; from an employee’s
viewpoint this means flexibility to choose time and place to accomplish the training.
Usually, commercial training Web systems mainly address hypermedia appeals to
keep online learners interested and self-motivated to proceed, but they leave a gap
between a learner’s cognitive profile and the assistance provided by the system to the
1 The Tapejara Project - Intelligent Training System in the Internet – is a consortium of two
Universities and a Telecom company supported by Brazilian Research Council, CNPq, under
the ProTeM-CC framework.

Towards an Adaptive Web Training Environment Based on Cognitive Style of Learning 339
remote user. To overcome this, many researchers have attempted to define which
individual differences actually influence learning in that system [4,6,7,10,11]. For
example, [10] investigates how individuals manage learning in interactive Web
learning environments designed to support learning orientations; [11] has attempted
to determine what combination of non-linearity and advance organizers work best for
learners with given levels of self-regulation.
The Cognitive Style of Learning (CSL) is an individual aspect that describes the
way in which a person habitually approaches or responds to the learning task [13].
According to [4, 5, 13], a person’s cognitive style is considered one of the most stable
user’s characteristics overtime, influencing a person’s general attainment or
achievement in learning situations. This stability is manifested in the use of
hierarchies’ processes in information treatment and strategies the learner uses when
acquiring new information through a hypermedia system.
Besides, matching the cognitive style to the domain content in hypermedia systems
is a pedagogical form to make the comprehension easier and lead the learner to
choose his/her preferred mode to process information [12, 13]. This occurs because
the cognitive style interacts with the content structure and processes the information
in some differentiated way, which means the use of a specific learning strategy for
each cognitive style [8]. Knowing the cognitive styles and their associated learning
strategy enable the instructional designer to develop learning resources close to the
learner’s profiles.
A wide number of cognitive style dimensions has been generated by psychologists
researchers such as the field-dependency-independency - Witkin and Asch, 1948 -
Wholist-analytic, verbal-imagery - Riding and Cheema, 1991 - etc. A review of these
styles can be found in [13]. We have found out in literature, at least two authors [3, 6]
that have been investigating the cognitive style construct as an important aspect to be
considered in the adaptation of the system characteristics to the user’s cognitive style
in an assisted hypermedia environment.
In the I-Help project, [3] has used cognitive style classification to integrate the
learner’s characteristics to be considered in the Intelligent-Help environment to match
learners who have help requests with potential peer helpers. Bull [3] has used both of
Riding and Cheema’s cognitive style dimensions: verbal-imagery and wholist-
analytic. These cognitive style dimensions were assessed using a five-item
questionnaire. The author doesn’t have any conclusive results about the use of the
learner’s cognitive style in the system yet.
Ford and Chen, [6], have investigated the correlation between field-
dependency/independency cognitive style and learning behavior, and learning
outcomes, while a group of learners were performing a learning task. This task
includes using a hypermedia system to learn how to design a web home page. The
Cognitive Style Assessment - CSA – (the instrument used in this experiment and
designed by Riding’s) was accomplished to determine levels of field-
dependency/independency. The authors also used Study Preferences Questionnaire
(SPQ) to assess levels of holist or serialist style dimension. The main conclusion of
the study is that field-dependency/independency cognitive style dimensions were
linked to strategic differences in navigation.
Like the works mentioned above, the Tapejara Project uses the learner’s CSL as
source of a user’s information to enable adaptation of a hypermedia system’s
characteristics to his/her cognitive style. Our approach is based on two psycho
pedagogical constructs: cognitive style of learning - Atkinson 1998; Riding 1991,
340 M.A.M. Souto et al.
Riding 1997 - and learning trajectory [9]. The CSL characterizes a pattern of
cognitive actions represented by the learning trajectory that can be concretely
observed through the index variables of performance behavior during the learning
practice. The study of these trajectories is accomplished through the analysis of the
learners’ interactions with the training environment during a free process of learning,
i.e. a learning phase without the interference of any artificial or human tutor [9].
In order to assess the target learner’s CSL, we have used the Ross test as described
in Sect. 3. The main objective of this test is approaching the cognitive superior
abilities, based on Bloom’s taxonomy of the educational processes in the cognitive
domain. According to this, we have aimed to take advantage of Bloom’s taxonomy,
which can make the tasks of planning learning experiences easier. Up to now, the
statistical results have showed five CSL groups, their style dimensions and the
correlation between each CSL group and their navigational behavior in the Web
training module. Based on these research outcomes, we have been developing the
training environment aiming to adapt the didactic resources to the learner’s profile. In
the next stage, we are planning to integrate in the environment an artificial agent
collection, whose roles will correspond to (i) monitoring a learner’s trajectories;
(ii) perform the learner’s cognitive diagnostic; and (iii) provide learning strategies to
adjust to the learner’s CSL.
This paper is organized as follows. In Sect. 2 we present the adaptive model, in
Sect. 3 we detail the empirical procedures of CSL class generation, in Sects 4 and 5
we detail, respectively, the domain content and the empirical results about CSL
classes behavior. Finally, in Sect. 6, conclusions and the next research activities are
described.
2 The Adaptive Model
According to [2], we presume that important issues should be considered in

developing an Adaptive Hypermedia System, such as: adapting where and why
(application area and adaptations goals); adapting what (user features); how (methods
and techniques), etc. Up to now, we have worked on the processes of design and
development of a Web training environment to support learners’ cognitive aspects
modeling the CSL and to adjust domain content to this important learner feature.
The purpose of the Tapejara Project Web training environment is to provide the
necessary support to the learner’s style of interaction with a specific knowledge space
represented by HTML pages with wide and rich didactic material. Based on this
interaction, the system identifies a learner’s cognitive style and starts adapting the
instruction to the needs of the style, presenting suitable paths to the learner’s
individual features of learning, looking for optimal learning results.
In order to support training adapted to learners’ cognitive style, the learner model
integrates information from the learners. For example: the Web log of the learner’s
interactions; a set of index variables obtained from the Web log; and the learner’s
most likely cognitive style pattern. These index variables highlight cognitive actions
and a learner’s behavior while browsing the training environment pages. Indexes are
organized in three categories: (i) navigational indexes (i.e. didactic resources and
preferable forms of presentation); (ii) temporal indexes (i.e. time spent in the whole
training module and in the module activities); and (iii) performance indexes (i.e. total
learner attempts to perform an exercise and the final evaluation grade). We assume
that each cognitive style presents a similar behavior related to the index values. This
means that we have a learning path pattern for each cognitive style.
The main requirements of a Web training adaptive environment is a computational
architecture that: supports the identification of a learner’s cognitive pattern based on
his/her interactions with the training environment; knows the domain organization in
the Web; knows the psycho pedagogical rules to support the learner’s cognitive
diagnostic; and knows the appropriate pedagogical strategies assigned to each
learner’s cognitive pattern. In this article we focus on important empirical results of
the Project towards the Web training adaptive environment architecture. These results
consist of: (i) the CSL classes; (ii) the domain content; and (iii) the CSL classes’
preferences.
3 The Cognitive Styles of Learning Classes
3.1 Applying the Psychological Ross Test
For learners’ styles classification we chose a psychological test – The Ross test [14].
The test was conceived to measure the psycho pedagogical abilities relevant to the
Upper Cognitive activities referred by [1]. The main focus of [1] was the Educational
Processes Taxonomy in the cognitive domain. The Ross Test assesses three psycho
pedagogical abilities (Synthesis, Analysis and Evaluation), underlying eight cognitive
processes: analogy, deductive reasoning, absent premises, abstract relations,
sequential synthesis, questioning strategies, relevant and irrelevant analysis of
information and attributes analysis. The main emphasis of the test is on the
individuals’ abilities to deal with abstractions from a verbal base. The ability to
elaborate concepts from abstractions with verbal elements is a major component of
the upper cognitive processes. The test requires the subject to perform a variety of
tasks that reflect these abilities.
The Ross test consists of a set of questions subdivided in two parts with four
sections each. The subjects had to answer on a separate answer sheet. The maximum
time for completion varied from 7 to 12 minutes, depending on the section. The test
was applied in a sample composed of 231 employees randomly chosen from different
departments of the Telecommunications Company.
Data collection through the Ross test lasted six months and collective sessions
were composed of at least 2 and at most 10 subjects in each meeting. Each meeting
lasted for one hour and a half, being divided in two successive sessions (of 45 minutes
and 32 minutes, respectively), with a break of 15 minutes between sessions. The
application setting was accomplished in a room at the company, especially prepared
for the application, with artificial illumination, good ventilation and excellent control
of external stimuli. The applicators were psychologists, trained in the application
procedure and responsible for all the initial report, as well as the reading of the test
instruction’s manual that was standardized for this procedure.
3.2 Analyzing Data from Ross Test
Data analysis was performed at three moments. At the first one, the Ross test was
assessed at three validity levels: (i) from the whole 105 test items (Alpha of
Crombach=0.923); (ii) from the eight Cognitive Processes (Alpha of
Crombach=0.8429); and (iii) from the three cognitive abilities (Alpha of
Crombach=0.8364). This means that the Ross test was consistent and trustworthy. In
order to validate the Ross test to the target population, the main adaptation consisted
of classifying the individuals according to their performance ranges on different test
sections, instead of considering the item numbers at each section.
At the second moment we have proceeded with factorial analysis. The factors
analyzed were related with the eight cognitive processes and the three abilities
assessed by the Ross test. The cognitive process Analogy, which underlies the
Analysis ability, explained 72.61% of the total variance. The processes abstract
relations and sequential synthesis, which underlie the Synthesis ability, explained
39.41% and 33.87%, respectively. The process deductive reasoning, which underlies
the Evaluation, explained 62.97% of the total variance. The processes mentioned
were the most predominant factors in each ability assessed.
In the cluster analysis the subjects were grouped by similarities in performance
levels in each of the cognitive processes and abilities assessed. This statistical
analysis has generated five clusters in a range of [0, 5] in the dendrogram.
The cognitive styles were generated based on empirical procedures, using the
factorial analysis results and the cluster analysis. The representative CSL classes
established are: (i) Analogue-Analytical; (ii) Concrete-Generic; (iii) Deductive-
Evaluative; (iv) Relational-Synthetic and (v) Indefinite style. The nomenclature used
to refer the five CSL classes highlights the most characteristically cognitive process
and psycho pedagogical ability.
Table 1 shows the performance levels in the three pedagogical abilities in each
CSL. The subjects that were assessed as ‘superior’ performance level (4 or 5) in
Analysis ability were classified as Analogue-Analytical style. The subjects that were
assessed as ‘inferior’ performance level (1 or 2) in Analysis ability were classified as
Concrete-Generic style. This could mean that the group tends to work with the whole
information and uses less abstraction than the Analogue-Analytical group. The
subjects that were assessed as ‘superior’ in Evaluation ability, ‘medium-inferior’ in
Synthesis ability and ‘medium’ Analysis ability were classified as Deductive-
Evaluative style and so on.
The fifth group was very heterogeneous, without a performance similarity in the
abilities and cognitive processes assessed by the Ross Test. To our surprise the Ross
test was not sufficiently refined to group subjects according to their performance in
the cognitive abilities. We are aware that the solution for this problem would be to use
another test that could assess the same cognitive aspects as the Ross test does, but that
would enable to utilize another grouping criteria. However, due to Project schedule
time limitations, we decided to investigate only the navigational pattern of this group
and labeled it as Indefinite CSL class.
Table 1. Cognitive Style Groups by Performance (4-5 superior, 3 medium, 1-2 inferior)
Psycho Pedagogical Abilities
CSL Subject Sample Ability 1 Ability 2 Ability 3
Class % (Analysis Synthesis (Evaluation
73%) 10,6% 14,02%)
Analogue-Analytic 109 47,19 4 or 5 - -
Concrete-Generic 60 25,97 1 or 2 - -
Deductive-Evaluative 24 10,39 3 1, 2 or 3 4 or 5
Relational-Synthetic 07 3,03 3 4 or 5 1, 2 or 3
Indefinite Style 31 13,42 3 1,2,3 e 4 1,2,3,4 e 5
4 The Domain Content
Up to now, the main objective related to domain content has been to develop a Web
experimental training module in order to analyze and generate the CSL classes
preferences. The design and development of this training Web module integrates a
multidisciplinary Project teamwork: psychologists, pedagogues, Web designers and a
domain content specialist.
According to training demands of the Telecommunications Company, the
Telecommunications course was chosen to be the first one. An initial module of
TDMA (Time Division Multiple Access) was chosen as experimental module. TDMA
has a medium difficulty level. This module was developed assuming that the learners
have as pre-requisite a working knowledge on multiplex technology.
The non-tutorial interface’s module of TDMA (Time Division Multiple Access)
was developed to prevent the learner inducement to a particular navigation style. This
approach enables the investigation about learners’ CSL classes preferences, behavior
and learners’ performance.
Figure 1 shows the module’s initial page, composed of two layers. In the top-left
position there are three buttons corresponding to the didactic resources, presentation
forms and final evaluation options. On the other part of the page there is a layer with
didactic resources offered to the learner such as concepts, examples, exercises and
evaluation review. The learner has to choose the didactic resources better suited to
his/her CSL. After a learner’s selection the presentation forms are presented using the
same graphical interface as the initial page.
The domain content was designed using traditional didactic practices like: concepts
presentation instruments, examples, exercises, evaluation review and final evaluation.
For each didactic resource, there is a presentation form set. These presentations
include textual, graphical and schematic forms or a mixture of all these. In a set, each
presentation form refers to the same content being explained. The presentation forms
varieties make possible to each learner to construct a customized trajectory which best
fits the specific CSL.
Fig. 1. The TDMA Initial Page
5 The Empirical Results about CSL Classes Behavior
In order to identify the generated CSL classes’ behavior or preferences, we have

carried out the investigation of learners while browsing the TDMA Web training
pages. The sample learners were selected from the most typical ones of each CSL
class. We obtained respectively from the representative groups: Analogue-Analytical
class - 10 subjects; Concrete-Generic class - 7 subjects; Deductive-Evaluative class -
8 subjects; Relational-Synthetic class - 3 subjects and Indefinite Style class–7 subjects.
Before TDMA training module submission, the learners were advised about (i) the
purpose of the investigation; (ii) the TDMA module characteristics; and (iii) their
flexibility to make choices while browsing the training module.
All training environment information was recorded in the application database. The
learner Web log was application-dedicated and the learners’ interactions were
sequentially recorded in a global log file. The log file recorded data/hour of the
interaction, learner’s and Web page visited identifications. Specifically, exercises a
learner attempted and final evaluation information were recorded in another database
file. The training environment used a client/server model. The client was developed
using Web pages and Javascript/PHP tools and the Web server was implemented
using servlets Java. The database access used JDBC connection.
There had been had a number of issues in preprocessing logged data before the
learner navigation patterns and preferences were analyzed. These include: (i)
cleaning/filtering the raw data to eliminate irrelevant items (i.e. learner menus
accesses, connection faults records and application information pages); (ii) grouping
individual page accesses into semantic units (i.e. learner session accesses); (iii)
integration of various data sources such as user CSL, age, instructional level and
gender; (iv) transforming alphanumeric codes into numeric; and (v) calculating time
spent in each Web page he/she visited.
In order to accomplish the statistical analysis, the learners’ Web log data were
grouped by CSL classes. In this investigation, the independent variable was the
learner’s CSL, and the dependent variables were learner’s preferences related to
didactic resources (i.e. concepts pages, exercises, examples and evaluation review)
and presentation forms (i.e. textual, graphical, schematic). The statistical analysis
included descriptive analysis (i.e. mean/standard deviation/range), ANOVA analysis
and Correspondence analysis. The ANOVA results have showed non-significant
differences between CSL classes concerned with didactic resources preferences (α >
0.05), except that the Indefinite CSL class have shown a significant preference for
exercises of filling the blanks presentation form compared with Concrete-Generic
CSL class (α = 0.016).
Comparing the total navigation time in the experimental module, the Relational-
Synthetic CSL class took more time than Concrete-Generic CSL class (α = 0.097 <
0.10) and comparing the mean navigation time in the exercises Web pages, the
Relational-Synthetic CSL class took more time than Concrete-Generic/Analogue-
Analytical CSL classes (α = 0.089 < 0.10). We have observed that the Deductive-
Evaluative CSL class took more time than Analogue-Analytical CSL class at
performing final evaluation (α = 0.051 < 0.10). Comparing the total pages accessed,
the Relational-Synthetic CSL class has accessed more pages than the Deductive-
Evaluative/Indefinite CSL classes (α = 0.058 < 0.10).
Nevertheless, the main differences between CSL classes were better observed from
the Correspondence analysis at the first twenty learner’s actions as Fig. 2 shows.
Figure 2 shows also the local associations among the cognitive styles (AA, CG, DA,
RS, EI) and Didactic Resources (1, 2, 3, 4, 5). These associations are outlined in the
graphical representation. The Analogue-Analytical and Indefinite CSL classes, (AA)
and (EI), have predominantly used the Concepts didactic resources (1). This
preference differentiates these CSL classes from the others. The Concrete-Generic
CSL class (CG) has predominantly used the exercise didactic resource (2), making
this preference a significant characteristic of the class. The Relational-Synthetic CSL
class (RS) has predominantly used the evaluation review didactic resource (4).
Finally, the Concrete-Generic and Relational-Synthetic CSL classes, (CG) and (RS)
respectively, have predominantly used the example didactic resource (3).
Before concluding the Correspondence Analysis, we have verified the learner
presentation forms preferences at the first twenty actions as Fig. 3 shows. The
Analogue-Analytical CSL class (1) has predominantly used concept pages that use
outlined texts and graphics (111) and evaluation review (402). The Concrete-Generic
CSL class (2) has predominantly used example pages that use comparative blocks of
texts and graphics (305), concept pages that use texts and graphics (107) and concept
pages that present schema with questions (109).
The Deductive-Evaluative CSL class (3) has predominantly used evaluation review
(401) and example pages with schema and graphics (303). The Indefinite CSL class
(5) has predominantly used filling the blanks exercises (204) and textual questioning
example (302). Finally, the Relational-Synthetic CSL class (4) has predominantly
used simple choice exercises (203), relating columns exercises (202) and example
pages with texts and graphics (301).
Fig. 2. The Correspondence Analysis Related to Didactic Resources Preferences
Fig. 3. The Correspondence Analysis Related to Presentation Forms Preferences
6 Conclusions
The adaptive Web training environment demands investigation of a multidisciplinary

team such as psycho-pedagogical researchers, Web designers, content domain
specialists, and technology information intelligent systems designers. This is a great
challenge not only concerning experts’ knowledge, but also concerning cooperation
requirements between them. In our investigation, we have attempted to contribute in
this area proposing a new methodology to generate the learner’s cognitive profiles to
be considered in the adaptation of systems characteristics to the learner’s cognitive
features. This work was totally based on empirical procedures that have worked with
a specific target population, the Telecommunications Company’s employees, and

partner of the Tapejara Project.
The study of CSL classes’ behavior consists of another contribution of our work.
This process has included a statistical analysis in order to identify the main features
associated to each CSL classes, concerning its pedagogical actions while browsing a
training Web module. All knowledge obtained from this investigation will give us the
background to develop Web training environment contents. Afterwards, we aim to
propose an artificial agent collection to assist the log distance training, suggesting to
the learner the strategy that best fits his/her cognitive style. We do hope that this
adaptation model will benefit the learner’s training over the Web, making his/her
learning process easier.
References
1. Bloom, B.: Taxonomia de Objetivos Educacionais – Domínio Cognitivo. Porto Alegre:
Ed. Globo, 1972.
2. Brusilovski, P.: Methods and techniques of adaptative hypermedia. In: Brusilovsky, Peter
et al. (Ed.). Adaptative Hypertext and Hypermedia, Netherlands, p. 1-43, 1995.
3. Bull, S., Greer, J.: Peer Help for Problem-Based Learning. In: International Conference on
Computers in Education, Vol. 2. Taipei, Taiwan (2000) 1007-1015.
4. Dufresne, A., Turcotte, S.: Cognitive Style and its Implications for Navigation Strategies.
In: World Conference on Artificial Intelligence and Education, AI-ED'97, 8, Proceedings
... Amsterdam: IOS, 1997.
5. Fierro, A.: Personalidad y Aprendizaje en el Contexto Escolar. In C. Copll; J. Palacios;
A.Marchesi (ed). Desarrollo psicológico y educación, II. Psicologia de la educación.
Madrid: Alianza, 1990.
6. Ford, N. & Chen, S.Y.: Individual Differences, Hypermedia Navigation, and Learning: an
Empirical Study. Journal of Educational Multimedia and Hypermedia 9(4). 281-311, 2000.
7. MacGregor, S. K.: Hypermedia navigation profiles: cognitive characteristics and
information processing strategies. Journal of Educational Computing Research, 20(2).
189-206, 1999.
8. MacLoughlin, C.: The Implications of the Research Literature on Learning Styles of the
Design of Instructional Material. Australian Journal of Educational Technology, 15(3),
222-241, 1999.
9. Madeira, M.; Diehl, E.; Verdin, R.; Wainer, R.; Fraga, T.: Learner Cognitive Modeling in
Intelligent Tutoring Systems for Distant Learning in the Internet. In: XXVIII
Interamerican Congress in Psychology. Chile, Santiago Del Chile, 2001.
10. Martinez, M.; Bunderson, V.: Building Interactive World Wide Web (Web) Learning
Environments to Match and Support Individual Learning Differences. Journal of
Interactive Learning Research, 11(3), 163-195, 2000.
11. Mcmanuns, T.F.: Individualizing Instruction in a Web-Based Hypermedia Learning
Environment: Nonlinearity, Advance Organizers, and Self-Regulated Learners. Journal of
Interactive Learning Research 11(3), 219-251, 2000.
12. Riding, R. & Grimley, M.: Cognitive Style and Learning from Multimedia Materials in
11- year Children. British Journal of Educational Technology, 30(1), 43-59,1999.
13. Riding, R. & Rayner, S.: Cognitive Styles and Learning Strategies. London: David Fulton
Publishers, 2000.
14. Ross, J. D.; Ross, C. M.: Test Ross of Cognitive Processes. São Paulo: Institute Pieron de
Psicologia Aplicada, 1997.
Automated Personalization of Internet News
Aditya V. Sunderam
2412 Harvard Yard Mail Center

[email protected]
Abstract. A systems approach to the automatic and adaptive person-

alization of Internet news is described. Implemented on the client side
as a lightweight, transparent software system, this approach is based
on implicit user feedback, thereby preserving privacy while avoiding the
constant recustomization needed in explicit schemes. The system con-
sists of two modules: (1) a profiling agent, which unobtrusively monitors
news reading patterns to track interest in different topics while acting as
a proxy server on the user’s computer; (2) and an action agent, which
uses the profile information to retrieve, filter, and present news articles.
A prototype of the system was implemented and evaluated over a two
week period. Precisions (the percentages of relevant articles returned by
the system) ranging from 60-95% were observed for profiles represent-
ing various combinations of interests. The system also responded very
well to simulated changes in user interests, returning rapidly increasing
numbers of articles relevant to newly developed interests.
1 Introduction
The Internet has the potential to be a highly effective and timely medium for
news, especially news that is personalized for individual users. Given its highly
interactive and customizable nature, in theory, users could construct their own
“newspapers” by assembling an ideal mix of articles from multiple websites.
However, creating such a personalized newspaper is currently impractical, due
to the sheer volume of news and the required expertise, tools, and computational
resources. Therefore, users must settle for generic front-pages that are created
by human editors at Internet news sites; readers scan these index pages and click
on articles of interest for in-depth reports. There are several major drawbacks to
this mode of newsreading. Manual selection of interesting articles is tedious and
time-consuming. Multiple sites must be scanned to ensure comprehensiveness.
Moreover, because news sites constantly update their index pages, scanning must
be frequent to avoid missing important articles. These drawbacks suggest that
new, automated approaches be pursued to make Internet news more effective.
A pragmatic scheme for the automated and adaptive personalization of Internet
news that addresses these issues is presented in this paper.

Automated Personalization of Internet News 349
1.1 The Basis for Automatic, Adaptive, Personalization
Generally, an individual’s interests determine the types of news articles that he

or she is likely to read. Automated approaches to news personalization may be
based on this premise; articles that a user recently found interesting could be
used to automatically gauge the relevance of new articles. Further, changes in
reading patterns could be used to detect and adapt to changing interests. For
Internet news read via a web browser, several parameters can be measured to
quantify the user’s degree of interest in an article. Combining these measures
with the characteristics of the articles themselves, a “profile” of the user could
be created, representing his or her interests. Such a profile could then be used as
a basis of comparison for future articles; articles that closely match the profile
are likely to appeal to the user.
1.2 Background
The system described herein draws upon the fields of software agents and infor-
mation filtering and retrieval (IF and IR). Agents are used within this project in
a heuristic and pragmatic manner; informal notions of inference, preference, and
persistence are utilized as in Letizia [12] and Apt [4]. Concepts from IF and IR
are used more explicitly, especially those employed in retrieval systems to select
relevant documents from a given set in real-time [1]. Specifically, news articles
are considered documents, characterized as in IF and IR systems by a set of
keywords K = {k1 , k2 , k3 , ..., kn }, and described by a corresponding frequency
vector: D = {d1 , d2 , d3 , ..., dn }, where di denotes the number of occurrences of
keyword ki . This project also uses a standard IR metric to express similarity
between two documents, viz. the angle, or cosine, metric,
n
xi yi
Similarity(X, Y ) ≡ cos θ = i=0 ,
| X || Y |
where x and y are the two document vectors, (the frequencies in each document
of all words in either document [6]). If each vector is visualized in n-dimensional
space, the angle between the vectors is small for similar documents and large for
dissimilar ones. Thus, higher values for cos θ imply greater similarity between
the documents.
1.3 Project Concepts
A key idea in this project is that the structure of “profiles” representing user
preferences can be designed in a manner that facilitates the selection of rele-
vant incoming news articles. The user profile is represented as a keyword-weight
vector, a set of numerical values indicating the user’s relative interest in topics
associated with corresponding words. For example, the profile of a user interested
only in politics would have high scores for words such as “Clinton”, “Congress”,
and “Republicans”, while words common to other topics would have lower scores.
350 A.V. Sunderam
The profile can thus be thought of as a composite representation of the user’s

ideal document (a canonical personalized newspaper), reflecting all the user’s in-
terests. Over time, the user’s interests and priorities may change. As this occurs,
the user’s profile should also change. A second key idea in this project is that
the profile is automatically maintained to be accurate and current by continually
monitoring reading patterns and adapting the profile as appropriate. Adaptation
is accomplished by comparing the user’s expected and actual level of interest in
an article and adjusting the profile accordingly.
Keyword Weight Keyword Frequency Keyword Weight

bush 4.52 president 13 bush 4.67
gore 3.68 + inauguration 9 gore 3.93
clinton 9.85 congress 16 clinton 9.91
president 7.60 New article User reads article; president 8.18
congress 8.71 (implicit) feedback congress 9.22
Article score = 9.657 (very high) is positive and
senate 5.79 senate 6.21
(high degree of similarity between reinforcing
filibuster 1.33 article and subset of profile) filibuster 1.28
legislation 2.68 Artcile scored Presented to user Profile updated legislation 2.75
Current User Profile inauguration 1.57
Updated User Profile
Fig. 1. News personalization schema
The concepts underlying the news personalization approach proposed in this

project are illustrated by the schematic example in Fig. 1. It shows the steady-
state profile of a user interested only in politics, the scoring and presentation
of the article to the user, and the eventual update of the profile with adjusted
scores and newly encountered keywords. In a different scenario, low-scoring ar-
ticles or those ignored by the user would result in attenuation of appropriate
keyword weights in the profile. The details of this process, including transparent
measurement of user interest, the scoring algorithm, profile maintenance, and
inclusion of new topics are discussed in the following sections.
1.4 Related Work
Among the many approaches to personalization, expert-system strategies are

the most automated. The Fishwrap and ZWrap systems developed at MIT were
successful implementations of such strategies and served as valuable foundations
for subsequent projects [2,3]. However, those systems were not adaptive; changes
in user interests were not automatically detected and reconciled. At the opposite
extreme are commercial systems, such as MyYahoo! [7] and MyExcite, which
are highly dependent on user input. These schemes tend to rely heavily on user
selection (from short, fixed menus) of news topics; at the portals, human editors
simply sort news articles into appropriate topic categories. In addition to being
too rigid, these approaches require the storage of personal preference data on
content-provider servers and lack the ability to adapt to changing interests. Most
other systems, including the popular PTV [9] and News Dude [5] services, also
require explicit user input.
2 Design and Implementation
This project proposes a new approach to adaptive, automated personalization

of Internet-based news that is implemented as an associated software system.
Several sources continually provide news on their web pages1 . Articles from
these sources typically include a headline, a byline, narrative text, pictures, and
embedded links. A user usually selects articles to read based on the headline and
the byline. Articles are read for varying lengths of time depending on interest in
the topic, length of the article, number of pictures, and other factors. The user
may also follow embedded links related to the article, indicating deeper interest
in the topic. The model proposed in this project is based on the premise that
observation of such behavior can be used to quantify the user’s interest in a given
article. This characterization is the foundation for selecting future articles.
2.1 System Design
The system was designed to consist of two software agents, small programs that
run transparently and unobtrusively on the user’s computer. The first is the
“profiling” agent, which monitors the user’s degree of interest in various topics.
The profiling agent characterizes articles based on keywords, as described earlier.
To estimate the user’s degree of interest in a given article, it uses measures of (1)
immediacy (time between seeing a headline and clicking on it), (2) detail (time
spent reading an article), (3) comprehensiveness (number of embedded links
followed), and (4) focus (percentage of topical articles read). These measures are
used to continually update the user’s profile, so that it always reflects the user’s
most current interests. As a result, when the user’s interests evolve or change,
the system adapts quickly and automatically to the new interests.
The second program is an “action” agent. The action agent uses the profile
to filter incoming news streams. Each new article is scored against the user’s
profile. News items likely to be of compelling interest to the user are presented
in a prioritized list. The profiling agent also monitors the user’s reaction to
the news presented by the action agent, in addition to monitoring news that
is read directly from the web. Both agents run on the user’s computer (i.e. on
the client side), where the profile is also stored. Therefore, content providers do
not have access to the user’s preferences, ensuring privacy. Further, analysis and
scoring of news articles against individual profiles is performed locally, thereby
distributing the computational load. Finally, client-side implementation allows
the user to obtain personalized news from multiple news sites and sources, in
contrast to server-side implementations which force loyalty to a single source.
1
At present, HTML-based webpages are assumed; it is expected that future XML
versions will assist the system and improve its efficacy.
352 A.V. Sunderam
2.2 Prototype Implementation

The action agent is implemented as a continuously running background process
on the client computer. One module within this program periodically (every 15
minutes) retrieves the front page of targeted news sites, by issuing HTTP GET
requests. The HTTP response is parsed to extract all the embedded URLs. Any
new URLs are piped to an analysis module that characterizes articles and scores
them against the user’s profile. This module first fetches the content of each new
article and parses the page by removing certain standard HTML tags (links,
javascripts, tables, images). From the remaining text, two vectors of length n,
K = {k1 , k2 , k3 , ..., kn } of keywords and A = {a1 , a2 , a3 , ..., an } of corresponding
frequencies, are created. Keywords appearing in the article but not in the profile
are set aside for later addition to the profile. Next, subvector P of the user profile
P is constructed, such that P = {p1 , p2 , p3 , ..., pn }, where pi is the profile weight
of keyword ki . The article score is computed as the cosine of the angle θ between
A and P . Pointers to high-scoring articles are then inserted into a custom-made
web page for later presentation to the user. New keywords encountered within
the article that are not in the profile are added to the profile (with a weight of
zero) after the article is scored. As the user reads more articles containing these
words, the profiling agent will appropriately update their weights. Note that
articles presented by the system, as well as those read directly from the Internet,
are included in this process. In this way, the system adapts automatically to (1)
newly developed interests in pre-existing topics and (2) user interest in newly
emerging news topics.
cnn.com
Retrieve New Fetch and Profile
front page score new
yahoo.com every 15 articles
URLs
minutes Update
Save Selections
URLs profile
Action Agent
Access and interest monitored Present selections,

observe level of interest Web
Pages from other sites simply passed through Appears as proxy browser
server to browser
Profile Agent
Fig. 2. Prototype Implementation
A schematic of the prototype implementation is shown in Fig. 2. In the

prototype, the profiling agent is positioned as an intermediary (on the client
computer) between the user’s browser and the Internet. The profiling agent acts
as a minimalistic web-proxy server and appears as such to the browser. This
scheme exploits the built-in proxy options of web browsers, thereby obviating
the need for software modifications and causing no disruptions to the user’s
browsing experience. The profiling agent reads in the current user profile and
executes an infinite loop, relaying HTTP requests and responses between the
browser and the Internet. For retrievals of interest (news from selected sites in
this case), it logs the reaction time rt (time between seeing a headline and clicking
on it), dwell ratio dr (time spent reading divided by article length), number of
embedded links followed lf , and rank rk (the order of selection of a given article
from a set of headlines). Note that the nature of the HTTP protocol facilitates
collection of this data through observation of the browser’s network connections.
Proxy servers, positioned between the browser and the web, can monitor both
the content and timings of all interactions and can thus transparently collect the
required measurements.
The data collected by the profiling agent is used to track the user’s interests
and changes therein. Low values of rt and rk, coupled with high values of dr and
lf indicate continued high interest or increased interest in the topic exemplified
by the article. Conversely, articles that are not viewed at all or are viewed for
a short duration indicate decreasing interest. To quantify the degree of interest,
an article score metric that indicates increasing, decreasing, or constant interest
is developed. It is computed as Score = c1rt×rk×dr×n
+ c2 × lf, where n is the length
of the article vector and c1 and c2 are adjustable constants. This implicit user
evaluation of the article is used to update the profile P = {p1 , p2 , p3 , ..., pm } as
follows. First A is scaled to A such that |A| = Score. For each aj ∈ A , such
that keyword kj is in the article and in the profile, pj is updated according to the
c3 ×pj +a
equation pj = c3 +1 j , where c3 is a constant. Informally, the relevant parts
of the profile are updated using a weighted average of the old profile and the
current article, which has been scaled to reflect the user’s degree of interest (as
measured by the implicit score for this article). Since the old profile represents
a historical measure of the user’s interest in the topic of the current article, this
weighted averaging process tracks the user’s degree of focus in this topic. Since
this scheme updates the user’s profile based on every article read, the profile is
always kept up to date.
3 Evaluation
Traditional measures [6] of IR and IF system performance were used to eval-

uate the prototype, since the selectivity of the system in personalizing news is
analogous to the retrieval of appropriate documents from a large set. The most
important performance metric is precision, defined as precision = {retr}∩{rel}
{retr}
where {retr} is the set of returned articles, {rel} is the set of all relevant articles
in the document set, and x represents the size of set x. Precision indicates the
probability that a retrieved document is relevant and is the most important mea-
sure because the primary goal of IF and IR systems is to return as many relevant
articles as possible. Another metric, which characterizes the proportion of rele-
vant articles that have been retrieved, is recall, defined as: recall = {retr}∩{rel}
{rel} .
In other words, recall is the ratio of the number of relevant articles returned to
the total number of relevant articles. Two other measures are sometimes used
to complement precision and recall. The first is f allout = {retr}∩{rel}
{rel}
, which
354 A.V. Sunderam
is the fraction of non-relevant articles returned. The second is the density of

relevant articles in the entire document set: generality = {rel}
{all} , an inherent
characteristic of the document set rather than a true performance measure. Gen-
erality provides an estimate of topic distribution in a given document set.
3.1 Methodology
During live operation, the system retrieves and filters articles for a user who
then provides implicit feedback through his or her reading patterns. In order
to simulate this process for several hypothetical users and evaluate the system,
the following methodology was adopted. First, a document set consisting of
3695 articles was assembled by continuously harvesting newly generated articles
from two popular news sources, cnn.com and dailynews.yahoo.com. As an
initialization process, the document set was analyzed to determine the list of
common words. A base profile was then created, consisting of all “content” words,
i.e. words in the document set that were not in the list of common words. An
initial weight of zero was assigned to all words in the profile. For the experimental
document set, the profile consisted of 25,000 words.
Table 1. Article counts and distribution by topic
Topic Russia Stocks Tennis Politics MidEast Total(5 topics) Entire set
No. of Articles 135 239 73 166 85 698 3695
Generality 3.65% 6.46% 1.98% 4.49% 2.30% 18.9% 100%
Five news topics were chosen for the evaluation process: Russian affairs, fi-
nancial news, tennis, domestic politics, and the Middle East. Table 1 shows the
number of articles corresponding to each topic in the document set of 3695 ar-
ticles, along with generality measures for each. Profiles were then created for
multiple users with one, two, and three interests and with a change from one in-
terest to another, by using “training articles”, simulating a new user’s selection of
articles in a topic of interest; in live operation, the profiling agent would accom-
plish this through observation of browsing patterns.. Training articles relevant to
a given user’s interest(s) were selected from the document set and assigned val-
ues for each of the feedback metrics. Realistic ranges for these metrics (reaction
time, rank, dwell ratio, and links followed) were obtained through empirical ob-
servation and by querying real users. Simulated values for experimental profiles
were then chosen randomly from within those ranges. For each of the hypotheti-
cal users, profiles were created (using the scoring formula developed in Sect. 2.2)
with 1, 5, 10, 15, 20, 25, 30 and 35 training articles. These profiles were then
used to filter the entire document set for articles relevant to each interest, com-
bination of interests, or change in interests. Precision and recall were computed
for each set of returned articles. The number of articles returned for a given test
profile ( {retr} ) ranged from 15 to 32 and was determined by the generality
of the topic(s) in the document set.
3.2 Results
The results of these experiments are very encouraging. Experiments using 35
training articles exhibited precisions in the range of 79-100%, and those using as
few as 5 training articles exhibited precisions of 60-95% in most cases. Values of
recall were mostly in the 60-70% range, and fallout was extremely low (less than
1%) in all cases; these positive results were obtained despite relatively low levels
of generality (ranging from 2% for tennis articles to 6% for articles concerning
the stock market). For space reasons, the remainder of the discussion concerns
only precision, the most important retrieval metric.
100.0 100.0
80.0 80.0
Precision (percent)
Precision (percent)
60.0 60.0
politics+mideast
politics politics+stocks
mideast russia+mideast
stocks russia+stocks
russia tennis+mideast
tennis tennis+stocks
40.0 40.0
20.0 20.0
0.0 0.0
0.0 10.0 20.0 30.0 40.0 0.0 10.0 20.0 30.0 40.0
Number of training articles Number of training articles
(a) One isolated interest (b) Two simultaneous interests
Fig. 3. Precision vs. Training Articles
Observed values of precision for varying numbers of training articles are

shown in Fig. 3a and 3b for each of the five news topics treated individually,
and for topics taken two at a time – representing users with one and two simul-
taneous interests respectively. Precisions are uniformly high in all cases, except
in a few isolated instances when only one training article was used. This indi-
cates that the filtering algorithm, though simple, is very effective. Results for
profiles simulating three or more simultaneous interests are also very good, as
may be seen in Fig. 4a. These results are particularly significant since they are
representative of the multiple, varied interests of the average news reader.
Fig. 4b plots precision versus the number of training articles when a hypo-
thetical user switches from one interest to another, measuring the system ability
to automatically adapt. The graph shows the selection of residual, undesired
articles about the original topic (dashed lines, unfilled symbols), as well as the
selection of desirable articles about the new topic (solid lines, filled symbols). The
general trends are clear; as the number of training articles (for the new topic)
increases, greater precision is observed in the number of “new-topic” articles
selected. Simultaneously, the numbers of “old-topic” articles retrieved decrease.
The rate at which new topic precision increases (and old topic precision de-
356 A.V. Sunderam
100.0 100.0
80.0 80.0
Precision (percent)
Precision (percent)
60.0 60.0
politics+mideast+tennis politics->mideast
politics+stocks+tennis oldtopic: politics
russia+mideast+tennis politics->stocks
russia+stocks+tennis oldtopic: politics
politics+russia+mideast russia->mideast
oldtopic: russia
stocks+russia+mideast
russia->tennis
40.0 40.0 oldtopic: russia
20.0 20.0
0.0 0.0
0.0 10.0 20.0 30.0 40.0 0.0 10.0 20.0 30.0 40.0
Number of training articles Number of training articles
(a) Three simultaneous interests (b) Adaptation: Topic1→Topic2
Fig. 4. Precision vs. Training Articles
creases) is high, indicating that the system adapts quickly and effectively. For
example, three of the five new topics achieve precisions of 60% or more with 20
training articles, while all achieve precisions of 80% or more with 35 training
articles. Overall, these graphs are consistent with the expected behavior of the
average user, who develops new interests not instantaneously but over a period
of several days.
4 Conclusions
An agent-based software framework for automatic, adaptive personalization of
Internet-based news was designed and implemented. By standard performance
measures, the system achieves a high level of efficiency and is effective in adaptive
filtering. The major strength of the system is its use of implicit relevance feedback
to determine user interests and changes therein. This project also emphasizes
other important issues. Privacy is achieved by a client-side implementation, and a
proxy-based profiling agent can be deployed with minimal effort and disruption.
During operation, the system is unobtrusive and transparent, and consumes
negligible resources.
The system has some limitations. The current version requires the analysis
of a few articles before it can handle entirely new topics. Thus, when presented
with news concerning an unusual, unexpected event (i.e. articles containing pre-
viously unencountered words), the first few articles may not be forwarded to
the user. However, if the user profile implies interest in this new topic or if
the user reads a few articles on this topic through direct browsing, the system
will adapt and subsequent articles will be presented. Techniques to detect and
include new topics, alternative similarity detection algorithms, and proactive
notification mechanisms are among the ongoing and future enhancements being
considered.
According to recent statistics, the Internet encompassed about 147 million

hosts in January 2002 [11] and is doubling in size every year [8]. With this growth,
the information overload problem is only worsening. Addressing this situation in
the context of text-based news is an important step forward. News represents a
large volume of constantly changing information, and individuals have selected
subsets of interests, which themselves change over time. This project has shown
that a reasonable solution can be developed in such an arena.
Acknowledgements
I would like to thank Dr. Walter Bender for giving me a summer internship
opportunity in the MIT Media Lab and for introducing me to news editing
systems. I am also indebted to the RSI AlumReader program for feedback and
comments on earlier versions of this paper.
References
1. Baeza-Yates, R., and Berthier Ribeiro-Neto. “Modern Information Retrieval.”
Addison-Wesley, 1999.
2. Bender, Walter. “Read All About It in the Daily You.” Communicating Busi-
ness, Forward Publishing, London, United Kingdom, 1995. Also available at:
http://nif.media.mit.edu/papers/forward.html.
3. Gruhl, D. and W. Bender. “A New Structure for News Editing.” IBM Systems
Journal 39, nos. 3&4, pp. 569–588 (2000).
4. Shearin, Sybil, and Henry Lieberman, “Intelligent Profiling by Example”, Pro-
ceedings of the International Conference on Intelligent User Interfaces (IUI 2001),
Sante Fe, NM, January 2001.
5. Hirsh, Haym, et al. “Learning to Personalize.” Communications of the ACM, 43(8),
pp. 102-106, August 2000.
6. Losee, Robert M. “Text Retrieval and Filtering: Analytic Models of Performance.”
Kluwer Academic Publishers, Boston, MA, 1998.
7. Manber, Udi, et al. “Experience with Personalization on Yahoo!” Communications
of the ACM, 43(8), pp. 35-40, August 2000.
8. Salus, Peter. “Internet Growth Rate.” Matrix News, 10(3), March 2000. Also avail-
able at: http://www.mids.org/mn/1003/growth.html.
9. Smyth, Barry and Paul Cotter. “A Personalized Television Listings System.” Com-
munications of the ACM, 43(8), pp. 107-111, August 2000.
10. Tryon, R. and D. E. Bailey. “Cluster Analysis.” McGraw-Hill, 1970.
11. Internet Software Consortium, “Internet Domain Survey, January 2002: Number of
Hosts Advertised in the DNS”, http://www.isc.org/ds/WWW-200201/index.html,
January 2002.
12. Lieberman, Henry. “Letizia: An Agent That Assists Web Browsing”, Proceedings
of the 1995 International Joint Conference on Artificial Intelligence, Montreal,
Canada, August 1995.
Conceptual Modeling
of Personalized Web Applications
Silvia Abrahão1 , Joan Fons1 , Magalı́ González2 , and Oscar Pastor1

1
Department of Information Systems and Computation
Valencia University of Technology
Camino de Vera s/n, P.O. Box: 22012, E-46020 - Valencia, Spain
{sabrahao, jjfons, opastor}@dsic.upv.es
2
Department of Informatics and Electronics
Catholic University of Asuncion
Tte. Cantaluppi y Tte. Villalón. C.C. 1683- Asuncion, Paraguay
[email protected]
Abstract. The demand for web applications that take into account the
different needs and interests of the users has been increasing. Personal-
ization can involve a wide-range of approaches and techniques to design
the end-user experience. This paper discusses the use of conceptual mod-
eling techniques in a software production process for the design of per-
sonalized web applications. This process is driven by an Object-Oriented
Web-Solutions Modeling approach (OOWS) that properly captures the
specific characteristics of web applications. We place special emphasis on
the primitives of a navigational model that provides personalization pat-
terns to capture and represent the semantics of this kind of application.
1 Introduction
The World Wide Web (WWW) has introduced a new software environment
characterized by rapid changes in technology and the need for defining personal-
ized applications for specific user requirements. Adaptation is the way to define
rules that allow web applications to give an appropriate response in accordance
with user profiles, navigation history, target devices, network connections, etc.
E-commerce applications are representative examples of systems where the adap-
tation to new environments and requirements is a critical factor. In this context,
automatic software production environments have become a significant way to
accelerate the development process and reusability capacities. Thus, today ther is
a need for dynamic adaptation of the applications with regard to the preferences
of the user.
Nowadays, the approaches that address some kind of personalization vary
widely [1]: from simple page generation strategies to complex content-based
prediction systems, pattern recognition of user behavior, machine-learning al-
gorithms, and data mining. Most of these approaches consider personalization

Research supported by the CICYT Project, with ref. TIC2001-3530-C02-01 and the
WEST Project (CYTED Program), subprogram VII.18)

Conceptual Modeling of Personalized Web Applications 359
issues from an implementation point of view. According to [12] the current tech-
niques for collecting user information are explicit profiling, implicit profiling and
legacy data. The most used widely techniques for analyzing information are rule
based and filtering techniques.
However, we consider the use of conceptual modeling techniques to be impor-
tant in the appropriate treatment of the inherent complexity of web applications.
In [3] and [7], customization issues in web application development are discussed
pointing out the need for modeling methods to develop customizable web ap-
plications. The WSDM method [8] is an audience-driven approach that defines
information objects based on the information requirements of the users to de-
velop a web application. In [4] the authors shows how to specify personalized
web applications by defining different scenarios according to user profiles or pref-
erences. In [2], a modeling language for designing web applications is proposed.
It provides a personalization model where users and groups can be explicitly
specified in the structure schema of the information.
In the same way, our proposal provides a contribution in this context claiming
that conceptual modeling should include personalization features for the devel-
opment of sound and robust web applications. In the context of the OO-Method
project [9,10], efforts have been oriented towards the development of a new model
to enrich this object-oriented software production method with the required ex-
pressiveness to specify web applications. This model proposes a set of high level
abstraction primitives to define the navigation semantics of applications. We fo-
cus on the integration of navigational design and conceptual modeling, which
together could be used as input for an automatic software production environ-
ment. All this information is used to provide precise methodological guidance
for going from the Conceptual Space to the Solution Space (represented by the
final software product). The objective of this paper is to show how the OOWS-
Modeling approach supports the specification of personalized web applications.
This paper is organized in four sections. Section 2 gives a general overview of
the OOWS approach. Section 3 presents the primitives used to express specific
features in conceptual modeling for customized web applications. Finally, Sect. 4
provides the conclusions and work in progress.
2 The OOWS-Modeling Approach
OOWS [11] is an extension of the OO-Method [9,10] approach to develop web

applications. The OO-Method is enriched with a navigational model to cap-
ture navigational semantics for different audiences. The OOWS approach com-
prises two major steps: Specifying the System and Developing the Solution. In
the first step, a conceptual model of an application is built in terms of the
problem space. At first, this step includes requirement collection using a Use
Case [6] approach. Then, the conceptual model is built using an Object Model
(object structure), a Dynamic Model (valid object lives and object interaction),
a Functional Model (behavior description) and a Navigational Model (navigation
structure). These models describe the object society from four complementary
360 S. Abrahão et al.
points of view within a well defined OO framework. In the Developing the So-
lution step, a strategy for component generation to integrate the solution (the
final software product) is defined. In this step, a functional equivalent to the
specification web-based application can be obtained in an automated way.
2.1 The Navigational Model
The Navigational Model allows for the definition of a navigational view over the
Object Model. It captures the navigation semantics of a web application using
a UML-like [5] notation.
The navigational model is essentially composed of a Navigational Map that
represents the global view of the system for an agent (potential end-user). It is
represented by a directed graph in which the nodes are the Navigational Contexts
and the arcs are the Navigational Links defining valid navigation paths.
A Navigational Context allows to define the content and presentation of a
user interaction. It represents the point of view that this user has over a subset
of the object model. It is composed by Navigational Classes and Navigational
Relationships. A navigational class defines which attributes and operations of an
object model class will be shown for the given agent. Navigational relationships
connect navigational classes and they describe valid navigation paths over object
relationships.
There are two kinds of navigational contexts: exploration contexts and se-
quence contexts. The first one are contexts that can be reached at any moment,
independently of the current context. Sequence contexts can only be reached
by following a predefined sequence of navigational links. Finally, presentation
patterns can also be specified to “format” the output presentation.
In OOWS it is possible to access the “connected user” with the application
in the navigational model by using the Active Agent primitive. With this primi-
tive, it is possible to define expressions that depend on part of the information
from this user. For example, in a population filter one can use the expression
#RegisteredUser#.ShoppingBasket to access only the ShoppingBasket for the
connected user (assuming there is a RegisteredUser class, a ShoppingBasket
class and an aggregation relationship between them which is defined in the Ob-
ject Model).
In addition, the Active Agent has a predefined Session property that auto-
matically collects information about user interactions with the application (such
as the navigation paths followed, the objects selected, the operations performed,
etc.) that can also be used to construct dynamic expressions.
3 Designing Personalized Web Applications
In the conceptual modeling phase of a web application [3,4], personalization can

be performed at three levels: the information content level (which information
is available for a given user interaction), the presentation level (how this in-
formation is shown) and the navigation structure level (navigation structure).
Conceptual Modeling of Personalized Web Applications 361
Personalization for each of these levels can be seen from two different aspects:
static and dynamic. Static personalization is fully defined at design time. Dy-
namic personalization depends on a runtime condition. Additionally, two levels
of granularity can be defined in accordance with the applicability to a group of
users (user profile) or to an individual user. Another characteristic of the per-
sonalization is the environment adaptation, changing or adapting the application
interface depending on the target resources (device, location, network connec-
tion, etc.).
Through its abstraction primitives the OOWS approach allows static person-
alization with respect to the content, presentation and navigation levels. For dy-
namic personalization it only allows personalization with respect to the content
and navigation level. It is possible to consider individual or group granularity.
3.1 Static Personalization

Navigation static personalization is obtained by linking a navigational map with
an agent, predefining its navigation structure at design time. Navigational con-
texts allow the content personalization for this agent by specifying its naviga-
tional classes (attributes and operations), navigational relationships and popu-
lation filters. Presentation patterns specified for a navigational context define
the static personalization for the presentation level.
Group personalization is achieved by providing a navigational map for each
user type (agent) detected in the navigational model. Therefore, agents of this
type share all the characteristics of this map. However, defining individual per-
sonalization is not allowed.
3.2 Dynamic Personalization

Dynamic personalization allows for the definition of rules and characteristics
at the conceptual level that must be evaluated at runtime depending on the
current interactive user. OOWS proposes the Active Agent primitive for this
purpose, enabling access to the “active user” and the user’s Session (collection
of interactions performed by the user with the application).
Content dynamic personalization is represented by defining population filters
over the navigational classes involving some properties of the current interactive
user (for example, accessing the user’s ShoppingBasket or showing the user’s
navigation path to the current context using the user’s Session property).
We consider that there is no need to specify navigation dynamic personal-
ization. However, the navigational model can implicity capture this semantic
enabling each user to define a subset of their navigational map. The applica-
tion could even automatically show the user’s most frequent navigational paths
(without disabling all other navigation possibilities).
362 S. Abrahão et al.
4 Conclusions
In this paper, we have presented how the OOWS approach deals with the de-
sign of personalized web applications. We argue that conceptual modeling of this
kind of applications is a practice from a software engineering point of view that
allow us to design applications which are more maintainable and extensible. Cur-
rently, we are extending our modeling language with new customization patterns
that permit the construction of flexible web applications. We are also defining
translation patterns for different target architectures, taking into account the
personalization aspects captured at the conceptual level. Future work involves
the extension of the modeling language to specify security features and integrity
validation of the navigational model with respect to the other elements of the
conceptual model.
References
1. Special Issue on Personalization. Communications of the ACM, 43(8), 2000.
2. Bonifati A., Ceri S., Fraternali P., and et al. Building Multi-device, Content-
Centric Applications Using WebML and the W3I3 Tool Suite. In 19th International
Conference on Conceptual Modeling (ER’00), Salt Lake City, USA, 2000. Springer-
Verlag.
3. Kappel G., Retschitzegger W., and Schwinger W. Modeling Customizable Web
Applications - A Requirement’s Perspective. In Proc. of Kyoto International Con-
ference on Digital Libraries: Research and Practice, November 2000.
4. Rossi G., Schwabe D., and Guimaraes R. Designing personalized web applications.
In Proc. of the WWW10, Hong Kong, May 2001.
5. Object Management Group. Unified Modeling Language Specification Version 1.4
draft. Technical report, www.omg.org, February 2001.
6. Jacobson I., Christerson M., Jonsson P., and Overgaard G. Object Oriented Soft-
ware Engineering, a Use Case Driven Approach. Addison -Wesley. Reading, Mas-
sachusetts, 1992.
7. Koch N. and Wirsing M. Software engineering for adaptive hypermedia applica-
tions. In 3rd Workshop on Adaptive Hypertext and Hypermedia, 2001.
8. De Troyer O. and Leune C. WSDM: A user-centered design method for Web sites.
In Proc. of the 7th International World Wide Web Conference, 1997.
9. Pastor O., Insfrán E., Pelechano V., Romero J., and Merseguer J. OO-Method: An
OO Software Production Environment Combining Conventional and Formal Meth-
ods. In 9th Conference on Advanced Information Systems Engineering (CAiSE’97),
pages 145–159, Spain, June 1997. Springer-Verlag.
10. Pastor O., Gómez J., Insfrán E., and Pelechano V. The OO-Method Approach for
Information Systems Modelling: From Object-Oriented Conceptual Modeling to
Automated Programming. In Information Systems, volume 26(7), pages 507–534,
2001.
11. Pastor O., Abrahao S., and Fons J. J. Object-oriented approach to automate web
applications development. In 2nd International Conference on Electronic Com-
merce and Web Technologies (EC-Web’01), Munich, Germany, Septiembre 2001.
Springer-Verlag.
12. IBM High-Volume Web site team. Web Site Personalization. January 2000.
On Evaluating Adaptive Systems for Education
Rosa Arruabarrena, Tomás A. Pérez, J. López-Cuadrado, J. Gutiérrez,

and J.A. Vadillo
Dpto. de Lenguajes y Sistemas Informáticos, Universidad del País Vasco (UPV-EHU)

Apdo. 649. 20080 San Sebastián, Spain
{jiparsar,jippefet,jiplocuj,jipgusej,jipvazoj}@si.ehu.es
http://www.ji.si.ehu.es/groups/hyper/
Abstract. In this paper we have gathered some interesting techniques to evalu-

ate educative systems. Our main purpose is to evaluate HEZINET, an adaptive
hypermedia system for education available commercially. Therefore, we also
include a system evaluation plan to identify which types and techniques of
evaluation will be needed in order to accomplish it.
1 Introduction
HEZINET is an Adaptive Hypermedia System [1] to learn Basque language. It auto-

matically compiles exercises adapted to the concepts covered by the nodes that the
student has already visited using the Item Response Theory and AI techniques [2].
The system has been validated and more than 2000 copies have been sold by now.
However, we think that we can improve HEZINET. In order to do that, we have
planned to perform different evaluations. At the moment we consider two uses of the
system that we want to differentiate. On the one hand, the system is used as a distance
learning system available on Internet to the subscribed students (we will call it
HEZINET-D for short). On the other hand, it is also being used in intranets by stu-
dents in Basque academia as an extra support to the traditional Basque classes (we
will call HEZINET-C for short).
Until now, we have already made a usability test using surveys. The conclusions
lead to some changes of location of functionality in the interface and to include some
other contents to adapt to very beginners in Basque [3]. Now, together with the com-
pany that sells HEZINET we are interested in evaluating the effectiveness of the sys-
tem and the advantages of the adaptation it provides [4].
This paper presents a wide study of strategies for software validation, types of
evaluation to be considered when assessing computer assisted learning and concrete
techniques to perform them. More detailed techniques are presented in [5]. A detailed
overview on techniques that focus mainly on adaptation can be found in [6]. Finally,
we present a plan to evaluate both versions of HEZINET.

364 R. Arruabarrena et al.
2 Validation Strategies
Software validation focuses on demonstrating that programs meet their requirements.

We call the processes that check this validation strategies. In [7], several validation
strategies are presented, such as (a) Proofs of Correctness, that consist of validating
and verifying programs using formal proofs of correctness; and (b) Criterion Based
Validation, that is commonly used in software engineering (here the system is consid-
ered successful if it displays no major inadequacies within its intended application
environment); (c) Expert Review, in which a person (the expert) observes the whole
system, parts of it and/or its behaviour and identifies its deficiencies; (d) Certification,
that is based on techniques for software quality and deterministic rules; (e) Empirical
Proofs, that use experience of user with the system as a measure.
3 Evaluation
The evaluation of an educational system is a process of data gathering to determine

the quality or the value of the instruction, its weaknesses and its strengths [8]. In
[7,9,10] different dimensions of evaluation of such systems are presented. Two of
them are the Evaluation Paradigms, that monitor what the evaluation goals are and
the Evaluated Elements that classify those parts of the educational system to be as-
sessed. Both are presented in detail in the following paragraphs.
There are mainly three evaluation paradigms: (1) Formative Evaluation, which is
oriented to detect weak points at early stages of the development of the product in
order to modify and improve it; (2) Summative Evaluation, that addresses punctual
assessments of effectiveness of the finished application. The evaluation parameters
are goals, benefits and costs; (3) Integrative Evaluation, which measures the level of
integration between the educational software and other learning materials and meth-
ods into the context overall.
If one looks at the elements being observed, three kinds of evaluation could be dis-
tinguished: (1) Internal Evaluation that implies that the system itself is assessed (for
instance, different components of the system can be evaluated, such as the architec-
ture, the intermediate processes, as well as the behaviours of and the relations among
the previous); (2) External Evaluation, that means that only elements out of the learn-
ing system are assessed, such as, student achievements or the affective impact of the
system; (3) Global Evaluation, that evaluates components and impact on users.
4 Evaluation Techniques
Evaluation techniques are concrete methods to carry out the validation of the system.
Four most common ones are [10,11]: (1) Comparison: is based on comparing the
characteristics of the system versus some standard or other system (for example, there
are methods, such as golden standard, theoretical corroboration, empirical corrobo-
On Evaluating Adaptive Systems for Education 365
ration, duplication, Turing test, sensitivity analysis and benchmarking). (2) Contact
with users whose objective is to collect data about how the user interacts with the
system, his/her behaviour and attitude (for instance, one can use interviews, surveys,
questionnaires, focus groups, nominal groups, and tests). (3) Data analysis, that con-
sists on reviewing, studying and assessing groups of data about certain characteristics.
The information collected, processed and reported should be systematically reviewed
and handled, and if any errors found, those should be corrected [12]. (4) Pilot testing
which involves studying the performance of the system with potential end users.
Depending on the number of users that are evaluated, we distinguish one-to-one test-
ing, small-group testing, field testing and beta testing.
5 Evaluation of HEZINET
To evaluate the effectiveness of both versions of HEZINET and taking into account
[6,13,14], we have established a plan to: (1) improve the tools associated to the sys-
tem; (2) assure that the adaptation results in better performance; (3) check the impact
of the system as a motivating element of learning; (4) attract users affectivity and
retain the interaction with them as long as possible; (5) achieve excellent performance
of students from the knowledge acquisition point of view, either using the system
alone or transferring this acquisition to other contexts different from the computer;
(6) guarantee that the system does not mislead students, independently from their
computational skills, and that students can locate the desired contents either in
HEZINET or in other related pedagogical material available.
We planned to perform several interface evaluation loops (internal summative
evaluation) to measure different features. Although it could be considered formative,
because the results will improve the product, we consider that the product has been
already developed, and that we are working on a new version of the system.
The plan is to create two groups of students (those involved with HEZINET-D and
the ones with HEZINET-C). First, we check how students use the elements of the
interface by means of (1) questionnaires detect weak points of the interface; (2) con-
tacts with experts in Human Computer Interaction for alternative solutions to the
identified problems and (3) empirical corroboration to choose the best prototype by
means of (4) pilot testing on both populations. Therefore, our aim is to check if the
new design solves the problems detected without interfering the rest of users. Once
the new interface is validated, and if it were required, it will be integrated into a new
version.
To assess the affective impact of HEZINET, we will use open-ended question-
naires, in which users should indicate which elements are more helpful and pleasant.
This evaluation (external and summative) could result in a reorganisation of the con-
tents or in an inclusion of more multimedia courseware.
To evaluate the system’s effectiveness, we are interested in finding the differences
among the learning results obtained from three populations: students from HEZINET-
D, from HEZINET-C and from those that learn attending regular classes at school but
without any tool of this sort. In order to conduct this (external summative) evaluation,
366 R. Arruabarrena et al.
we should analyse the user model. We will compare the data obtained from each
population to check if any type of correlation among them exists. Concretely, when-
ever it would be possible, we will handle data from the system logs, from lists of
results in tests completed at schools, in an anonymous way, and from the list of marks
obtained in official tests of linguistic capability. As a result of the analysis, the hy-
potheses that correspond will be established.
To verify the suitability of HEZINET-C within an institution like an academy of
languages, it is important to assess whether the educational methods include the ap-
plication as one more tool inside the classes or as an additional supporting service.
We will conduct an (external integrative) evaluation using (1) interviews to students
and teachers and (2) data gathering by direct observation.
We consider that it is very complicated to evaluate the integration of HEZINET-D
students and the learning material they may handle. Nevertheless, we could give some
advice about the work environment for a good learning.
6 Conclusions
Although it is not usual to evaluate systems after they are already in use, the goals of
the education can vary and, perhaps, it is necessary to make certain adjustments to the
initial proposed topics.
It is important not only to evaluate the system but also to value the integration
within the context where the learning will take place. Some synergies can be obtained
thanks to the integration, which can do the system’s potential be even higher than the
expected one if the application were considered only separately.
In this paper we have presented a wide study of types of evaluations that appear in
educational systems. In the same way, it focuses on, possibly, the most interesting
elements to conduct an evaluation of an educational adaptive hypermedia system. We
have also discussed different goals we can consider and the system elements on which
centre the study. Moreover, we also report about the techniques to be used to obtain a
successful evaluation. Among these techniques, types of evaluations and strategies we
have picked out the most interesting ones to evaluate HEZINET.
Future work, as well as conducting the proposed evaluation plan, includes projects
to adapt HEZINET to other languages such as Spanish and French, since the system is
successful.
References
1 Brusilovsky, P. (2001). “Adaptive Hypermedia”. User Modeling and User-Adapted Interac-

tion.Special Issue on: Ten Year Anniversary Issue 11(1/2): 87-110
2. Pérez, T.A., López, R., Gutiérrez, J., González, A.: Learning Basque in a Distance-adaptive
way. Computers and Education in the 21st Century. Kluwer Academic Publishers,
Dordrecht, The Nederlands (2000) 251-262
On Evaluating Adaptive Systems for Education 367
3. Villamañe, M., Gutiérrez, J., Arruabarrena, R., Pérez, T.A., Sanz-Lumbier, S., Sanz-
Santamaría, S., López-Cuadrado, J., Vadillo, J.A.: Use and Evaluation of HEZINET; A Sys-
tem for Basque Language Learning . In: Lee, C.-H., Lajoie, S., Mizoguchi, R., Yoo, Y.D.,
du-Boulay, B. (eds). Proceedings of the 9th International Conference on Computers In Edu-
cation (ICCE/SchoolNet2001), Seoul (South Korea) (2001)
4. Karagiannidis, C. and D. G. Sampson (2000). Layered Evaluation of Adaptive Applications
and Services. International Conference on Adaptive Hypermedia and Adaptive Web-Based
Systems: AH2000, Trento (Italy), Springer-Verlag Berlin Heidelberg.
5. Arruabarrena, R., T. A. Pérez, et al. (2001). Compendio de técnicas para evaluación de
sistemas hipermedia adaptativos. Simposium Internacional de Informática Educativa, SIIE,
Viseu, Portugal.
6. Weibelzahl, S.: Evaluation of Adaptive Systems. UM2001: Proceedings of the eighth Inter-
national conference on User Modeling, Springer: Berlin (2001)
7. Mark, M.A., Greer, J.E.: Evaluation Methodologies for Intelligent Tutoring Systems. Inter-
national Journal of Artificial Intelligence in Education, 4(2/3) (1993) 129-153
8. Tessmer, M.: Planning and Conducting: Formative Evaluations. Kogan Page Limited. Lon-
don (1993)
9. Draper, S.W., Brown, M.I., Henderson, F.P., McAteer, E.: Integrative evaluation: an emerg-
ing role for classroom studies of CAL. Computers and Education, 26, (1-3) (1996) 17-32
10 Murray, T.: Formative Qualitative Evaluation for “Exploratory” ITS Research. International
Journal of Artificial Intelligence in Education, 4(2/3) (1993) 179-207
11 Harvey, J.: LTDI Evaluation Cookbook. Learning Technology Dissemination Initiative.
Edinburgh (1999)
12 Frechtling, J., Sharp, L.: User-Friendly Handbook for Mixed Method Evaluations. NSF,
Arlington, VA (1997)
13 Chin, D. N.: Empirical Evaluation of User Models and User-Adapted Systems. User Model-
ing and User-Adapted Interaction 11(1-2) (2001) 181-194
14 Shute, V.J., Regian, W.: Principles for Evaluating Intelligent Tutoring Systems. Interna-
tional Journal of Artificial Intelligence in Education, 4(2/3) (1993) 245-271
Recommending Internet-Domains
Using Trails and Neural Networks
Tobias Berka, Wernher Behrendt, Erich Gams, and Siegfried Reich
Salzburg Research - SunTREC

Jakob-Haringer-Strasse 5/III
5020 Salzburg
AUSTRIA
{tberka, wbehrendt, egams, sreich}@salzburgresearch.at
Abstract. This paper discusses the use of artificial neural networks,

trained with patterns extracted from trail data, as recommender systems.
Feed-forward Multilayer-Perceptrons trained with the Backpropagation
Algorithm were used to assign a rating to pairs of domains, based on the
number of people that had traversed between them. The artificial neural
network constructed in this project was capable of learning the training
set to a great extent, and showed good generalizational capacities.
1 Introduction
In today’s information society, people are faced with the problem of navigating
information spaces every day. This creates a need for effective navigational aids.
Recommender systems provide means for assisting users in the decision mak-
ing process (for a discussion of recommender systems see [5]). These systems
are the technical response to the fact that we frequently rely on other peoples’
experience when confronted with a field where we have little or no knowledge.
It is a recent development to view the process of navigation in the Internet not
as an isolated activity of a single user, but to make the combined knowledge of
individual users available to others.
The notion of a trail is an established concept in the field of hypertext navi-
gation. A trail is a sequence of trailmarks, each consisting of a node (representing
a document), the activity performed by the user and other properties such as
time and duration.
T railmark = (node, activity, time, duration, user)
For a concise definition of trails and trail-based systems see [3].

The recommender system described in this article has been developed for the
Trailist research project, as proposed in [4], which aims at assisting the user in
navigation with a variety of trail based systems. These system are unified in the
TrailBlazer framework, an agent based platform for integrating various compo-
nents addressing trail assessment, recommending and linking the platform into
the users’ applications. We used an early stage of this framework for prototyping.

Recommending Internet-Domains Using Trails and Neural Networks 369
If we focus on the documents in a set of trails, they form a “universe of nav-

igational behaviour” for the respective community, consisting of n documents.
These documents reside in k Internet-domains. For our purpose of developing a
recommender system which is handling the “What’s Related” problem from a
trail based approach, we have a significant reduction of complexity if we reduce
the trail induced dependencies between documents to dependencies between do-
mains (since k n). Once these domain dependencies have been established we
can reduce the relationship between two documents, which are to be compared,
to the domains, in which they are residing, thus increasing the probability of
hitting a part within the knowledge base.
We take a somewhat different approach of constructing a recommender sys-
tem than the classical collaborative filtering algorithms (see e.g. [1]) by using
Artificial Neural Networks, or ANNs, to extract information about rules in trails.
This approach utilizes a characteristic property of ANNs, which is the distributed
encoding of information in the weights of the connections. The hypothesis is that
the ANN will be able to generalize over the set of domain pairs, thus being ca-
pable of rating and recommending appropriate target domains.
2 Using ANNs
for Comparative Internet-Domain Name Rating
The ANNs used for our experiments were all feed-forward Multilayer-Perceptrons
trained with the Backpropagation Algorithm (as proposed in [6]). They were
designed and trained using the Stuttgart Neural Net Simulator (SNNS v 4.2, see
http://www-ra.informatik.uni-tuebingen.de/SNNS/).
We used proxy access logs as an initial source of trail data. We extracted
all successful GET accesses of the research and development team of an IT
enterprise and split them into (anonymous) user trails. These user trails were
then analyzed in order to generate training data as described below:
If N (d1 , d2 ) is the number of times the domain names d1 and d2 appear as
neighbors in a trail, then the relation rating r computes as follows:
n
1
r(d1 , d2 ) = 1 − ,
1 + N (d1 , d2 )
where n has to be adjusted to sufficiently stretch the output interval for a better
differentiation between the results. If the function s assigns the binary encoding
to every pair of domain names, a learning task L is defined as follows:
L = ( (s(d1 ), s(d2 )) , r(d1 , d2 ) )

The learning behaviour of the ANN suggests monthly re-training update
steps. Since the training sets represent a statistic evaluation of the trail data
of an observation period, too short update steps will not allow the generation
of representative training sets. Too long update steps would result in too large
training sets, which implies unacceptable learning times. The ANN was always
370 T. Berka et al.
initialized randomly before re-training to avoid that the algorithm gets trapped
in local minima. We extracted some 30,000 training tasks per monthly log file.
After training with the training set generated from one month, the ANN correctly
classified some 86.5% of this training set. The same net correctly classified as
much as 84.5% of the training set generated from the following month, even
though these training sets only contain 15.1% of equivalent domain name pairs.
This suggests a reasonable generalizational capacity, though further analysis will
be necessary to study the interpolation and extrapolation behaviour.
3 Experiments
In order to test the ANN, we integrated a sidebar into Netscape 6.2 to display
the recommender’s results in a user-friendly manner, and used a breadth-first
traversal in the hypergraph-neighborhood of the current HTML document to ob-
tain a search space, an algorithm which is time consuming. It is obvious that the
response times may vary greatly based on the speed of the Internet connection
and the performance of the user’s computer.
First test runs performed by ten users with varying IT skills and educational
backgrounds with the prototype of the TrailBlazer framework indicated that the
algorithm’s time consumption is acceptable, if it can be used in a “single shot”
manner, being launched by the user to find related domains to the current HTML
document. Another important issue is the presentation of the algorithm’s results.
Since the algorithm parses a great number of documents, simply displaying the
domain name or title of the HTML document residing at that address may lack
transparency, a problem which can be avoided by presenting the results in a
manner similar to search engines (by displaying link target title, meta-data or
excerpts).
The advantage of the algorithm is that it can exploit the generalizational
capability of the ANN, and thus operate on unknown and yet uncharted regions
of the net that are not part of the knowledge base, which clearly distinguishes
it from other approaches to the “What’s Related” - problem (see e.g. [2]).
4 Perspectives
Further features of the system will incorporate other methods of search-space

generation. These algorithms will be based on user trails, bookmarks, brows-
ing history, link databases, web connectivity charts, search engine results and
other sources of URLs, and thus Internet-domains, and will provide a significant
increase in system performance. Due to the fact, that we were “short of trail
data” at this early stage of the Trailist project, and had not yet connected the
system to existing online databases, we had to generate the hyper-graph neigh-
borhood of an HTML-document seed by hypertext parsing, even though it is an
extremely time consuming way of search-space generation, and loses the content
independence of the trail-based approach.
Recommending Internet-Domains Using Trails and Neural Networks 371
Another possible use of the ANN, besides recommending related domains,

could be part of a “distance measurement system” for trailmarks, which could
be useful for trail selection in a certain radius of the current document, thus
increasing the probability of encountering other trails in the process of hyper-
media navigation. The ANN would be performing the distance measurement on
the domain name level of the trailmarks’ nodes.
Summarizing, we have shown how ANNs trained with patterns extracted
from user trails can be used to recommend Internet-domains, described the first
experiments and outlined future extensions of the system. Even though the be-
haviour of the ANN has to be studied more thoroughly, with a special focus
on the interpolation and extrapolation behaviour on partly or totally unknown
Internet-domain pairs, first results indicate that the system is able to operate
on yet unknown domains.
Acknowledgements
This work has been supported in part by the Austrian Fonds zur Förderung der
wissenschaftlichen Forschung (FWF) under grant No. P14006-INF.
References
1. John S. Breese, David Heckerman, and Carl Kadie. Empirical analysis of predictive
algorithms for collaborative filtering. In Proceedings of the Fourteenth Conference
on Uncertainty in Artificial Intelligence, pages 43–52, July 1998.
2. Jeffrey Dean and Monika R. Heinzinger. Finding related pages in the world wide
web. In Proceedings of the Eigth World-Wide Web Conference, pages 1467–1479,
1999.
3. Siegfried Reich, Leslie A. Carr, David C. DeRoure, and Wendy Hall. Where have
you been from here? Trails in hypertext systems. ACM Computing Surveys — Sym-
posium on Hypertext (published as electronic supplement), 31(4), December 1999.
4. Siegfried Reich and Erich Gams. Trailist - focusing on document activity for as-
sisting navigation. In Proceedings of the Twelfth ACM Conference of Hypertext and
Hypermedia, pages 29–30, August 2001.
5. Paul Resnick and Hal R. Varian. Recommender systems. Communications of the
ACM, 40(3):56–58, March 1997.
6. D. Rumelhart, G. Hinton, and J. McClelland. Learning internal representations,
1986.
Learning Grammar with Adaptive Hypertexts:
Reading or Searching?
Angela Brunstein, Jacqueline Waniek, Anja Naumann, and Josef F. Krems1
Technische Universität Chemnitz, Institut für Psychologie

Allgemeine und Arbeitspsychologie, D-09107 Chemnitz, Germany
{Angela.Brunstein,Jacqueline.Waniek,Anja.Naumann,Josef.Krems}@ph
il.tu-chemnitz.de
Abstract. Using adaptive hypertext as a learning tool, the present study

addresses the question of the effects of different processing goals, especially the
goal of general reading a hypertext as opposed to the goal of searching for
specific information, on learning content and skills at English grammar. Twenty
students with German as mother tongue processed the present continuous
chapter of the Chemnitz InternetGrammar. It has been shown that readers
answer a higher amount of questions about details and, more importantly, they
answer them in more detail than searchers. Nevertheless, searchers tend to
produce more elaborated answers to complex questions than readers. Both
groups performed better on skill tests after the session, showing no effect of the
performed task. Based on this experimental evidence, the requirements on
adaptive learning tools are discussed.
1 Introduction
Adaptive hypertext has some interesting properties that make it suitable as a learning
tool [1], e.g. it’s possibility to adapt the lecture online to the learner. In the domain of
foreign language education web based systems exist, that provide features beyond
conventional learning material, striving to impart declarative and procedural aspects.
For example,the Chemnitz InternetGrammar (http://www.tu-chemnitz.de/phil/Internet
Grammar) combines grammatical rules, actual examples from politics and grammar
exercises adapted to the proficiency of the learner.
It is often reported that the processing goal is an essential factor for navigating and
learning with hypertext, e.g. in [2]. Also, the influence of the learner’s goal
sometimes dominates the influence of hypertext design [3,4, but 5]. Or it is at least
postulated that different forms of design may be appropriate for different goals [6]. So
the study of Dee-Lucas and Larkin [7] has shown that a high degree of segmentation
of the hypertext led to a deeper understanding of the content while a low degree of
segmentation led to a broader knowledge with subjects searching for details and
solving problems.
1 This study was supported by German Research Foundation Grant KR 1057.

Learning Grammar with Adaptive Hypertexts: Reading or Searching? 373
It is assumed here that the guidance of different instructions associated with more
or less specific goals is responsible for reported effects of both goals and interaction
between goal and design by constraining the mental representation of the content.
Two plausible goals for using hypertext systems are reading chapters of a learning
system and searching for details within them. Reading a hypertext without specific
instruction requires to decide which information is essential. However, there are only
few navigation decisions. Searching for details within the hypertext requires to decide
where to go next to find the desired information. However, searchers do not have to
separate central and secondary information already given by the task [7].
Therefore readers should develop a broader mental representation of the content
and searchers should demonstrate a deeper understanding of the content [4].
Therefore readers should answer more factual questions in more detail. In contrast,
searchers should answer questions in less detail altogether. However, they should
answer complex questions more completely. This pattern should be shown only for
declarative knowledge since the specific searching instruction focuses on rules and
not on skill acquisition. So the benefit of the hypertext should be the same for
searchers and readers in respect to application of the learned facts.
2 Methods
20 students of the Chemnitz University of Technology took part in the study. All
students were native speakers of German. On average they solved 10 out of 21 items
of a performance test on English grammar before processing the text.
One chapter dealing with the present continuous of the Chemnitz InternetGrammar
was adapted to the experimental setting. Learning the present continuous is a
challenging task for German students, since the German language lacks a verbal
representation of the feature (+continuous).
The hypertext was created using Hypercard 2.4.1. The stack contained 75 cards for
rules, examples and exercises. The headings of the topics were presented as a
navigation tool on the left side of each card. Tabs above link the card to the
corresponding rules’, examples’ or exercises’ card. Buttons below lead respectively to
the previous and next card.
The study was conducted on iMacs with 64 MB RAM. Factual knowledge was
measured by a questionnaire consisting of 24 detailed questions, skill level was
measured by performance tests each consisting of 14 items. The questionnaires were
presented as paper and pencil tests.
All subjects processed the chapter for 30 minutes and navigated freely within the
hypertext. The time seems to be enough to read all cards once, but it doesn’t prove to
be sufficient to get all details.
The study was conducted in group-sessions of up to six subjects at a time. The
searching group of 10 subjects was instructed to answer the questions corresponding
to the text online. The reading group was instructed to process the chapter to learn
about the present continuous and afterwards to answer the questions by recall. Both
groups performed a skill test before and after processing the chapter. Altogether a
session lasted about one hour.
374 A. Brunstein et al.
3 Results
Factual Knowledge. As expected, readers answered a higher amount of detailed

questions than searchers. The former on average answered 21 out of 24 questions,
while the latter had an average of 14 answers. This difference is statistically
significant: F(1,32) = 14.65, p < 0.001. Interestingly, missing data occurred more
often within the second half of the questionnaire for the searchers (10 at the
beginning, 4 at the end) but not for the readers (each 11) (F(1,32) = 9.95, p < 0.001).
Also as predicted the readers answered detailed questions more accurately than the
searchers according to the number of answers (F(1,16) =10.59, p < 0.005). The
readers attained an average of 19 marks out of 48, whereas the searcher reached only
13 marks. Looking only at the answers where the subjects report all necessary details,
the effect disappears: In this case the readers gave a complete answer in 7 out of 24
cases while the searchers did so in 9 cases.
Skills. Both groups performed better after processing the chapter (66% of the items
answered correctly) than before (56% of the items answered correctly) (F(1,16) =
4.16, p < 0.05). The processing goal had no influence whatever on this gain (F(1,16) =
0.44). However, subjects performed better only in respect to hits (correctly inserted
verb forms), but not in respect to correct rejected verb forms. Readers and searchers
on average gained 29% and 13% respectively (F(1,16) = 7.62, p < 0.01) in respect to
hits. The difference between readers and searchers was not remarkable (F(1,16) =
1.13, p = 0.31). In respect to correct rejections subjects performed at a level of 77%
correct answers before and 76% after processing the continuous chapter.
100
90
80
70
60
50
40
30
20
10
0
Number Accuracy Completeness Hits Correct Rejections
Factual Knowledge Skills
Fig. 1. Performance of readers (dark columns) and searchers (light columns) on factual
knowledge and skills test in percent. Factual knowledge is illustrated by the number of answers,
accuracy and completeness of answers. Skill level is illustrated by hits before and after
processing and correct rejections before and after processing of the chapter.
Learning Grammar with Adaptive Hypertexts: Reading or Searching? 375
4 Discussion
Discussing intelligent and adaptive educational systems is usually aimed at adapting

the system to the learner. Therefore it is essential to characterize constraints on the
learner’s behavior as precisely as possible, e.g. the processing goal.
This study has shown that knowledge acquisition is affected by goals even with
exactly the same hypertext design: Readers answered more factual questions in more
detail while searchers tended to give more complete answers. There was no
difference between both groups with respect to acquired skills. The subjects improved
their hit rate but not their rejection rate.
For adaptivity in eLearning the results show the following: First, it can be useful
not only to manipulate the appearance of the system but also to guide learners through
the material by instruction relevant to their goals.
Second, the learning system should adapt itself to the type of relevant knowledge
and to the complexity of desired information. That concerns adaptive sequencing,
presentation and navigation.
Third, the acquired knowledge doesn’t need to develop consistently as seen with
performance tests. This means that there is a need for segmenting and differentiation
of the given feedback for problem solving and solution support.
And last, adaptive systems should manage omissions differently depending on if
they indicate lack of knowledge or insufficient time management. So they should
offer more information for hard tasks and less information for time problems.
Altogether, this study shows differentiated effects of the learner’s goal on
knowledge acquisition and the need for intelligent guidance by an adaptive learning
system.
References
1. Brusilovsky, P.: Adaptive and Intelligent Technologies for Web-based Education. Künstliche
Intelligenz, Vol. 4/99. arenDTap Verlag, Bremen (1999) 19-25
2. Chen, C., Rada, R.: Interacting with Hypertext: A Meta-Analysis of Experimental Studies.
Human-Computer-Interaction 11 (1996) 125-156
3. Foltz, P.W.: Comprehension, Coherence, and Strategies in Hypertext and Linear Text. In:
Rouet, J.F., Levonen, J.J., Dillon, A.P., Spiro, R.J. (eds.): Hypertext and Cognition.
Erlbaum, Hillsdale, NJ (1996) 109-136
4. Wiley, J.: Supporting Understanding through Task and Browser Design. Proceedings of the
Twenty-third annual Conference of the Cognitive Science Society. Erlbaum, Hillsdale, NJ
(2001) 1136-1141
5. Shapiro, A.M.: The Effect of Interactive Overviews on the Development of Conceptual
Structure in Novices Learning from Hypermedia. Journal of Educational Multimedia and
Hypermedia 9 (2000) 57-78
6. Dee-Lucas, D.: Instructional Hypertext: Study strategies for different types of learning tasks.
Proceedings of the ED-MEDIA 96. AACE, Charlottesville, VA (1996)
7. Dee-Lucas, D., Larkin, J.H.: Hypertext Segmentation and Goal Compatibility: Effects on
Study Strategies and Learning. Journal of Educational Multimedia and Hypermedia 9 (1999)
279-313
SIGUE: Making Web Courses Adaptive
Cristina Carmona, David Bueno, Eduardo Guzman, and Ricardo Conejo
Departamento de Lenguajes y Ciencias de la Computación

Universidad de Málaga
{cristina,bueno,guzman,conejo}@lcc.uma.es
Abstract. Most of the information of the WWW is not adaptive, rather it is

dispersed and disorganized. On the other hand, the creation of an adaptive
course is a complex task. SIGUE is an author tool that makes possible to build
adaptive courses using existing web pages. This means that if there is a lot of
information on the web about the same topic the author doesn’t have to design
the content of a specific course, he can reuse these pages to build his own
course, taking the best pages for the concepts he wants to explain. The author
can also construct adaptive courses reusing previously non-adaptive ones.
SIGUE provides an enhanced interface for the student, controls his interaction,
and annotates the visited links in a student model.
1 Introduction
Nowadays, there are many tutorials on the WWW that let students acquire knowledge
on a specific subject. Most of them are not adaptive. Because of the freedom and the
disorganization of the Web, a user can navigate freely in the pages of a course, and
even in pages not related with the course. These problems have been stated several
times and have been clearly summarized in [1], which points out four main problems:
Disorientation, Cognitive overload, Narrative flow, and Content readiness. One of the
characteristics that make a tutorial system adaptive is the possibility of modifying the
sequencing or presentation of the course for each user according to their knowledge or
preferences. Also the ability to recommend which unit they should do next. In this
article we present SIGUE. This system is an author tool on the Web that can convert
courses that weren’t adaptive into adaptive ones. The original course could be a set of
static pages build in as a course or existing pages from different sources. This lets
authors make courses with the best information on the Web about a subject. SIGUE
provides a student's interface that assists the students with adaptive tools when they
navigate through the web pages. Another aspect of adaptivity is the multilingual
interface of SIGUE, currently available in Spanish and English. SIGUE can also be
integrated as a component of MEDEA [2]. MEDEA is an open system to develop
Intelligent Tutorial Systems. It is composed of independent educative modules
coordinated by a core that controls the instruction of the student. MEDEA can decide
with is the best module to explain a concept in each moment. There are other
modules, like SIETTE [3] which create tests to assess the estimated knowledge of the
student. There are several tutorial systems on the Web that are adaptive and can guide
the student through the learning process (i.e. ELM_ART [4], or Metalinks [5], just to

SIGUE: Making Web Courses Adaptive 377
mention a few). The main difference between these systems and SIGUE is that
SIGUE has not been developed as a content authoring tool with adaptive features, but
as an authoring tool to reuse previously developed non-adaptive material, turning it
into adaptive material.
2 The SIGUE System
SIGUE has two main modules, the author tool (SIGUEAUTOR) and the student tool
(SIGUE), which is used by the student to study the course. Both of them are
accessible through a WWW interface
2.1 The Author Module (SIGUEAUTOR)
The author module lets human tutors to create their Web courses through SIGUE. The
domain model of the system and the way that users can create adaptive courses are
explained in the following subsections.
Domain Model. The author tool can create a course defining the domain model
associated to it. For a specific course the domain model is a hierarchy of concepts.
Various URLs can be associated with each concept and the relationship to other
concepts can also be defined. For SIGUE two relationships have been defined:
“belongs to” and “prerequisite of”. This gives an idea of the order in which concepts
should be visited, which is the sequence of the curriculum. The number of pages
(URL) to describe a concept is unlimited. It is necessary to indicate the kind of
information that each page contains (theory, examples or exercises) and the difficulty
(easy, normal or hard).
Creation of Courses. Once the domain model is created and all the concepts and
relations have been defined, the author has to associate pages to each concept. The
authoring tool lets the user navigate freely, but supervised. The pages visited by the
user are processed by a parser that modifies the links in accordance with SIGUE. The
parser modifies regular links and forms. When the author finally finds an appropriate
page for a concept he only has to click a button to add it to the course. The author can
also define a glossary of terms for the course with synonymous. When a term defined
in the glossary appears in any page of the course, a link is automatically included to
show the definition of this term. Courses in SIGUE can have different modes of
operation. The author decides how adaptive the course will be and must set the mode
of operation accordingly. There are four predefined modes: (a) Disable all the links in
the pages shown to the student. He will be able to do the course by accessing
documents only through the concept tree; (b) Leave all the links of the pages, this lets
users navigate freely even in pages not related to the course; (c) Enable only the links
that give references to pages that belongs to the course. (d) Full adaptive. The links
will be enabled according to the user model. Only those links corresponding to
concepts that the user is ready to learn will be activated. To construct an adaptive
course, it is necessary to define the prerequisite relationship of the concepts. It is
378 C. Carmona et al.
important to remark that the author of a course in SIGUEAUTOR doesn’t necessarily

need to be the author of the HTML pages that he associates to his courses. One of the
main ideas of this system is to reuse and compile pages that are related to a topic, but
are located in different sites on the WWW, and add adaptive features to create a
course with them.
2.2 The Student Module (SIGUE)
Once the course has been developed with SIGUEAUTOR, it will be available to
students in SIGUE. While connected to the web through SIGUE, the student will see
the hierarchical structure of concepts created by the author. Each time he selects a
node he will see the main page associated with this concept. Before showing any page
of the course, the HTML document is parsed and the links and submission forms are
redirected to the SIGUE engine that acts as a supervisor between the user and the
WWW. All links and forms are modified according to the strategy selected by the
author for the course. The system will also include new references for the terms of a
glossary defined by the author. If the course is fully adaptive, a student model is
created for each user. In this model SIGUE stores the student’s navigation path for
each concept of the course. For each concept it shows two indicators. (a) The
estimated background of the student visiting this concept, that is, if he’s prepared or
not to read it, according to previous pages that have been visited. (b) The percentage
of pages related to that concept that he has already visited. SIGUE makes estimations
of the knowledge based on the percentage of visited pages for each concept. So, if the
percentage of URLs visited for a concept is less than the minimum the status of the
concept is “empty”; if this percentage of visited pages is bigger than the maximum,
the status is “full”. The intermediary case is shown as “half-full”. The status of a
concept is associated to the percentage of pages the user has studied within a concept.
This information is shown to the user by a progress bar that appears next to each
concept. The level of preparation necessary for a user to visit a concept is reflected by
using colors in the nodes of the concept tree. They can be green, red or orange. The
color of the node is decided using the status of the prerequisites in this way: Green: A
concept has this status when all the prerequisites are shown to have their status as
“full”. The user is ready to learn it. Red: This will be the color if at least one of the
prerequisites of a concept is “empty”. It indicates that the user is not ready to study
the concept. Orange: This indication will appear when no empty prerequisites exist.
Nodes with this color could be studied but finishing all the prerequisites first is
recommended. The student model is updated after each interaction, every time the
user visits a page. This update is reflected in the concept tree, modifying the status
and color of nodes accordingly. The aim of SIGUE is to guide the student’s
navigation, and support it with adaptive annotation, but at the same time let him move
freely through the pages of the course. This is why access to any page is permitted for
the user even if this is not recommended (red). When a concept is red, it won’t change
its color until its prerequisites have been visited, even if the user has already visited
the documents associated to this concept. The student also has the freedom to visit
pages through the concept tree or through the activated links of each page. The
activated links correspond to the glossary and to the recommended concepts (green).
This means that if there is a link in one page to a page of a concept that is not
recommended (red, orange) the authors can decide whether this link should be
SIGUE: Making Web Courses Adaptive 379
activated or disabled. Another important characteristic used to guide students is the

inclusion of a simple planner that the student can use to get the best recommendation
for the next document to view from the concept he is studying. This mechanism is
designed for “two buttons users”. These are the NEXT and BACK buttons. An
evaluation module can replace the simple mechanism used by SIGUE, when SIGUE
is used as a component of the MEDEA architecture.In this case the navigation frame
provided by SIGUE is replaced by MEDEA’s navigation frame. The SIGUE
component receives a temporary student model each time it is activated by a call from
MEDEA, and this information is used to control the links on the main page.
3 Conclusions and Future Works

The web is a great source of information, but it is dispersed and hard to find. The
advantage of SIGUE is the possibility of taking the best parts of each tutorial;
compiling them in the order the author thinks is best; organizing them; adding
relationships and glossary terms; and providing a learning environment in which the
student is always oriented because someone (the course author) has followed the path
before him and has provided some marks. The author can also decide the degree of
freedom that the user should have while navigating, according to the characteristics of
his expected audience. The adaptive features of SIGUE can also be used to convert
static tutorials into adaptive tutorials. SIGUE is an independent and adaptive
component, with author and student modules that can be integrated into MEDEA. If a
better measure of the knowledge is needed it can be taken from MEDEA. At present
the prototype of SIGUE is fully functional and can be used freely. Future work on
SIGUE will be directed towards improving the estimation of user understanding by
analyzing their navigation behavior. SIGUE could also a workbench to compare the
results and the satisfaction of users studying a course in an adaptive way or not. It is
available for trial at http://www.lcc.uma.es/SIGUE
References
1. Murray, T. Shen, T. Piemonte, J. Condit, C. Tibedau, J.: Adaptivity for conceptual and
narrative flow in hyperbooks: The Metalink system. Adaptive Hypermedia and Adaptive
Web-based system. Lecture Notes on Computer Science 1892, Berlin (2000) 155-166
2. Trella, M., Conejo, R., Bueno, D.: MEDEA: Una arquitectura basada en componentes para
el desarrollo de Sistemas Tutores Inteligentes en Internet. CAEPIA 2001. (2001) 469-478
3. A.Rios, E.Millán, M.Trella, J.Perez-de-la-Cruz, R.Conejo, Internet Based Evaluation
System, en: Artificial Intelligence in Education, IOS Press, Amsterdam (1999) 387-394
4. Weber, G., Specht, M.: User Modeling and Adaptive Navigation Support in WWW-Based
Tutoring Systems. Proceedings of the Sixth International Conference on User Modeling,
UM97. Springer Wien New York (1997) 289-300
nd
5. Henze, N. Nejdl, W.: Adaptivity in the KSB Hyperbook System. Proceedings of the 2
Workshop on Adaptive Systems and User Modeling on the WWW. (1999)
An Ontology-Guided Approach to Content Adaptation
in LEO: A Learning Environment Organizer
John W. Coffey
The Institute for Human and Machine Cognition

The University of West Florida
Pensacola FL 32514
[email protected]
Abstract. This paper describes new capabilities of a nonlinear course

presentation program entitled LEO: a Learning Environment Organizer. A
Knowledge-based Instructional Assistant combines knowledge of the learner’s
attainment with knowledge of various attributes of the instructional media in
order to make recommendations of media that might benefit the learner. This
work describes an ontology to represent the basic conceptualization of the
learning system, and how that representation is used to match potentially useful
instructional content to student profiles. Recommendations of media are based
upon multi-parameter similarity measures between media content and capability
attainment.
1 Introduction
This work describes new enhancements to a program entitled LEO, a non-linear

Learning Environment Organizer [6]. LEO is an advance organizer [1] that presents
links to instructional media that contain content pertinent to the topic. This basic
system is augmented by a Knowledge-based Instructional Assistant that exploits
annotations of the content in a course and a student's emerging attainment profile in
order to suggest instructional materials that might be beneficial to the student.
The approach in this work is based upon the idea of an ontology [9] for a learning
environment that tracks student progress. The ontology creates a common parlance
that is utilized by the various parts of the system including the testing component and
the Instructional Assistant. The remainder of this paper presents a brief description of
the Knowledge-based Instructional Assistant, an ontology for an instructional system
that tracks student progress, and the means through which the ontology is used to map
attainment deficiencies to instructional resources.
2 Recent Related Work on Adaptive Learning Systems
A substantial amount of literature exists on adaptive learning systems. Brusilovsky [3]

presents a survey of adaptive educational hypermedia. He enumerates several current
systems and describes adaptation in terms of adaptive navigational support and
adaptive presentation. He further describes adaptation relative to a learner's

An Ontology-Guided Approach to Content Adaptation in LEO 381
knowledge, interests and individual traits. De Bra & Ruiter [7] describe a well-
realized, general adaptive hypermedia engine that can be used to adapt instructional
content for students. Their system, AHA!, determines user interest levels in topics and
also tracks what users have read to determine preparedness for subsequent items.
Castells & Macias [5] describe PEGASUS, an adaptive hypermedia modeling system
that is based upon a domain-specific ontology that is enumerated by the author. Their
system has a domain model that is rendered in XML that describes a semantic
network of domain objects. An explicit presentation model is updated from
presentation templates and the user model as the user traverses through system.
Fig. 1. The Knowledge-based Instructional Assistant.
3 A Knowledge-Based Instructional Assistant
The ideas in this work leverage the integration of LEO with the multimedia
knowledge modeling software suite "CMapTools" [4]. Knowledge models created
with CMapTools are comprised of many learning resources such as texts, graphics,
video and web links, etc. Figure 1 presents a graphic depiction of LEO with links to a
CMapTools knowledge model and with the Knowledge-based Instructional Assistant
indicating recommended media. The user starts with a pretest to establish baseline
knowledge for the course. The baseline is updated as the student works through the
course within LEO, submitting deliverables and taking tests that are evaluated.
On the basis of the attainment profile and catalogued attributes of the instructional
content, the Knowledge-based Instructional Assistant determines instructional content
to recommend. The student can access recommended content through the links
between topics and content in LEO itself, or from a comprehensive list of
recommended resources that displays by individual topic, or by all the topics in the
course.
4 An Ontology-Guided Approach to Content Adaptation
Cognitive and computer scientists borrow the notion of ontology, study of the nature
of being or existence, from the philosophy field. In computer/cognitive science
382 J.W. Coffey
context, an ontology is generally held to be a formal representation of a

conceptualization of a domain [8], a common parlance or vernacular for a knowledge
domain.
In the modeling of an instructional system, generic attributes such as students,
testing resources and learning resources form templates for domain-specific topic
areas and descriptors of competencies in these topic areas. The ontology is comprised
of five basic entities: Student, Competency, Attainment, LearningResource, and
TestResource.
This basic representation provides a framework for domain-specific areas of
competency and specific dimensions of competency within an area. Ratings for these
measures can be any range that the user of the system chooses. For example, Likkert
scale values on the range from 0 to 5 might be used.
5 Identifying Relevant Instructional Content
The process of matching possibly useful instructional materials to a user profile that is
employed here was suggested by weighted nearest-neighbor [11] similarity measures
utilized in case-based reasoning approaches. The fundamental idea is to employ a
matching procedure [2] to identify a set of potentially useful instructional materials,
and then to rank-order the set [10].
In an instructional system, the goal is to create matches on the basis of student
strengths and weaknesses in topic areas and to suggest resources that vary along the
various dimensions of competency in the area. As an example, a competency area in a
Data Structures class would be an understanding of linked lists. Various competency
dimensions within the area might include a theoretical understanding of the ideas
surrounding the data structure, the ability to program various operations on the data
structure, the ability to analyze various algorithms from code, etc.
The method described here creates matches between test questions and
instructional resources. This approach is utilized to make recommendations from a
simple, minimal student profile that records student performance on the questions. For
questions on which the student has performed poorly, the similarity measure can be
applied to find the learning resources that are most appropriate. Initially, the search
space of relevant learning resources is reduced by a string match on the descriptors of
the test question of interest (TRj) and the set of resources (LR) to identify LR' ⊂ LR.
A similarity measure of the following sort is then applied to assess the similarity of a
given resource from LR' to the test question under consideration:
n
Difference(TRj,LR'k) = Σ WCDi * (TRjDi – LR'kDi)2 (1)
i = 1
In formula 1, TRj is the jth TestResource, LR'k is the kth Learning Resource from
the string-matched subset of LR, WCDi is the weight assigned to the ith competency
dimension, TRjDi is the value of the jth test question with regard to competency
dimension i and LR'kDi is the value of the kth learning resource with respect to
competency dimension i. This analysis is a summation of squared, weighted distances
between the various dimensions of competence supported by a learning resource and
the same dimensions in a test question.
An Ontology-Guided Approach to Content Adaptation in LEO 383
6 Summary and Future Work
This work describes a new approach to online course structuring and instructional
content adaptation that is embodied in LEO, a Learning Environment Organizer. LEO
is enhanced with an instructional assistant that suggests content that might be
appropriate to a particular student. The student may access recommended materials by
noting recommendations in navigational links to content materials from individual
topics in the organizer or by viewing a list of suggestions pertaining to a given topic.
The adaptation approach starts with a basic ontology for common elements in an
adaptive instructional domain. The approach identifies orthogonal ontological
elements of content areas and dimensions of attainment within an area. The common
ontology provides a generic framework within which context-specific vernacular can
be mapped.
The approach described here employs a multi-parameter similarity measure to
match test questions to content. This is a minimal approach that is the most subject to
automation, but the least flexible. Future work will address ways to allow for a more
comprehensive student profile that will allow the assessment of student attainment
both across the various dimensions of an individual content area as well as across
content areas on an individual attainment dimension. Also, the approach will be
applied to additional knowledge domains in order to draw conclusions regarding the
sorts of competency areas and dimensions that are identified.
References
1. Ausubel, D.P.: Educational Psychology: A Cognitive View. Rinehart and Winston, NY

(1968)
2. Borg, I., Lingoes, J.: Construction of SSA Representations. In Multidimensional Similarity
Structure Analysis. Springer-Verlag, New York: (1987)
3. Brusilovsky, P.: Adaptive Hypermedia. User Modeling and User-Adapted Interaction. Vol
11. 87-110, Kluwer Academic Publishers, The Netherlands (2001)
4. Cañas, A. J., Coffey, J. W., Reichherzer, T., Hill, G., Suri, N., Carff, R., Mitrovich, T., &
Eberle, D.: El-Tech: A performance support system with embedded training for electronics
technicians. Proceedings of FLAIRS’98, Sanibel Island, FL, (1998)
5. Castells, P., Marcias, J. J.: An Adaptive Hypermedia Presentation Modeling System for
Custom knowledge Representastions. Proceedings of WebNet 2001, (2001)
6. Coffey, J.W.: LEO: A Learning Environment Organizer to accompany Constructivist
Knowledge Models. Unpublished doctoral dissertation, The University of West Florida,
Pensacola, FL. (2000)
7. De Bra, P., Ruiter, J.: AHA! Adaptive Hypermedia for All. Proceedings of WebNet 2001,
World Conference on the WWW and Internet. Orlando, FL, October 23-27, (2001)
8. Genesereth, M.R., & Nilsson, N.J.: Logical Foundations of Artificial Intelligence. Morgan
Kaufmann, San Mateo, CA (1987)
9. Gruber, T.A.: A Translation approach to Portable Ontology Specifications. Knowledge
Acquisition, Vol 5 (1993) 199-220
10. Kolodner, J., Leake, D. A.: Tutorial Introduction to Case-based Reasoning . in Case-based
Reasoning, Experiences, Lessons and Future Directions. D. Leake (ed.). AAAI Press,
(1996)
11. Spalazzi, L.: A survey on case-based planning, AI Review, (16) 1 (2001) 3-36
A Scrutable Adaptive Hypertext
Marek Czarkowski and Judy Kay
Basser Department of Computer Science,

University of Sydney, Australia
{marek,judy}@cs.usyd.edu.au
Abstract. This paper describes Tutor/ADAPT, a working model and frame-

work for scrutable adaptive hypertext. This enables the user to scrutinise the
system to determine how the document was adapted. They are also able to con-
trol the adaptation by altering the user model at any time in the interaction. The
user can see which parts of the user model have been used to control aspects,
which appear in the document. The system has three major parts: an authoring
tool; the web based interface; and ATML is the language used for adaptive hy-
pertext source documents. This paper describes the user view, especially the
scrutability support.
1 Introduction
This paper describes the Tutor/ADAPT framework for constructing and displaying
scrutable hypertext. By scrutable we mean that the reader can explore the interface to
determine what has been adapted. There has been considerable work in making user
models accessible [1,2,3,4,5,6]. This work takes a step beyond that, making the proc-
esses also scrutable.
The hypertext author creates adaptive teaching material using an editor, which cre-
ates pages in the language ATML, a mark-up to HTML 4.0, which conforms to XML
1.0 [7]. Later the student can use the web-based interface to make use of the teaching
materials. This is like a conventional adaptive hypertext system. The novel aspect of
Tutor is that the user can scrutinise each part to see what has been customised to
them.
2 System Overview
On starting a new course, users are asked questions. This very simple mechanism
establishes their initial user-model. Next, students see an interface with teaching ma-
terial like that shown in Figure 1. The icons across the top offer various facilities,
including a course map, teacher's news page, a notes editor, glossary, the personal
profile discussion room, log out and, at the far right, help. Figure 1 shows the

A Scrutable Adaptive Hypertext 385
hypertext as it would appear if a student had answered none to the question How
many examples would you like to see for each new concept.
The system requires material marked up in ATML (Adaptive Teaching Mark-up
Language). ATML provides a standard framework for authoring adaptive teaching
hypertext material, allowing the author to include adaptive content and adaptive navi-
gation features in the material. The author creates a course in an editor, which aids,
and in some cases automates, the creation of ATML documents.
The author uses simple ATML syntax to create highly dynamic HTML content.
When the user requests an ATML page, the system processes the user’s profile and
translates the ATML content into HTML 4.0, including the required JavaScript1.2
and CSS1.0 code which implement the dynamic HTML features. The resulting
HTML is displayed in the browser.
Fig. 1. Sample lesson as seen by user who does not wish to be shown examples.
3 Scrutability
At the very bottom of each page of adapted material is a link, How was this page
adapted to you?. As in Fig. 1 this is below the navigation arrows and a horizontal
line. If there has been no adaptation, this is indicated with the message There was no
adaptation of this page. Where there has been adaptation, the bottom of the screen
expands to list the set of page adaptations with a separate entry in the adaptation ex-
386 M. Czarkowski and J. Kay
planation. Each entry indicates if content (lesson text) or navigation (navigation ar-
rows) has been included or excluded. The text is also colour coded with descriptions
for content displayed in green and those for excluded content in red.
The user is able to query the system to scrutinise exactly what was included or ex-
cluded. If the user clicks on the text [highlight] the interface indicates the part of the
page content adaptively included by highlighting the background of the relevant ma-
terial in green. If the user wanted to see what they missed out on in their adapted
hypertext, they select the Show all text hyperlink. The system reloads the page ignor-
ing the adaptation criteria and the user profile to display all the content available on
that page (see Fig. 2). The user can select [hide explanation] at the bottom of the
screen. It then reverts to revert to the usual form as in Figure 1.
Fig. 2. This depicts the scenario where a user has activated the Show all view on the page from
Fig. 1 and then clicked the provided hyperlinks to highlight the first, second and fourth adap-
tive sections. The first and fourth adaptive sections were excluded from the page hence are
highlighted with a red background (dark grey). The second section was included and hence is
highlighted with a green background (lighter grey).
A Scrutable Adaptive Hypertext 387
4 Discussion and Conclusions
The current work extends the basic of notion of adaptive hypertext with support for
scrutability so that the reader can determine how the document was customised for
them. The current system comprises of an authoring tool to support the author of
meta-hypertexts and a web interface to enable the reader to interact with their custom-
ised and scrutable hypertext. Basic explanations of the adaptivity are presented at the
bottom of each page so that the user can identify which triggers caused the presenta-
tion of each adaptive element on the page. These also enable the user to see where
optional material was omitted and the user model values causing this. This enables
the user to scrutinise the detailed adaptation. In addition, there is support for seeing
the global adaptation: the user can easily alter their profile and see the consequent
hypertext pages.
This approach can be used for what-if experiments, enabling the user to see what
the system would do if they answered user modelling questions differently. When a
user does this, they can scrutinise the effect of the large grain-size.
It offers a coherent and elegant structure for the author of the adaptive, individual-
ised hypertext material. The current trial of out system demonstrates interesting, use-
ful, though modest adaptation and represents a foundation for continuing work in
supporting and delivering scrutable, adaptive hypertext.
References
1. Corbett, A.T., Anderson, J.: Knowledge tracing: modeling the acquisition of procedural
knowledge. User Modeling and User-Adapted Interaction. 4 (1995) 253—278
2. Crawford, K., Kay, J.: Metacognitive processes and learning with intelligent educational
systems. In: Slezak, P., Caelli, T., Clark, R., Perspectives on Cognitive Science, Ablex
(1993) 63-77
3. Fischer, G., Ackerman, D.: The importance of models in making complex systems com-
prehensible. In: Tauber, M.: Mental models and Human-computer Interaction 2. Elsevier
(1991) 22—33
4. Höök, K., Karlgren, J., Waern, A., Dahlbeck, N., Jansson, C.G., Lemaire, B.: A glass box
approach to adaptive hypermedia. User Modeling and User-Adapted Interaction 6:2-3
Kluwer (1996) 157—184
5. Paiva, A., Self, J., Hartley, R.: Externalising learner models. Proceedings of World Con-
ference on Artificial Intelligence in Education. AACE, Washington (1995) 509—516
6. Self, J.: Bypassing the Intractable Problem of Student Modelling: Invited paper. Proceed-
ings of the 1st International Conference on Intelligent Tutoring Systems. Montreal (1988)
18—24
7. Czarkowski, M.: An adaptive hypertext teaching system. Honours Thesis, Basser Dept of
Computer Science, University of Sydney (1998)
AHA! Meets AHAM
Paul De Bra, Ad Aerts, David Smits, and Natalia Stash

PO Box 513, NL 5600 MB Eindhoven
The Netherlands
[email protected]
Abstract. After many years of hypertext research, the Dexter model was de-
fined [7] to capture the features of most existing hypertext systems in a single,
formal reference model. Likewise, the AHAM model [5] (based on Dexter) de-
scribes most features that are found in adaptive hypermedia systems (ahs). In
the AHA! project funded by the NLnet Foundation we are extending the simple
AHA system [4, 6] with the condition-action rules that were introduced in
AHAM [8]. This results in a more versatile adaptation engine, with more intui-
tive and more powerful rules.
1 Introduction
In 1994 we started a course on Hypertext, consisting of lectures and a Web-based

course text. From 1996 on the lectures were discontinued and the course text was
augmented with adaptive content and linking. Our adaptive software later became the
AHA! system [4,6], for Adaptive Hypermedia Architecture. Also in 1996, Peter Bru-
silovsky published a survey article [2] on adaptive hypermedia systems (ahs). This
article was updated in 2001 [3] to include more recent developments in this research
field. The overview articles show (informally) which kinds of features are offered by
the different systems. In 1999 we developed a reference model for adaptive hyperme-
dia, called AHAM (Adaptive Hypermedia Application Model) [5]. Based on the for-
mal Dexter model for hypermedia [7] AHAM provides a framework to express the
functionality of any ahs. Hence, AHAM is not the model behind the AHA! system (as
some people have asked us in the past) but rather the model behind almost every ahs.
In [8] we studied the adaptation rules and the behavior of adaptive engines that can be
expressed in AHAM. In this paper we describe our current efforts to bring the simple
AHA! system closer to the powerful adaptation systems that AHAM suggests. This
effort is part of a project, funded by the NLnet Foundation, to turn AHA! into a gen-
eral-purpose (Web-based) ahs.
In Sect. 2 we briefly describe the overall architecture of the AHA! system. Section
3 discusses the new rule system and adaptive engine. Section 4 provides an outlook to
the future of AHA!.

AHA! Meets AHAM 389
2 The AHA! System

According to Brusilovsky [2] there are two types of adaptation in ahs: adaptive pres-
entation and adaptive navigation support. Each time the user visits a page a user
model is updated. This user model is used to determine how the presentation of the
“next” page should be adapted and which advice should be given to the user regard-
ing the links to follow next. While some systems take sequences of pages into ac-
count, and possibly even the behavior of user groups, AHA! is one of the simpler
systems that work on a page (access) by page basis. The main functions of AHA! are:
• For every user there is a user model that represents how the user relates to the
Web-pages and to (higher-level) concepts. Each time a user visits a page a set of
generate rules determines how the user model is updated. We describe these rules
below.
• Each page contains (zero or more) fragments that are conditionally included in the
presentation. The requirement rules for inclusion are also described below. The
conditional inclusion of fragments is AHA!’s main adaptive presentation tech-
nique.
• For each page there is a similar requirement rule. It is used for adaptive navigation
support. AHA! uses different colors for link anchors to indicate the “desirability”
of the links. Depending on the choice of colors AHA! can be configured to use the
link hiding technique (with black undesirable links) or the link annotation tech-
nique (with all links shown in visible, different colors).
The AHAM reference model proposes a rich user model, with for each page or higher
level concept a set of attribute/value pairs, and with an adaptation model based on
condition-action rules. The (old) AHA! system on the other hand, as described in [4,
6], has a very simple user model with for each page or concept a single numerical
attribute (called knowledge), and with simple rules, mostly intended to describe the
propagation of knowledge from pages to sections to chapters. In this paper we de-
scribe how the “new” AHA! system adopts user model and adaptation model ele-
ments introduced in the AHAM model. The result is a new AHA! that tries to pre-
serve its initial simplicity while making the adaptation more versatile.
The “new” AHA! system uses multiple (user model) attributes per page or concept.
The author of an application can choose attribute names that match what attributes are
used for. It thus becomes possible to indicate e.g. interest in a certain concept as well
as knowledge about a concept. Also, the attributes can have Boolean, numeric or
string values.
For every page and fragment there is a requirement to decide whether links to the
page should be shown as “desirable” or whether the fragment should be included in
the presentation. The requirement is a Boolean expression in (user model) attribute
values for concepts (or pages). An example, using the non-existent website on Bel-
gian products we used in [6]:
beer.interest > 70 and chocolate.interest < 30
While browsing the site, visits to pages about beer contribute to the system’s confi-
dence that the user is interested in beer and not in chocolate. So a user who is inter-
ested in beer can be shown additional information that is hidden for chocolate lovers.
390 P. De Bra et al.
For every page or concept there is a set of generate rules. Each rule has a condition
that is checked to see whether the associated action should be performed. (It is also
possible to specify an alternate action to be performed when the condition is not
satisfied.) Each rule is associated with an attribute of the page or concept, and the rule
is “triggered” by an update to that attribute. For a page there are also rules that are
triggered by an access to the page. (This is treated as an update to an access attrib-
ute).
The following rule expresses that when the user reads a page the “knowledge”
about a concept increases by an amount that depends on a condition. Assume that the
interest in beer should be high to understand everything on the page about “Duvel”:
C: beer.interest > 70
A: duvel.knowledge := 100
AA: duvel.knowledge := 35
As another example, the following rule, to be associated with the interest attrib-
ute of beer, expresses that if the user shows a low interest in chocolate, then any
increase of the user’s interest in beer will induce a small (namely 20% of that in-
crease) decrease in interest for chocolate. If the user has already shown great interest
in chocolate then reading about beer no longer affects that.
C: chocolate.interest < 50
A: chocolate.interest -= 20
There is no limit to the number of (condition-action) rules that can be associated with
an attribute of a concept.
AHA! uses XML to represent the generate rules, but that syntax is too verbose to
write down in this short paper. Furthermore, the actual syntax is irrelevant as we are
also developing authoring tools that will hide the syntax from the author, and
(mySQL) database support that will use XML only for import and export of rule sets.
3 The AHA! Adaptation Engine

When a user accesses a page the generate rules for (the access attribute of) that
page are executed (as if access was changed from 0 to 100). AHA! distinguishes
between two kinds of updates that can be defined in actions:
• absolute: a value, specified in the action, is assigned to an attribute of a page or
concept. (symbol := used above)
• relative: the specified value is used as a percentage of the triggering update to be
added to or subtracted from the attribute; relative updates only make sense for nu-
merical attributes of course. (symbol += or -= used above)
The adaptation engine keeps track of which attributes of which pages or concepts are
updated by rule actions, and “triggers” the execution of their associated rules. This
process continues until there are no more rules to execute.
From research in active databases [1] we know that the rule execution process is
not guaranteed to terminate and may also not be confluent (meaning it may generate
AHA! Meets AHAM 391
different results depending on the order in which triggered rules are executed). In [8]
we showed how such problems can be detected at authoring time.
4 Conclusions / Future Work

The “new” AHA! user model and adaptation engine greatly improve the versatility of
AHA!. The extensions are based on the rule system presented in [8] for the AHAM
reference model. In the future we want to bring AHA! even closer to AHAM through
at least the following new features:
• links to concepts (not just pages), and a method to “select” the best page to present
when following such a link;
• a way to express generic rules, so that rules don’t need to be replicated for every
page or concept they apply to.
Also, authoring tools for the concept structure and the adaptation rules will be ex-
tended with static analysis methods to warn authors of potential termination and con-
fluence problems.
Acknowledgement
The development of the AHA! system is supported by a NLnet Foundation grant,

through “Adaptive Hypermedia for All!” project (conveniently abbreviated to AHA!).
More information about AHA! can be obtained at the Website http://aha.win.tue.nl/.
References
1. E. Baralis, and J. Widom. An algebraic approach to static analysis of active database rules.
ACM Transactions on Database Systems, Vol. 25, nr. 3, pp. 269−332, 2000.
2. Brusilovsky, P. Methods and Techniques of Adaptive Hypermedia. User Modeling and
User-Adapted Interaction, Vol. 4, pp. 1−19, Kluwer academic publishers, 1996.
3. Brusilovsky, P. Adaptive Hypermedia, User Modeling and User-Adapted Interaction,
Vol. 11, nr. 1−2, pp. 87−110, Kluwer academic publishers, 2001.
4. De Bra, P., A. Aerts, G.J. Houben, and H. Wu. Making General-Purpose Adaptive Hyper-
media Work. Proceedings of the AACE WebNet Conference, pp. 117−123, San Antonio,
Texas, 2000.
5. De Bra, P., G.J. Houben, and H. Wu. AHAM: A Dexter-based Reference Model for Adap-
tive Hypermedia. Proceedings of the ACM Conference on Hypertext and Hypermedia,
pp. 147−156, Darmstadt, Germany, 1999.
6. De Bra, P., and J.P. Ruiter. AHA! Adaptive Hypermedia for All. Proceedings of the AACE
WebNet Conference, pp. 262−268, Orlando, Florida, 2001.
7. Halasz, F., and M. Schwartz. The Dexter Hypertext Reference Model. Communications of
the ACM, Vol. 37, nr. 2, pp. 30−39, 1994.
8. Wu, H., E. De Kort, and P. De Bra. Design Issues for General-Purpose Adaptive Hyperme-
dia Systems. Proceedings of the ACM Conference on Hypertext and Hypermedia,
pp. 141−150, Århus, Denmark, 2001.
Adaptive Real Time Comment Generation
for Sail Racing Scenarios
1 1 2 1
Andrea Esuli , Antonio Cisternino , Giuliano Pacini , and Maria Simi
1
Dipartimento di Informatica, Università di Pisa, Corso Italia 40, 56125 Pisa, Italy
{esuli, cisterni, simi}@di.unipi.it
2
Accademia Navale di Livorno, Livorno, Italy
[email protected]
Abstract. We describe a comment generation subsystem developed as part of

an expert authoring system for sail racing scenarios. The result of the expert
system analysis is sent to a multimedia presentation subsystem composed of a
3D player and a speaker agent. Comments output by the text-to-speech
component must be carefully generated so that they are concise, relevant, non
repetitive and timely since they must be synchronized with the 3D animation.
The comment generation process is adaptive with respect to the time available
and to the user profile and preferences.
1 Introduction
As part of our work in developing intelligent tutoring systems for sail racing, we
designed and implemented a system for the multimedia presentation of scenarios
involving dynamic objects in general [1]. The architecture of the system, (see Fig. 1)
was inspired by the simulated soccer system within the RoboCup initiative [2].
Each module is implemented as a program that interacts with other modules using a
TCP connection and a text protocol.
Presentation System
Simulator Expert NLG 3D Player
/Editor System System
Fig. 1. The architecture of the expert authoring system
The authoring of sail scenarios is performed by means of a visual editor, working

in close cooperation with a simulator of boats movements. The editor offers a number
of high-level functionalities specific of the sail racing domain (such as combining
trajectories); the simulator ensures that the dynamic scenarios comply with physical
constraints. The simulation history, output of the Simulation/Editor module, consists
of a stream of instantaneous snapshots describing the boats positions (bi-dimensional)
and status (direction, rudder and sails).

Adaptive Real Time Comment Generation for Sail Racing Scenarios 393
The role of the Expert System module is to analyse and annotate with comments
histories received by the editor, using knowledge about the racing rules. In some
cases, in order to verify some conditions, the expert system needs to ask the simulator
how a given hypothetical situation will possibly evolve. The expert system analysis
results in a structured representation of all the facts worth commenting about, marked
with their importance. The expert however is not in the position to do a detailed
planning of the comment since it does not know about the rendering media, time
constraints or user preferences.
The presentation system (see Fig. 2) consists in a 3D player that is responsible for
showing the simulation and uttering the comments generated by the comment
generation module. A Web browser with a VRML plug-in is used to render the virtual
scene, under the control of a Java applet through the External Authoring Interface.
The main control functions are written in Jscript. The Microsoft Agent technology is
used to implement the animated character [3].
Fig. 2. Player interface
The comment generation sub-system needs to perform a complex natural language

generation task in close cooperation with the visual rendering and the user. It is
responsible for generating the most informative messages that can be pronounced in
the available time and to pass control to the Microsoft Agent, who will output them
using a text-to-speech component.
The rest of the paper will give more details about the strategy of adaptive
comment generation implemented.
2 The Comment Generation Subsystem
The language generation task consists in producing meaningful natural language

comments from comment structures provided by the expert system for each snapshot
of the scenario. This process has to take into account temporal constraints imposed by
the dynamics of the scene and other parameters such as the level of expertise of the
users and their preferences, expressed by suitable controls in the interface.
394 A. Esuli et al.
As traditional in NL generation [4,5] this task is performed in two phases. First, a

contents selection phase is performed in order to discard repetitions, select and
organize the concepts to be communicated. In the second phase, concepts are
rendered as natural language text [6]. This two phases however are not strictly
sequential; in fact contents selection needs to be performed again, after text
generation, when the time available does not allow completion of the generated
comment. We regard this interplay of the contents selection and generation phases as
one of the main characteristics of our approach.
2.1 Contents Selection
A comment is represented by a structure whose basic elements are primitive facts,

represented as sentences composed of a subject (s), a predicate (p) and an object (o).
Two parameters are associated to each primitive fact:
• the importance: a value rating the relevance of the fact in the present context;
• the difficulty level: specifying the language complexity class for which the fact is
appropriate, ranging from a highly specialized language to an ordinary language.
Here is an example of primitive fact, represented in Jscript syntax:
{s:”B”, p:”is”, o:”beyond herd to wind”, i:400, d:[0,1]}
where i is the importance value and d assigns difficulty levels to the fact.
Each comment structure is composed of a number of primitive facts related by
predefined relations, namely:
• Facts-Additional facts (FA): additional facts are complements to the main facts
• Facts-Explanations (FS): explanation facts further describe main facts.
• Cause-Effect (CE): cause-effect relation between the two groups of facts.
The expert system produces a comment for each fired rule, thus the same primitive
facts may appear in more than one comment. A pre-processing step takes care of
collapsing such duplicate facts into one instance. As a consequence, facts belonging
to different rules may become related.
The main task in contents selection is the selection based on the richness level, as
selected by the user through the interface. This is based on a threshold value (similar
to [7]), computed as the weighted sum of a static factor and a dynamic one. Primitive
facts that have an importance level above the threshold pass the selection.
The static factor is a percentage, which depends on the richness of the comment,
of the maximum importance value of the facts in the current snapshot. The dynamic
factor is a percentage of the threshold value of the previous snapshot and it is used for
allowing the completion of important comments at the expense of successive less
important ones (Fig. 3). This aspect is strongly related to the extension strategy
described next.
2.2 Comment Generation
In principle selected comments should be all presented but often this is impossible
due to time constraints. The generation phase consists in a cycle where comments are
progressively discarded until temporal constraints are satisfied. At each iteration two
main steps are performed for comment generation:
Adaptive Real Time Comment Generation for Sail Racing Scenarios 395
• Linearization: the selected facts are ordered according to their relations in order to
construct a plan of the comment sentence (text design).
• Realization: the natural language sentence is produced according to the previous
linear structure (text realization).
Fig. 3. Evolution of the threshold.Discarded comments are darker.
At the end of each iteration if the generated text comment can be output during the
animation time (according to an experimental estimation) the cycle is terminated and
the sentence is sent to the text-to-speech synthesizer of the Microsoft Agent.
Otherwise an attempt is made to extend the comment to successive snapshots.
This extension process activates the selection phase on the next snapshot; if no
comment remains after selection, this snapshot is marked as silent and the time
available to pronounce the comment is extended with the animation time of this
snapshot. If the time is still not sufficient, the function is called recursively on
successive snapshots until the end of the planning window is reached.
If the necessary time cannot be found by this strategy, two alternatives remain:
• In case of a reproduction modality that we call guided by the comment, the
commentator slows down the animation by increasing by a small percentage the
animation time and repeats the extension cycle. No comment is discarded.
• If the reproduction modality is fixed, the commentator is forced to discard the least
important comment in the current snapshot and the generation cycle is repeated.
The problem of real time comments has been tackled by RoCCo (RoboCup
Commentator) in the context of simulated soccer [8]. Our approach to comment gen-
eration however is not based on text templates but on concept structures more similar
to the ones used for example in the ILEX system [9].
References
1. Cisternino A., Del Cioppo E., Marotta P., Pacini G., Silenzi A., Simi M.: An expert
authoring tool for dynamic scenarios, Proc. of ED-MEDIA, (2001)
2. Kitano, H., Asada, M., Osawa, E., Noda, I., Kuniyoshi, Y., Matsubara, H.: RoboCup: A
Challenge Problem for AI, AI Magazine, Vol. 18, No. 1, (1997)
396 A. Esuli et al.
3. Microsoft Agent Software Development Kit, Microsoft Press, Redmond, (1999)

4. Cole, A., et al: Survey of the State of the Art in Human Language Tecnology, Kluwer,
Dordrecht, (1996)
5. Wahlster: Intelligent Multimedia Interface Agents, Tutorial MICAI 2000, (1999)
6. Feiner S.K., McKeown K.R.: Generating Coordinated Multimedia Explana-tions,
Proceedings of the 6th IEEE Conference on A.I. Applications, (1990)
7. Fraczak L., Lapalme G., Zock M.: Automatic generation of subway directions: salience
gradation as a factor for determining message and form, Proceedings.of the Ninth
International Workshop on Language Generation, Canada, (1998)
8. Voelz D., André E., Herzog G., Rist, T.: RoCCo: A RoboCup Soccer Commentator System,
Lecture Notes in Computer Science, Vol 1604, (1999)
9. Oberlander J., O’Donnell M., Knott A., Mellish C.: Conversation in the museum:
experiments in dynamic hypermedia with the intelligent labelling explorer, New Review of
Hypermedia and Multimedia, Vol. 4, (1998) 11-32
A Web-Based Selling Agent That Maintains Customer
Loyalty through Integrative Negotiation
Michael Grimsley1 and Anthony Meehan2
1School of Computing & Management Science, Sheffield Hallam University, Pond Street,
Sheffield S1 1WB, UK
[email protected]
2 Faculty of Mathematics & Computing, The Open University, Walton Hall, Milton Keynes
MK7 6AA, Buckinghamshire, UK

[email protected]
Abstract. In many transactions, the seller's objective includes promoting cus-

tomer loyalty in order to increase the likelihood of further transactions. Integra-
tive bargaining styles foster positive relationships between parties. This short
paper describes a protocol for a web-based selling agent that adopts an integra-
tive selling style to identify alternative sales contracts that match customer pri-
orities and hence promote customer satisfaction.
1 Introduction
Part of the value of many sales transactions is realised by securing repeated business.
Thus, customer loyalty is an important outcome of any transaction. We describe a
protocol for an adaptive web-based selling agent that is designed to engender loyalty
in a customer by adopting a selling style that promotes confidence and trust. This is
achieved through an integrative negotiation style using cues from the customer's in-
teraction to infer customer priorities and respond positively to them.
Negotiation theory provides a framework within which to understand selling. A
successful sale can be viewed as a contract negotiated between a vendor of goods or
services and a prospective customer. An integrative bargaining style seeks an agree-
ment that allows all parties to feel their priorities have been acknowledged and incor-
porated in the agreed deal. Integrative behaviour has a stabilising effect on relation-
ships: parties enjoy the relationship and seek to help the relationship persist [1].
An important tactic in integrative negotiation is logrolling. This involves trading
concessions on different issues, typically, when an issue is marginal to one party but
significant to another. For a selling agent to engage in logrolling, it must infer the
relative importance of issues from the earliest stages of an interaction with a new
customer.

398 M. Grimsley and A. Meehan
2 Selling Agent Protocol
A high-level description of an integrative protocol for a selling agent is as follows.

During the sales dialogue, the selling agent infers the customer's priorities from the
(short) sequence of contract proposals. The agent searches for alternatives that are
most similar to those proposed by the customer but which maintain value for the
vendor. The measure of similarity improves on conventional measures of similarity
[2,3] by incorporating the inferred the priorities of the customer. The logrolling tactic
is realised by selecting those similar alternative proposals that reflect the issue priori-
ties of both vendor and purchaser.
To implement this protocol, we specify the following: a means to represent sales
contracts for each agent; a metric which infers the relative importance of issues for a
customer; a means of expressing relative issue importance as weights in a weighted
similarity measure over pairs of alternative contracts; the construction of a set of
alternatives that reflect the relative priorities of both vendor and customer.
An agent interprets alternative proposals on the basis of a set of issues or attributes,
some public, some private [4]. An agent-centred representation of a sales contract is
c = (..., p i ,..., q j , ...); ( p i , q j ∈ P s ). (1)
s
P is the set of public (pi) and private (qj) issues for a selling agent s. (Note: we do not
require agents to share representations.). Table 1 provides an example of a sequence
of eight contract proposals over six issues. Issue 1 is conceded most reluctantly, issue
2 is conceded gradually from the outset, as perhaps is issue 5, but in this case rela-
tively larger concessions are made earlier on. Issues 4 and 6 seem to vary without any
obvious intent. (Note: the issues are not necessarily independent of each other.)
Table 1. Example sequence of sales contract proposals from a customer.

Issue (p)
Contract (i) 1 2 3 4 5 6
1 700 610000 7900 10 300 40
2 600 940000 10000 100 200 30
3 700 1100000 10000 60 400 40
4 900 1250000 12000 200 1500 200
5 1500 1800000 25000 110 1400 240
6 1650 1980000 27500 121 1540 264
7 2400 1800000 43000 120 2700 300
8 18000 2330000 20200 1000 3300 20
Having found no work in the literature on assessing the relative importance of is-
sues over such short temporal interaction sequences we have developed an empirical
metric that scores issues for importance over a short sequence of multidimensional
proposals.
A Web-Based Selling Agent That Maintains Customer Loyalty 399
The function, Pen, penalises relatively large early concessions and/or seemingly
non-goal directed variation in the values taken by issue p over a sequence of k con-
tract proposals:
ak  ( k −1)  (2)
Pen p , k = max  c k ( p ) − c i ( p ) i =1 .
Rp  
Rp is the known range of values for issue p. ak is chosen to be decreasing over 1..k
so as to penalise relatively early concessions.
The obvious cumulative penalty, CPen, for issue p after k proposals is simply,
∑ (3)
k
CPen p , k = Pen p ,i .
i =2
There are many ways to express this measure of issue importance as a weight. We
wish high penalties to map to low weights and so choose
CPen p , k − min (CPen i , k )ip=1 (4)

wp = 1−
max (CPen i , k )ip=1 − min (CPen i , k )ip=1
Given any two contracts, ci, cj, the similarity of these contracts for selling agent, s,
may be given as
Sim s (c i , c j ) = F p∈P s w p Sim sp (c i ( p ), c j ( p)). (5)
F is a function the choice of which is domain dependent. (The choice Fp∈Ps= Σp∈Ps is
common but assumes independence of the issues.) wp is a weight associated with an
issue, p, by the agent (see below) and may be normalised (Σp∈Pwp=1). Simp is a spe-
cific similarity measure for issue, p [5]. In the context of sales contracts not all issues
take values that are numeric. Hence, a fuzzy measure of similarity is appropriate
[6,7]. Following [5] and [6], a fuzzy similarity function can be defined as
Sim sp (c i , c j ) = Inf 1≤i ≤ n ( hi (c i ( p)) ↔ hi (c j ( p))). (6)
where Inf is infimum and h (ci)↔hi(cj) is a fuzzy equivalence relation determined by

the comparison function, hi. There are many possible choices of hi, examples include
[7,3]
(7)
h ( c i ) ↔ h ( c j ) = 1 − h ( c i ) − h (c j )
and
h(ci ( p )) ↔ h(c j ( p )) = min(h(c i ( p )) / h(c j ( p)), h(c j ( p)) / h(c i ( p))). (8)
400 M. Grimsley and A. Meehan
Given the expressions above and a current contract proposal, ck, from the cus-
tomer, a set of alternative contracts can be generated which reflect both the vendor's
and the customer's priorities.
L(c k ) = Sim s (c k , c j ) ∩ Sim c (c k , c j ) (9)
L(ck) contains contracts whose strength of membership is given by fuzzy intersec-

c
tion [8]. Sim incorporates the vendor's estimate of the customer's relative priorities
s
using expressions 2-4. The issues and weights for Sim may be elicited directly from
the seller. (This is often a significant knowledge elicitation task and there may be
significant difficulties if there are many highly correlated issues [9].)
The membership function of the set of 'logrolling' contracts in L(ck) reflects the is-
sue priorities of both the selling agent and the customer. The selling agent can now
select alternative proposals from L(ck), choosing those that have strong membership
and at least maintain, and possibly increase, value to the vendor.
3 Conclusion
This paper has presented a protocol for an adaptive web-based selling agent. It has
focused on using inferences about a customer's priority issues to help the selling agent
find alternative sales contracts that are integrative and thus give the customer a sense
that their priorities are acknowledged and responded to. This offers the prospect of
enhanced customer confidence and loyalty.
References
1. Pruitt, D.G.: Negotiation Behaviour. Academic Press (1981)

2. Hahn, U., Ramscar, M.,: Similarity and Categorisation. Oxford University Press (2001)
3. Everitt, B.S., Landau, S., Leese, M.: Cluster Analysis. Arnold, London (2001) 35-53
4. Barker, R., Holloway, L.P., Meehan, A.: A negotiation model to support material selection
in concurrent design. In Proceedings of the 14th International Conference on Engineering
Applications of AI and Expert Systems. Monostori. L, Vancza, J. and Ali, M. eds. Lecture
Notes in Artificial Intelligence 2070; Springer-Verlag, (2001) 697-707.
5. Sierra, C., Faratin, P., Jennings, N.R.: Deliberative Automated Negotiators Using Fuzzy
Similarities. Proc EUSFLAT-ESTYLF Joint Conference on Fuzzy Logic, Palma de Mal-
lorca, Spain, (1997) 155-158.
6. Valverde, L.: 1985 On the structure of F-indistinguishability. Fuzzy Sets and Systems
17:313-328 (cited in Sierra, et al 1997 above.)
7. Yao, Y.Y.: Qualitative similarity, Soft Computing in Industrial Applications, 4th Online
World Conference on Soft Computing in Industrial Applications (WSC4), September 21-
30 (1999) 339-348.
8. Zadeh, L.A.: Fuzzy Sets. Information and Control. 8 (1965) 338-353
9. Grimsley, M.F., Meehan, A.S.: Perceptual Scaling in Materials Selection for Concurrent
Design. FLAIRS 2002, Pensacola Florida. AAAI Press, (2002)
Adaptive Content for Device Independent
Multi-modal Browser Applications
Jennifer Healey, Rafah Hosn, and Stéphane H. Maes
Human Language Technologies Group

IBM T.J. Watson Research Center
Route 134, Yorktown Heights, NY 50198
{jahealey,rhosn,smaes}@us.ibm.com
Abstract. Adapting content appropriate to the device and modality of

a user’s preference becomes more important as users begin to expect
universal access to information, whether they are on the phone, on a
desktop or using a PDA. This paper outlines the design of a travel ap-
plication authored using an XForms compliant language and deployed
using a DOM-based MVC multi-modal browser. The travel application
authored in a language of conversational gestures that can be transcoded
into multiple synchronized views for access via a variety of devices.
1 Introduction
The next generation of human computer interaction will implement multiple
forms of input and output. People will expect the information they need to
be available to them where ever they are and from whatever device they are
currently using. This paper describes a device independent application written
using an XForms [9] based meta-language that is adaptively transcoded into
multiple synchronized presentation languages (e.g. HTML, VXML and XHTML-
MP etc.) using XSLT [10] rules. The XSLT transformation can adapt content
and style to suit the mode of interaction and the restrictions of the device.
The XForms application is served using a DOM-based MVC [1] multi-modal
browser (MMBrowser) [4] with an added forms based dialog manager (FDM).
The MMBrowser contains a controller module that acts as a hub to all modules
inside the MMBrowser as well as a mediator between the devices, the application
server and the FDM. This module is also responsible for device registration
and synchronization. The added FDM is an adaptation of a previous dialog
manager [7] developed for the DARPA Communicator competition [2] where
XForms is now used for the data and interaction models, DOM level two events
are now used for communication with the MMBrowser and VoiceXML is used
as the speech input/output voice channel.
2 Multi-mode and Multi-channel

Computer systems are being developed that support a variety of modes of in-
teraction including speech, keyboard, handwriting, gesture, facial expression,

402 J. Healey, R. Hosn, and S.H. Maes
context and affect. These multiple modes of interaction can be used to facilitate
both recognition [6] and communication [5,11]. Often these modalities require
specialized channels to allow access modalities such as cameras, microphones and
sensors. A multi-modal multi-channel system faces the challenge of accepting in-
formation from any input method and delivering information through the appro-
priate output methods. In this paper, we show how a declarative language based
on XForms [9] encodes user interaction as a set of conversational gestures [8]
and binds these gestures to an XForms data model using XPath expressions.
The interaction is presented to the user after being transformed by XSLT [10]
rules adapted to the user and the device. The dialog manager insures that an ap-
propriate sequence of inquiry and response gestures continues the conversation.
The MMBrowser uses standard DOM events to synchronize modalities, allow-
ing the user to switch modalities during a conversation so that voice, keyboard
or mouse can be used interchangeably. The framework [3] derives its power by
separating application content from its presentation and providing a controller
that manages the interaction with the user while updating the application data
following the model view controller programming paradigm. The framework also
leverages evolving XML-based industry standards for modeling application con-
tent and representing user interaction as well as for communicating the results
of user interaction among various components of the system.
3 Authoring Applications
There are three components to application authoring in this framework: the user
interaction; the data model and the presentation. This section describes each of
these components and the relationships between them.
3.1 User Interaction: Encoding Conversation
The multi-modal application is written in a simple, declarative language de-

signed to encode conversation at the highest level consistent across modalities.
A set of primitive conversational gestures including “dialog”, “message”, “in-
put”, “select” and “submit” is used to specify the behavior of the application.
For example each message or input that the system uses to present or collect
data is simply defined and can be bound to information the data model using
XPath. To illustrate, the following is an excerpt from a travel application written
for the system:
<dialog name="trip/air" action="submit">

<message>
Welcome <value name="trip/user/firstName"/>
<value name="trip/user/lastName"/>!
</message>
<input name="arrivalCity">
Adaptive Content for Device Independent Multi-modal Browser Applications 403
<grammar src="travel.gram"/>
<caption> Where will you be flying to?</caption>
</input> ...
<submit target="http://localhost/DB"/>
</dialog>
This example shows a “Welcome” message gesture bound to the firstName

and lastName data model instance nodes for output presentation and an “ar-
rivalCity” input gesture bound to the arrivalCity data model instance node. The
message uses full XPath expressions such as “trip/user/firstName” while the
input constructs the path “trip/air/arrivalCity” from both the binding in
the nearest dialog gesture name=”trip/air” and the binding inside the input
gesture name=”arrivalCity”.
3.2 Data Model Specification
The data model is specified in XML and is distinct from the presentation and
the controller. An example of a data model containing the passenger and air
travel information could be specified as:
<?xml version="1.0"?>
<trip>
<user>
<firstName/>
<lastName/>
<password/>
</user>
<air>
<departureCity/>
<departureDate/>
<arrivalCity/>
<returnDate/>
</air>
</trip>
Each of the slots in this model such as “firstName” and “departureCity” will
be populated through interaction with the user. The dialog structure is flexible
and allows users to fill multiple slots, in various order, in a single interaction. The
dialog manager uses the data model to keep track of the conversation context to
determine which slots have been filled with information from the user and which
slots still need to be presented to complete the form.
3.3 Presentation Rules
Speech, text and gesture all serve as a means to the same end, but the methods
are inherently different. In a visual presentation the application author might
404 J. Healey, R. Hosn, and S.H. Maes
wish to have strict control over the spatial presentation of information, however
in speech conversation the system should adapt to receive whatever information
the user chooses to present and respond appropriately. In this application, the
XSLT rules allow the author to specify the details of data presentation in a par-
ticular modality while the XForms language allows information to be collected
from the user in a flexible order across modalities.
Fig. 1. An example of the synchronized text modality response synchronized to the

voice input “I want to fly to New York on third Friday in July and return a week later
on Monday”
For example the designer can exactly specify the exact order, font, size, color
and placement of visual data by including special tags in XSLT rules for HTML:
<xsl:template match="dialog[@format_key=’D2’]">
<p>I want the best round-trip travel:</p>
<table><tr><td><b>From:</b></td>
<td><xsl:apply-templates select="input[@select_key=’I4’]"/>
</td><td><b>To:</b></td>
<td><xsl:apply-templates select="input[@select_key=’I5’"/>
</td></tr> ... </table> ...
</xsl:template>
In addition, individual users with special needs might use special XSLT rules to
show larger fonts, bigger buttons or to allow selection input through gesture or
alternative input devices.
Presentation customizations are entirely separate from the interaction logic
layer and the data model. As a result, a visual presentation might show as output
particularly placed “From:” and “To:” fields while the voice channel might say
as output “Where would you like to go?.” The separation of presentation from
the underlying interaction logic also allows users the flexibility to express input
in different ways through different modalities. For example in the presentation
shown in Fig. 1, the “From” field precedes the “To” field, since visually the user
might think of a right to left line describing the trip. However, the user can still
verbally express the information in whatever order they prefer, e.g. “I want a
flight to New York, leaving from Boston.” Since the gestures are synchronized
and bound to the same data model using XPath, when the user speaks this
response the information will appear in the reverse order in the appropriate
fields of the HTML display.
Adaptive Content for Device Independent Multi-modal Browser Applications 405
4 Discussion
This system separates data from presentation which allows ubiquitous adapta-
tion of information to the mode and device of the user’s preference. Multiple
modalities can be used collectively to enhance user experience or modalities
can be used independently to allow access through a greater variety of devices.
Applications can be simply written using a declarative XForms language and
additional sophistication can be added to presentation by creating specialized
formatting using XSLT. A single application and data model are used to model
the conversation synchronously across modalities. A dialog manager keeps track
of the conversation so that information is presented at a level that best suits the
user’s modality, channel and context. This system allows content to be adapted
to a variety of devices with different modes of access, allowing the user the great-
est freedom of expression when available and the greatest access to information
from the most convenient device.
References
1. S. Burbeck. Applications programming in smalltalk-at: How to use model-view-
controller - mvc. http://st-www.cs.uiuc.edu/users/smarch/st-docs/mvc.html.
2. M. W. et al. Darpa communicator dialog travel planning systems: The june 2000
data collection. In Proc. of Eurospeech ’01, 2001.
3. R. Hosn, T. Raman, and S. H. Maes. Single application model, multiple synchro-
nized views. In Proc. ICME ’01, Tokyo, Japan, 2001.
4. J. Keindienst01. A dom-based mvc multi-modal e-business. In Proc. ICME ’01,
Tokyo, Japan, 2001.
5. R. Lau, G. Flammia, C. Pao, and V. Zue. Webgalaxy – integrating spoken language
and hypertext navigation. In Proc. Eurospeech ’97, pages 883–886, Rhodes, Greece,
1997.
6. C. Neti, G. Potamianos, J. Luettin, I. Mathews, H. Glotin, and D. Vergyri. Large
vocabulary audio-visual speech recognition - a summary of the johns hopkins sum-
mer 2000 workshop. In Proc. IEEE workshop on Multi-media signal processing,
pages 619–624, Cannes, France, 2001.
7. K. A. Papineni, S. Roukos, and R. T. Ward. Frames based dialog management using
forms. In Proc. Eurospeech ’99, volume 3, pages 1411–1414, Budapest, Hungary,
September 1997.
8. T. Raman and S. H. Maes. Multi-modal interaction in the age of the information
appliance. In Proc. ICME ’02, New York, July, 2000.
9. W3C. Xforms 1.0 working draft. http://www.w3.org/MarkUp/Forms, 2001.
10. W3C. Xslt. http://www.w3.org/TR/xslt, 2001.
11. V. Zue, S. Seneff, J. Polifroni, M. Phillips, C. Pao, D. Goodine, D. Goddeau, and
J. Glass. PEGASUS: A spoken dialogue interface for on-line air travel planning.
Speech Communication, 15:331–340, 1994.
Behavioral Sequences: A New Log-Coding Scheme
for Effective Prediction of Web User Accesses
1 2
Rushed Kanawati and Maria Malek
1
LIPN-CNRS UMR Q 7030, Av. J. B. Clément, F-93430 Villetaneuse
[email protected]
2
LAPI-EISTI, Av. du Parc, F-95011 Cergy
[email protected]
Abstract. Mining web site logs for predicting user actions is a central issue in
the field of adaptive web site development. In order to match the dynamic
nature of today web sites we propose in this paper a new scheme for coding
Web server log data into sessions of behavioral sequences. Following the
proposed coding scheme the navigation sessions are coded as a sequence of
hypothetical actions that may explain the transition from one page to another.
The output of a prediction algorithm will now be an action that can be evaluated
in the context of the current navigation in order to find pages that to be visited
by the user.
1 Introduction
One hot topic in the field of adaptive web site development is the problem of mining
usage log data in order to predict future user actions [1,3,5]. Web-user access
prediction can be used for a variety of different applications such as: prefetching web
pages [9] and recommending links and pages [4,7,8]. In almost all-existing
approaches log data are pre-processed to form a set of navigation sessions [2,6]. A
session is defined as a chronologically ordered sequence of Web pages visited by a
given user. Most web user access approaches apply some sequence prediction
algorithm on navigation sessions. By the very nature of used data, existing prediction
systems can only predict access to pages that have been visited earlier by some users
(e.g. pages that have been registered in the log file). This feature is a serious
limitation of the usability of prediction systems for adapting today highly dynamic
Web sites. In order to match the dynamic nature of web sites we propose in this paper
a new scheme for coding log data into sessions of behavioral sequences. A behavioral
sequence is defined as an ordered list of elementary behaviors. Each elementary
behavior describes a visited Web page associated with a set of hypothetical actions
that may explain how the user acts to reach the next visited page. Using this coding
scheme the output of a prediction algorithm will consist on predicting the action to be
made by the observed user. Actions can be evaluated in the context of the observed
user navigation in order to find pages to be visited later by that user. Next, we
describe briefly the concept of behavioral sequences.

Behavioral Sequences: A Log-Coding Scheme for Prediction of Web User Accesses 407
2 Turning Log Data into Behavioral Sequences
For sake of clarity, we start by introducing some notations that are used later in this
paper.
- Web page (P). A web page P is defined as a triple P = <U, L, I>. Where U is
the page address (i.e. the page URL), L is the list of links contained in the page P,
and I is some information record about the page (i.e. the record I may contain a
description of the page content, the page author, etc). We denote by L(P)
(respectively U(P) and I(P)) the U (respectively L and I) component of the
page P.
- Web site (S). A web site, denoted S, is a set of n web pages. We write S = {Pi,
0<i<= n}.
- Navigation (N). A user navigation session in a site S denoted N, is a sequence of

steps si. We write N = [si]i=0..m where m is the navigation length.
- Step (s). A step s is defined as a couple s = <Pi ∈ S, SI> where Pi is a web page,
and SI is some information structure about the step. The SI structure can contain
entries about request result (i.e. successful access, page not found, etc), page
th
visualization time, user satisfaction from the page, etc. The i step in a navigation
j j j
Nj is designated by Nj[i] = s i. We use the notation P(s i) (respectively I(s i)) to
designate the page (respectively information) component of a step.
j
- Transition (T). We call a transition T i a sequence of two pages visited
j j
sequentially in navigation Nj at steps i and i+1. Formally, we write T i= [P(s i),
j
P(s i+1)].
- Navigation Base (NB). We define a navigation base NB as a set of past

navigation sessions that have occurred in the site S.
- Navigation action template. We define a navigation action template, denoted

j
NATk, as a generator of hypothesis that can explain a transition Ti occurring in
navigation Nj. The set of all used templates is denoted by NAT. The composition
of the NAT set is site-dependent. Next, we give some examples of different
templates based only on the definition of a hypermedia navigation process. The
given list is far from being exhaustive and it should be taken just as a simple
example of hypothesis we can generate to explain transitions in a navigation. The
four examples are the following:
1. NAT1 : Follow a link. If page Pi contains a hypertext link that points to Pi+1,
then one plausible hypothesis that can explain the transition from Pi to Pi+1 is
that the user has clicked that link. Two sub-types can be derived from NAT1
type: Follow a link with anchor A (NAT11) and Follow a link with rank X
(NAT12).
2. NAT2. Move backward. If page Pi+1 has already been visited in the same
navigation then a plausible transition explanation would be that the user has
408 R. Kanawati and M. Malek
made a backward move. Once again, we can derive several sub-types of that
action type such as moving backward n steps, moving up to the start, etc.
3. NAT3. Page Reload. If page Pi+1 is the same as Pi then one plausible
explanation is that the user has reloaded the same page.
4. NAT4. Follow an URL. A simple transition explanation is that the user has
simply typed the address (i.e. the URL) of the page Pi+1 after visiting page Pi.
The application of a template, NATi, on navigation Nj allows to generate for each

j
transition Ti a set of hypothetical actions, defined by the template NATi that may
explain the transition. Let Ai be the set of concrete actions from type NATi that can
j j
explain the transition Ti , we write: Ai= Apply(Nj,NATi, Ti )
The generated navigation action set can be the empty set. For example consider the
web site illustrated in Fig. 1. Consider a navigation Nj in that site. Nj is composed of
the following steps:
Nj = [<P1, SI1>, <P3, SI2>, <P4, SI3>, <P4, SI5>]
P2
P1
L1 A
L2 B P4
P3
L3 A
L4 C L1 A
L2 D
Fig. 1. Example of a simple Web site
The following table illustrates the result of applying the different templates defined
above on each simple transition in the navigation Nj.
Table 1. Example of generated actions by applying navigation action templates
Transition NAT11 NAT12 NAT3 NAT4

Nj[1,2] follow link with anchor B follow link 2 - follow URL: P3
follow link with anchor A follow link 3
Nj[2,3] follow link with anchor A follow link 1 - follow URL P4
follow link wit anchor D follow link 2
Nj[3,4] - - Reload follow URL P4
- Action evaluation function (Eval). We define the function of action evaluation

as follows:
eval : A x NB x S → ℘(S)
Behavioral Sequences: A Log-Coding Scheme for Prediction of Web User Accesses 409
Where A is the set of all actions, NB is the navigation base and S is the web site. The
evaluation of an action a ∈ A is made in the context of a given navigation Ni ∈ NB
and on a given page Pi ∈ S. The result of the evaluation is either a set web page {Pk}⊆
S or the null (φ) value. For example consider the action Ax: follow link with anchor
equal to "A" to be evaluated for each page in the navigation example given above
(Fig. 1). Evaluation results are given in the following table.
Table 2. Evaluation of the action follow link with anchor A in the context of the example
illustrated in Fig. 1.
Page Evaluation result

P1 {P2, P3}
P2 φ
P3 {P4}
P4 φ
- Elementary behavior (EB). Given a navigation Nj. With each step s i ∈ Nj we can
j
j j j
associate an elementary behavior EB i, defined as follows: EB i = <s i, LAi, CFi>
where:
• LAi is the set of all actions from all types that verify the following
condition: ∀ aj ∈ LAi : eval (aj, Nj, P(si )) = P(s i+1).
j j
• CFi is a vector that represents the system confidence in each action listed
in LAi.
Using the concept of elementary behavior we can now represent a navigation session
Nj as a sequence of elementary behaviors Nj= [EBi]i=0..N where each EBi represents the
j
elementary behavior associated to the step si in Nj. Turning web access log into
behavioral sequences requires modifying the Web server. The following algorithm is
to be executed on each ongoing navigation: No =[s1,…, sm].
LA ←Φ; // LA the list of actions

For each NATi ∈ NAT do: // Computing all actions of all types that
m
m
A = Apply(NATi,No,To ) // could explain the transition To
LA ← LA ∪ A
N = |LA| // N is the number of found actions
For i=0 to N do: // Computing the default confidence in
CF[i]=1/N // each action
EB = <Sm, LA,CF> // Setting the elementary behavior of rank
No[m] = EB // m in navigation No.
Applying a prediction algorithm on the obtained behavioral sequences allow

getting as output an action or set of actions. These actions can be evaluated, using the
eval function, in order to indicate pages to be visited by the user. An example of using
this scheme is detailed in [4].
410 R. Kanawati and M. Malek
3 Conclusion
In this paper we have introduced a new scheme for coding web log data. The
proposed scheme transforms log data into behavioral sequences. A behavioral
sequence represents a sequence of pages visited by the user. Each page is associated
with a set of hypothetical actions that can explain the user motivation to go from one
page to another. The main advantage of this new coding scheme consist on allowing
to apply sequence prediction algorithms to predict actions to be made by the user
rather than pages to be visited by the user as the majority of existing systems do. This
feature is essential for predicting user actions in dynamic web sites.
References
1. P. Brusilovsky, Methods and Techniques of Adaptive Hypermedia. In User Modelling and

User-Adapted Interaction 6: 87-129, Kluwer academic publishers, 1996.
2. R. Cooley, B. Mobasher and J. Srivastava. Data Preparation for Mining World Wide Web
Browsing Patterns. In Knowledge and Information Systems 1(1). pp. 5-32, 1999.
3. P. De Bra, Design Issues in Adaptive Web-Site Development. In Proceedings of the
Second Workshop on Adaptive Systems and User Modelling on the World Wide Web, pp.
29-39, Toronto and Banff, Canada, 1999. (Editors P. Brusilovsky and P. De Bra), 1999.
4. M. Malek, R. Kanawati. COBRA: A CBR-Based Approach for Predicting Users Actions
th
in a Web Site. In (D. Aha and I. Watson eds.) Proceedings of the 4 International
Conference on Case-Based Reasoning (ICCBR 2001), (LANI 2080) Springer, Vancouver
BC., Canada, July/August 2001. pp. 336-346.
5. A. Mobasher, R. Cooley and J. Srivastava. Creating Adaptive Web sites Through Usage-
Based Clustering of URLs. In proceedings of the IEEE workshop on Knowledge and Data
Engineering (KDEX'99), 1999.
6. J. Pitkow. In search of reliable usage data on the WWW. In Sixth International World
Wide Web pages 451 463, Santa Clara ,CA, 1997.
7. M. Spiliopoulou, C. Pohle and L. C. Faulstich. Improving the Effectiveness of a Web Site
with Web Usage Mining. In proceedings of KDD workshop WebKDD'99, San Diego,
August 1999.
8. B. Trousse, M. Jaczynski, and R. Kanawati. Using User Behaviour Similarity for
Recommendation Computation: The Broadway Approach. In H-J Bullinger and J.
Ziegler, editors, proceedings of the HCI International (HCI’99), Munich. pp 85-89.
Lawrence Erlbaum Associates, august 1999
9. I. Zukerman, David W. Albrecht, and E. Nicholson, Predicting users' requests on the
WWW, Proceedings of the Seventh International Conference on User Modelling, 1999.
A Fuzzy-Based Approach to User Model Refinement
in Adaptive Hypermedia Systems
Cveta Martinovska
Electric Engineering Faculty, Karpos II

1000 Skopje, Macedonia
[email protected]
Abstract. This paper presents an approach for refinement of the user model in
hypermedia applications using fuzzy adaptation rules. The rules register
relevant user browsing actions and this way change the strength of relationship
between the user model attributes and the concepts of the knowledge domain.
The process of fuzzy inference is formalized with fuzzy automaton. Using
appropriate authoring tool fuzzy automata provides a method for detecting
conflicts and inconsistency in the set of rules, in an automated fashion.
1 Introduction
Managing personalized views over the information spaces is the main problem in
developing adaptive hypermedia systems. Models that represent relevant user aspects
as goals, interests, cognitive abilities, decision styles or way of learning enable user
oriented adaptivity. The system gathers the knowledge about the user through the
initial interview and/or continuous monitoring of the user's browsing behavior.
Hypermedia systems offer freedom in the navigation through the hyperspace but
more of the users prefer guided over unguided browsing. The user will lose the
interest if the links he/she follows are meaningless, and needs to go back to the links
related to the area of interest. In the hypermedia systems used in education the system
has to keep track what the user has read in order to avoid presenting redundant
information or pages that the user can not understand. There are several ways to
overcome the navigation problems in adaptive systems like restructuring the link
configuration according to the user models, changing the content of the nodes,
supporting the user browsing by showing the links that are most suitable in certain
state. An overview of methods and techniques for adaptive hypermedia systems can
be found in [1].
In adaptive systems user model is usually independent from the domain
knowledge. A discussion about modeling an application domain and user modeling in
adaptive hypermedia systems and especially in Web-based systems is presented
in [2]. The domain model is concerned with the content and structure of the
information that has to be presented to the user. It deals with the link relationships
between the concepts as well as with the decomposition of the concepts in a hierarchy
of sub-concepts and atomic information like paragraphs of text, sounds, videos, and
images. Part of the adaptive hypermedia systems is a mechanism that performs
adaptation based on the user model and updates the user model. This mechanism
might be in the form of adaptive rules as described in [3]. The if -then rules might be

412 C. Martinovska
built in the adaptive hypermedia systems when they are not dependent on the domain
or provided by the author of the hypermedia application when they are domain
dependent.
Interactive systems for modeling user expectations, motives, abilities or interests
have to deal with imprecise knowledge, because users prefer to express this type of
information in terms of vague concepts. Several systems for user modeling applying
uncertainty techniques are described in [4].
In this paper is proposed refinement of the user model with fuzzy adaptation rules.
The formal description of the process is modeled with fuzzy automaton that maintains
a dynamic characterization of user knowledge and interests through the process of
browsing.
This paper is organized as following: Section 2 discusses the basic characteristics
of the fuzzy-based user model and the format of the rules for its revising. Section 3
gives a formal definition of fuzzy automaton corresponding to the fuzzy adaptation
rules. The concluding remarks and next research are presented in Section 4.
2 User Modeling and Rule-Based Adaptivity

In this section some relevant descriptions of the user model and the fuzzy adaptation
rules are provided.
2.1 User Model
The adaptive system has to construct and maintain a matrix containing the concepts
and their constitutional parts, and attributes like the information about the relevance
of the concept for the user and whether the user is familiar with the concept.
The natural way to characterize the relationship between the attributes in the user
model and concepts of the domain is using fuzzy linguistic labels. Fuzzy linguistic
labels are obtained from the term set T (A) of the linguistic variable A [5]. For
example, let A=interest then T(A)={very interested, quite interested, more or less
interested, not interested,...} where the hedges very, quite and more or less are used to
modify the meaning of the notion.
This paper is not concerned with the initialization of the user model. The user is
asked to select linguistic value for domain concepts in the initial dialog or he/she is
classified into some stereotypical category. In the following text the updating of an
individual user model is discussed representing single user characteristics relevant for
the knowledge domain.
2.2 Fuzzy Adaptation Rules
The user model is modified by fuzzy adaptation rules with the following general
pattern: if x1 is A1 and x2 is A2 and ... and xn is An or ... then y is B, where Ai and B are
linguistic variables from the universes of discourse U and V respectively, xi is an
instance in the universe U corresponding to membership value µAi(xi), y is an instance
in V with membership value µB(y).
A Fuzzy-Based Approach to User Model Refinement in Adaptive Hypermedia Systems 413
Condition part of the rule might for example include external events like selection
of certain links in the presented pages, access of some pages or checking the user
model for the relevance of the concept. Action part might specify new values for the
attributes in the user model or might activate other rules.
The operators and,or,not are defined as fuzzy operators. The author has to choose
appropriate membership functions and operator definitions. For example, he/she must
determine whether it is appropriate a single attribute to block the application of the
rule giving the operator and meaning minimum of the membership values.
More than one rule can be activated with the execution of some events leading to
conflict, which has to be signaled to the author, or the author might add a rule that is
inconsistent with the rest of the rules. Analyzing the process of fuzzy inference using
fuzzy automaton may help in detecting undesirable and ambiguous situations. That is,
this representation can be used as a basis for an authoring tool, which will discover
conflicting results or potential infinite loops.
3 Fuzzy Inference
The process of fuzzy inference is formalized with fuzzy automaton FA=(I,S,O,f,λ)
where I is a set of input linguistic variables, S is a set of automaton's internal states
and O is a set of output linguistic variables. Transition function f is defined from
SxIxS to {0,1} and output function λ is a function from SxIxO to {0,1}, where
f(si,ip,sj)=1 if there is link from state si to sj, and f(si,ip,sj)=0 in other cases, and
λ(si,ip,op)=1 if op is the output at state si when input is ip and λ(si,ip,op)=0 otherwise.
The fuzzy automaton has to be deterministic. That is, for a given input and current
state there is only one next state and output. So, some constraints are imposed:
function f has value 1 for exactly one next state sj being in state si when the input is ip
and λ has value 1 for only one output ok being in state si when the input is ip.
To obtain next states the automaton computes max-min operations from the current
state and inputs. Let current input x has membership values I=[µi1(x),... ,µip(x)] for
every input linguistic variable ik and S is the current state of the automaton distributed
over several states, where the degree of activation of the states is defined with value
'
in the interval [0,1]. The next state S is computed as fuzzy composition
S =S ° max[min(µi1(x),f(si,i1,sj)),...,min(µip(x),f(si,ip,sj))]
'
(1)
As an example, Figure 1 shows a fuzzy automaton which corresponds to the following
fuzzy rules:
if in subdomain D12 user selection of links is low then
set the attribute value12 to not interested
State input output next state

s1 medium more or less interested s2
s1 low not interested s3
Fig. 1. An example of fuzzy automaton based on fuzzy rules
414 C. Martinovska
if in subdomain D21 user selection of links is medium then

set the attribute value21 to more or less interested
The automaton has 3 internal states { s1, s2, s3 }, two input variables {low, medium}
and two output variables {not interested, more or less interested} and s1 is the initial
state. The membership functions for input and output variables are µmedium(x), µlow(x),
µnot interested(x) and µmore or less interested(x). For example, µlow(x) is a membership function
mapping from number of selected links to a value in [0,1] describing to what degree x
is a member in the set.
4 Conclusion
To customize the process of navigation through the hypermedia space user modeling
component may have the possibility to register the relevant browsing events and to
modify the user model to adequately reflect certain user interests, familiarity with the
concepts, domain expertise and his/her cognitive abilities.
This report presents an approach for updating the user model using fuzzy
adaptation rules formalized with fuzzy automaton. The set of rules that revise the user
model are domain specific rather than built in the adaptive hypermedia application.
Developing the user modeling component the author has to verify whether the set of
rules satisfies his/her expectations and to detect the potential conflicts in the rule set.
For analyzing the set of fuzzy rules an authoring application will be needed that
allows an author to express properties without having to be concerned with the
methods and techniques of fuzzy inference.
References
1. Brusilovsky, P.: Methods and Techniques of Adaptive Hypermedia. UMUAI, Vol.
4. Kluwer academic publishers (1994) 21-45
2. De Bra, P.: Design Issues in Adaptive Web-Site Development. Proceedings of the
2nd Workshop on Adaptive Systems and User Modeling on the WWW (1999),
29-39 (Workshop proceedings http://wwwis.win.tue.nl/asum99/, TU/c CNS99/07)
3. Wu, H., De Bra, P., Aerts, A., Houben, G.J.: Adaptation Control in Adaptive
Hypermedia Systems. Proceedings of the AH2000 Conference, LNCS Vol. 1892.
Springer-Verlag, Berlin Heidelberg New York (2000) 250-259
4. Jameson,A.: Numerical Uncertainty Management in User and Student Modeling:
An overview of Systems and Issues. UMUAI, Vol.5. Kluwer academic publishers
(1996)
5. Zadeh, L.A.: Outline of a New Approach to the Analysis of Complex Systems and
Decision processes. IEEE Transactions on Systems, Man, and Cybernetics,
vol.smc-3,1,(1973) 28-44
Towards an Authoring Coach
for Adaptive Web-Based Instruction
Judith Masthoff1
University of Brighton, UK
[email protected]
Abstract. A new approach to adaptive web-based instruction is outlined and

advocated, which is based on models and principles that allow the course
material to organize itself. This is particularly useful when course material
originating from multiple authors is to be combined into a coherent whole.
1 Background
In previous research, an “artificial teacher” has been developed, which adapts

instruction to the learner [1]. This artificial teacher needs an abstract description of
the course material (often called metadata). Additionally, there needs to be sufficient
variation in the course material to be able to reach a high level of adaptation. We are
developing a system (‘Authoring Coach’) that coaches authors to provide both the
metadata and the variation needed for the adaptation to be effective. It shall
• Provide an easy user-interface for entering course material and metadata.
• Stimulate authors to enter metadata by clarifying its purpose and consequences.
• Stimulate authors to provide variation by indicating the amount of adaptation
possible with the current material and how this can be increased.
• Enable multiple authors to contribute without need for coordination. An essential
aspect of the World Wide Web has been that is has organically grown: authors
from around the world (without coordination) have contributed material. We use
the same principle for the authoring of courseware.
• Generate personalized web-based lesson books from the material provided by the
authors, which are easy to use, have a good narrative flow, and allow the artificial
teacher to monitor and optimally support student learning.
2 Existing Systems
To model the teaching domain, both InterBook [2] and NetCoach [3] use concepts,
which are organized in a network, with links reflecting different types of relationships
between them. Concepts in InterBook are “elementary pieces of knowledge for the
1 The author is supported by Nuffield grant NAL/00258/G.

416 J. Masthoff
given domain”. All examples mentioned are noun phrases, like “production rule”. In
NetCoach concepts are ”internal representations of pages”, like “Chapter-2-1-2”. In
InterBook, the author provides an electronic textbook that is hierarchically organized
(chapters, sections, etc). Each page has a set of outcome concepts and a set of
prerequisite concepts associated with it (analogous to pre- and post-conditions in
programming). These are used to support adaptive navigation and hyperlink
annotation. NetCoach also uses prerequisites, but as relations between concepts. In
MetaLinks [4], authors provide the hierarchical relationship between pages.
3 Proposed Authoring Coach
We will describe here a number of aspects of our Authoring Coach that are vital for
supporting authoring by multiple authors, in a self-organizing way.
Use of action verbs in outcomes and prerequisites. The use of concepts to express
outcomes, as in InterBook, is not enough to make the outcomes sufficiently precise
and unambiguous. It can result in multiple pages in an electronic textbook that cover
apparently the same outcome. In the pedagogical literature, outcomes are always
described as containing an action verb, indicating what the student will be able to do
[5]. We will express each outcome as a combination of an action verb with a concept.
For instance, an outcome of a page will not be "search methods", but something like
"explain search methods" or "implement search methods". This will allow authors to
specify the content of a page more accurately.
Lists of recommended verbs exist, often classified into six levels of learning
according to Bloom's taxonomy [5]. This allows the Authoring Coach to advice
whether there is enough material at a suitable level. For example, in a third year
module, the students should display higher levels of learning, like analysis and
evaluation. The coach could give advice like "Please add pages that encourage critical
thinking, using verbs like evaluate". The verbs will also inspire authors about which
pages could still be added (comparable to the thematic links in MetaLinks). The
Authoring Coach will coach the authors, by using requests like "Please write a page
that covers outcome X", and "Please write a page that uses prerequisite X".
Generation of page titles, exercises, etc. Some aspects of the presentation of

material can be generated automatically on the basis of underlying information, such
as verbs and concepts (cf. [6]). A simple example relates to the generation of titles.
Table 1 shows some example mappings between verbs and titles. Similarly exercises
could also sometimes be generated. In both cases, an advantage of using generation is
that it allows better support of narrative flow and consistency in terminology.
Table 1. Example mappings between verbs and titles

Verb Title
implement CONCEPT How to implement CONCEPT
advocate CONCEPT Why use CONCEPT
explain CONCEPT What are CONCEPT
compare CONCEPT1 CONCEPT2 How do CONCEPT1 and CONCEPT2 compare?
Towards an Authoring Coach for Adaptive Web-Based Instruction 417
Generation of a hierarchy. In authoring tools like InterBook, MetaLinks, and

NetCoach, authors explicitly provide the textbook hierarchy (like page1 has
subsection page1-1). In contrast, Authoring Coach will generate the hierarchy on the
basis of the concept network and the prerequisites and outcomes associated with
pages (and some additional information, see below). The concept network provides
information about which concepts are 'children' of which other concepts. Note that
there are (at least) two different kinds of parent-child relationships: in AI, "Blind
search" is a special case of "Search Methods" (an aggregation relationship), while
"Criteria for evaluation" is a subtopic within "Search Methods" (a 'uses' relationship).
A concept will be included in the hierarchy if it has either (1) at least two children
that are included in the hierarchy, (2) two pages that use it in their outcomes, or (3)
one page that uses it in its outcome and one child that is included in the hierarchy.
The included children of an included concept and the pages that use it in their
outcomes will be shown as subsections or pages below that concept.
A concept can be a child of multiple other concepts. Additionally, a page can use
multiple concepts in its outcomes. In those cases, the generated hierarchy may contain
multiple instances of a concept (as section title) or page. This will allow the student to
more easily find material, and to follow different paths through the course.
Making a custom book. Books can be generated that are tailored to individual needs.
Authors, teachers and students can choose the outcomes to be covered by a book. The
system will only include pages that are related to achieving those outcomes, either
directly or indirectly (needed for prerequisite).
Dealing with alternatives. Different authors may have different views on how a
specific outcome should be taught. So, different pages can exist with the same
outcomes. We call these pages alternatives. As alternatives have the same outcomes,
they will share the same generated titles (see above). The decision on which page out
of a set of alternatives to include in the textbook can be postponed till the moment
that the student (or artificial teacher) clicks on the title for the first time. Several
strategies can be used. A page can be chosen on the basis of:
• Author: written by the same author as the previous page in the section (if exists),
or by the same author as the last visited page. This may support narrative flow.
Students (or teachers) may also provide an ordered list of their favorite authors.
• Its learning style. This would require the author to annotate learning style
related characteristics of a page. Alternatively, deductions could be made based
upon the page content, like use of images, formulas, words like "example".
• A quality measure. Experts could review (alternative) pages and indicate a
quality rating. Students could also rate pages seen. Alternatively, deductions
could be made based upon the time spend on the page (not very reliable), links
clicks, and the students performance on exercises after having visited the page.
• Its prerequisites being most closely met. Different authors may have different
views on the relative order in which outcomes need to be addressed. This can
lead to pages sharing the same outcomes, but having a different set of
prerequisites. When links are annotated, like in [7], the annotation should be the
most favorable one, i.e., if there is a page with the outcome whose prerequisites
have been met, then the "ready to be learned" indication ought to be given.
The student can select an alternative, via links available from the page.
418 J. Masthoff
Ordering siblings. Pages can have outcomes that use different verbs on the same
concept. For instance, a page with outcome "implement loops", and another page with
outcome "explain loops". We call such pages siblings. When generating the hierarchy,
siblings are ordered in such a way that a page with outcome X will precede all
siblings with prerequisite X. Within this constraint, different strategies can be used to
support narrative flow.
• A page with outcome X is followed immediately (or as soon as possible) by a
page with prerequisite X.
• A page is followed by another page by the same author.
The relative level of verbs in Bloom's taxonomy can also aid the ordering process. For
instance, "explain" precedes "design", which in its turn precedes "evaluate".
4 Conclusions
This paper describes an alternative approach to authoring adaptive web-based

instruction systems. Though this work is only in its starting phases, it is clear that the
choice to generate hierarchies automatically poses interesting problems and
opportunities. If we want web-based instruction systems to grow organically, while
maintaining narrative flow and user guidance then this seems a way forward.
References
1. Masthoff, J. An agent-based interactive instruction system. Ph.D. thesis, Eindhoven

University of technology. (1997)
2. Brusilovsky, P., Eklund, J., and Schwarz, E. Web-based education for all: A tool for
developing adaptive courseware. Computer Networks and ISDN Systems, Proceedings of
Seventh International World Wide Web Conference (1998) 291-300
3. Weber, G., Kuhl, H-C., Weibelzahl, S. Developing adaptive internet based courses with
the authoring system NetCoach. Proceedings of the third workshop on adaptive hypertext
and hypermedia, UM2001, TU/e Computing Science Report 01/11. (2001)
4. Murray, T. MetaLinks: Authoring and affordances for conceptual and narrative flow in
adaptive hyperbooks. Submission to the International Journal of Artificial Intelligence in
Education. As accessed on http://helios.hampshire.edu/~tjmCCS/cv.html December 2001.
5. Bloom, B.S. (Ed.). Taxonomy of educational objectives: Book 1 Cognitive domain.
Longman, London (1979)
6. Reiter, E. and Dale, R. Building natural language generation systems. Cambridge
university press (2000)
7. Brusilovsky, P., Schwarz, E., and Weber, G. ELM-ART: An intelligent tutoring system on
World Wide Web. In C. Frasson, G. Gauthier, & A. Lesgold (Eds.): Third International
Conference on Intelligent Tutoring Systems, ITS-96. Lecture Notes in Computer Science,
Vol. 1086. Springer Verlag, Berlin (1996) 261-269
Generation of Personalized Web Courses Using RBAC
Susana Montero, Ignacio Aedo, and Paloma Dı́az
Laboratorio DEI. Computer Science Department

Universidad Carlos III de Madrid
Avda. de la Universidad 30. 28911 Leganés, Spain
[email protected], [email protected], [email protected]
Abstract. Role-based access control policies model the users domain by means of
complex structures where the roles assumed by the users are specialized into more
concrete subroles which inherit properties and authorizations from their parents. In
this paper, we describe how to combine these principles and adaptation to provide
personalized access to the different types of users of a web-based course. We also
present Courba, a platform that generates personalized web-based courses using
XML to support the definition of access policies.
1 Introduction
In a hypermedia educational application, there are two candidates to be adapted to the

user learning style: the browsing structure (navigation adaptation) and the content de-
livered to the user (presentation adaptation) [1]. In order to adapt the hyperdocument,
course developers have to build a user model gathering relevant information about the
user knowledge, learning style, goals and needs. In predictive statistical models, two
approaches are applied for user modeling: content-based and collaborative- based learn-
ing [6]. Content-based learning uses an individual model based on the user’s behavior
whereas collaborative-based learning uses group based models on the assumption that
users behave in a similar way than other users. In order to specify such group-based
user models, we propose in this paper the use of Role Based Access Control (RBAC)
policies.
2 Role Based Access Control Modeling
RBAC policies [3,4] model users by means of roles which represent functions or re-
sponsibilities. To gather the complexities of most organizations, hierarchies of roles are
defined as a DAG (Directed Acyclic Graph) [3], where general roles are specialized into
more specific roles that inherit attributes and authorizations from their parents. Each
role is granted permission to perform some operations. Finally, each user is assigned a
number of roles, whether static or dynamically, so that during the application operation
each user exercises the abilities specified for the role(s) she belongs to.
RBAC can be applied within the context of educational applications, as far as a
number of roles with different access and manipulation abilities are identified (e.g.
lecturers, students, course coordinator, etc.). Moreover, even though RBAC is oriented
420 S. Montero, I. Aedo, and P. Dı́az
towards modeling security policies, it can be extended to support the adaptive access
to the courseware, translating authorizations rules to access rules that will state the
contents, navigation facilities and interface of each node for each existing role, using a
collaborative-based personalization mechanism as explained in the next section. Such
rules can be explicitly specified or derived for each role, depending on whether the
system is an adapted or adaptive one.
3 A RBAC-Based Hypermedia Model to Specify Personalization
The RBAC model proposed for providing safe access to hypermedia applications in [2]
has been extended in order to provide a rational framework to specify the personalized
access policy of web applications in terms of a number of access rules. Such rules
specify the access category of a subject for a given object and how the object will be
presented (presentation specifications): Access rule = (subject, object, access category,
presentation specifications)
Subjects are active entities that execute actions on the objects (e.g. retrieve a node,
modify a link or create a new content). Two kinds of abstractions are identified as subjects:
roles and teams. While a role is an organizational position or job function, a team is just
a group of users established whether to represent collaborative group as in [5] or to
simplify the administration tasks. Roles and teams support composition mechanisms to
be able to deal with complex user structures.
Objects are passive entities that receive the effects of the actions executed by the
subjects. Nodes and contents are considered as objects, and links which always resolve
into a node or content, inherit the rules that apply to the elements making up their
definition.
Access categories embody the access capabilities for a given hyperdocument or
course. Four categories are considered: No access, Browsing, Personalizing and Editing.
4 Courba: A Platform
to Generate Personalized Web-Based Courses
Courba is a platform devoted to create web-based courses that will be accessed following
a RBAC policy. The current version does only deal with personalized courses where
access rules are hard coded for the roles so that personalization of the course structure,
contents and interface is supported.
4.1 Courba Architecture and Implementation
Courba is made up of a series of modules where structure, content and presentation

features have been separated to make possible the personalization to different users
roles in a web-based course (see Fig. 1).
The Information Base holds information on the course as well as on the users. First,
there is a course DTD which defines a logical structure where a course has a number of
roles, access rules and a set of pages. Each page can be a cover page, index, theory page
Generation of Personalized Web Courses Using RBAC 421
Fig. 1. Courba Architecture
or exercise. In turn, a index page has some contents and some operations which can be
performed on that page (e.g. go to the parent page). The course contents are held in an
XML document and the information about the presentation features (color, style and so
on) is contained in separated style sheets. Finally, the roles as well as the assignment
of their users are maintained in a separated file. When a user has to be included or
removed from a course, only this latter users file is modified while the course DTD,
XML document and style sheets remain unchanged.
The Application Manager is responsible for answering to the user’s requests. Their
tasks are to divide the XML document of the course into a series of pages (or nodes)
taking into account the structure of the course represented in the DTD and to execute
the manipulation operations that will be supported by the course. Current version of
Courba supports the navigation operations where are established the target of each link
according to the role of the user who is accessing the node. Access rules determine
which operations will be available for each role. Finally, this module creates the web
page corresponding to a users request taking into account the information about the page
(structure, contents, interface and operations) as well as the user’s role.
4.2 Personalized Access in a Courba Courseware
In order to create a personalized hypermedia courseware, developers have to create the

XML document containing the course components, to specify which roles will be used
in their course and which users hold that roles. Finally, they have to establish the access
rules for the course components that will be nodes (that is, pages) and contents (that is,
each information item - text, image, plug-in or operation-) included in a page.
422 S. Montero, I. Aedo, and P. Dı́az
The Courba platform has been tested in a course on Hypermedia Design. In the XML
document has been included an index of topics, two different roles (the novice and the
expert learner) and a set of access rules are specified to let build a dynamic personalized
index. Figure 2 shows output created by Courba. In Fig. 2.A, it can be seen that a novice
learner in this subject will have access to four topics. Once a learner has some experience
(see Fig. 2.B), she is also given the opportunity of accessing sections about more specific
problems and basics concepts are hidden. The targets of links Previous (“Anterior” in
the figures), Parent (“Padre” in the figures) and Next (“Siguiente” in the figure) are
calculated in runtime by the Courba Operational Specification module depending on the
user’s role.
Fig. 2. Table of contents of the subject “User Interface” for a novice and an expert learner
Generation of Personalized Web Courses Using RBAC 423
5 Conclusions
In this paper, we have discussed how to combine the basic principles of adaptation for
hypermedia applications with RBAC policies in order to support personalized access
to a web-based courseware. With this purpose, a platform to support personalization in
web courses, called Courba, based on a modular architecture has been proposed.
Acknowledgements
We’d like to thank Jose Ángel Cruz for his cooperation in the development of Courba.
The RBAC model used in Courba is part of a project funded by “Dirección Gen-
eral de Investigación de la Comunidad Autónoma de Madrid y Fondo Social Europeo
(07T/0012/2001)”.
References
1. Brusilovsky, P. Adaptive hypermedia. User Modeling and User-Adapted Interaction, 11:87–
110, 2001.
2. Dı́az, P., Aedo, I. and Panetsos, F. Modelling security policies in hypermedia and web-based
applications. In Web Engineering: Managing diversity and complexity of web application
development, volume 2016 of Lecture Notes in Computer Science, pages 90–104. Murugesan,
S. and Deshpande, Y. Eds. Springer, 2001.
3. Ferraiolo, D., Barkley, J. and Kuhn, R. A role-based access control model and reference
implementation within a corporate intranet. ACM Transactions on Information and System
Security, 2(1):34–64, 1999.
4. Shandu , R.S., Coyne, E.J., Feinstein, H.L. and Youman, C.E. Role-based access control:
A multi-dimensional view. In Proc. of the Tenth Annual Computer Security Applications
Conference, pages 54–62, Orlando, Florida, 1994.
5. Wang, W. Team-and-role-based organizational context and access control for cooperative
hypermedia environments. In Proc. of Hypertext’99, pages 37–46, Darmstad.
6. Zukerman, I. and Albrecht, D.W. Predictive statistical models for user modeling. User
Modeling and User-Adapted Interaction, 11:5–18, 2001.
An Automatic Rating Technique
Based on XML Document
Hyeonjeong Mun1 , Sooho Ok2 , and Yongtae Woo1

1
Changwon National University, Changwon Kyungnam 641-773, Republic of Korea,
{mun,ytwoo}@sarim.changwon.ac.kr
2
Kosin University, Yeongdo-gu Pusan 606-701, Republic of Korea,
[email protected]
Abstract. In this paper, we propose an automatic rating technique to

collect the interest of users on the contents, such as e-books and elec-
tronic catalogs that are composed of XML documents, for developing a
personalized recommender system. Our approach focuses on a method
to collect implicit rating values for the elements in an XML document
accessed by a user, when the content is converted into HTML format.
In general, existing implicit rating techniques collect rating values after
analyzing access logs in batch mode, however our method can collect the
rating values in realtime. As a result of experimentation, we show that
the implicit rating values collected by the proposed method are strongly
correlated with explicit rating values.
1 Introduction
The main purpose of a recommender system is to provide personalized sug-
gestions about the items in an E-commerce system. Therefore, a recommender
system requires an intelligent user interface that can determine the interest of
a user. The common solution is to use explicit ratings, where users tell what
they think about a piece of information to the system, is well understood and
fairly precise [1]. Even though the explicit rating technique can collect the exact
interest about items from users, it can make troublesome overloads for them
and deteriorate the efficiency of the recommendation if the rating information is
insufficient or rating features had been changed.
Several implicit rating techniques have been studied to collect rating values
automatically, where a rating is obtained by a method other than obtaining
it directly from the user. Claypool has studied the correlation between various
implicit ratings and the explicit rating for a single item, and the impact of
implicit interest indicators on user privacy [2]. Morita has proposed an implicit
rating technique to accumulate a user’s preference for information based on user
behavior monitoring by measuring the time spent for each article [3]. Nichols
has sought to construct a comprehensive view of implicit feedback, with a focus
on its use in information filtering systems. He has presented a list of potentially
observable behaviors; adding purchase, assess, repeated use, refer, mark, glimpse,
associate, and query to those mentioned above [4]. Oard and Kim have built upon

An Automatic Rating Technique Based on XML Document 425
work by Nichols by categorizing implicit ratings, dividing them into the following
categories: Examination, where a user studies an item; Retention, where a user
saves an item for later use; and Reference, where a user links all or part of an
item into another item [5].
Recently, XML(eXtensible Markup Language), a data format for structured
document interchange is promoted actively in various application areas such as
e-books and electronic catalogs [6]. In this paper, we propose a new technique to
collect users’ interest automatically for a content based on the hierarchical struc-
ture among the elements in an XML document for personalized recommender
system.
2 Implicit Ratings Using XML Hierarchical Structure

XML is way of marking up data, adding metadata, and separating structure from
formatting and style. An XML document is composed of a hierarchical set of
elements depending on the root element. When users access an XML document,
which is converted into HTML using styles defined in the XSL script, instead of
accessing XML documents directly.
We suggest a new method to collect rating values of the elements implicitly
in an XML document accessed by a user when the document is converted into
HTML document. Figure 1 shows a conceptual diagram of the proposed system.
XML Docu m ent HTML

XML Parser Docum ent
Im plicit Rating Engine
User Rating
XML Generator Rating Process Record
Request
Rating
Evaluation
XML Elem ent

DTD Content Rating
Rating
Database Database Database
Fig. 1. The proposed implicit rating system based on XML document
Initially, all elements in an XML document can be assigned predefined rating

value in accordance with their priority. For example, the elements related to
purchasing intention such as price or ordering products are given higher value
than such elements as product description or an image in an E-commerce system.
Whenever a user accesses the XML document, The Implicit Rating Engine is
activated to evaluate rating values related to that content. The Rating Process
retrieves rating values of those elements in the Element Rating Database, and
426 H. Mun, S. Ok, and Y. Woo
generates rating records for those elements. Afterwards, the Rating Evaluation
calculates a rating value for the content, and saves the value into the Rating
Database.
The total rating value for an XML document can be defined as an accumu-
lated value multiplied level weight by rating value of each element accessed by
the user. A level weight is assigned to each element in accordance with its level in
the XML hierarchy. In our approach, we have defined a formula, Rc to evaluate
total rating value for an XML document as follow:

Rc = wi · ri ,
i∈V
where V is a set of elements of an XML document accessed by a user, wi is a

level weight, and ri is a rating value assigned to each element. Therefore, Rc
means that the total rating value for an XML document and a degree of interest
on the specific content accessed by a user.
We have performed an experiment with a total of 2,030 users from a recruit-
ing site to verify the effectiveness of our approach. In the experiment, we have
collected explicit rating grades from 1 through 5, implicit ratings with the pro-
posed method, and reading time from the access logs on 326 XML documents.
Correlations between the explicit ratings and the reading time, as well as the
explicit ratings and the implicit ratings, have been calculated. Figure 2 shows
the results of comparison with each method.
ͣ͢͡͡ ͦ͡
͢͡͡͡ ͥ͡
Im plicit Rating (Rc)
Reading Tim e (sec)
ͩ͡͡
ͤ͡
ͧ͡͡
ͣ͡
ͥ͡͡
ͣ͡͡ ͢͡
͡ ͡
͡ ͢ ͣ ͤ ͥ ͦ ͧ ͡ ͢ ͣ ͤ ͥ ͦ ͧ
Explicit Rating(grade) Explicit Rating(grade)
(a)Explicit Rating Vs. Reading Time (b)Explicit Rating Vs. Proposed Method
Fig. 2. The results of experiment
Figure 2(a) shows the results of correlation between the explicit rating values
and the reading time for the same set of contents. As the result of experiment, a
correlation coefficient between the explicit rating values and the reading time was
An Automatic Rating Technique Based on XML Document 427
0.41. The result shows weak correlation due to the difference in understanding
the content among the users or network speed.
Figure 2(b) shows the results of correlation between the explicit rating val-
ues and the implicit ones collected with the proposed method. Correlation co-
efficients show a range from 0.75 to 0.83 depending on various rating values
and weights. The implicit rating values collected with the proposed method,
therefore, have shown strong correlation with explicit ones.
4 Conclusion
In this paper, we have proposed a new method to collect users’ interests auto-
matically for content in order to develop a personalized recommender system.
Our approach focuses on collecting implicit rating values of the elements in an
XML document accessed by a user. Our method can collect implicit rating val-
ues in realtime mode, while the existing implicit rating techniques collect rating
values after analyzing access logs in batch mode. The result of our experiment
shows that the implicit rating values collected with the proposed method has
more strongly correlated with the explicit rating values than the correlations
between the explicit ratings and the reading time. The proposed method can be
thus expected to contribute more to the development of an efficient personalized
recommender system as it collects users’ interests automatically to substitute
explicit ratings.
Acknowledgement
This work was supported by grant No. 2001-111-2 from the University funda-
mental research program of the Ministry of Information & Communication in
Republic of Korea.
References
1. Konstant, J., Miller, B., Maltz, D., Herlocker, J., Gordong, L., Riedl, J.: Grou-
pLens: Applying Collaborative Filtering to Usenet News. Communications of the
ACM., 40(3) (1997) 77–87
2. Claypool, M., Brown, D., Le, P., Waseda, M.: Inferring User Interest. Technical
Report WPI-CS-TR-01-97 (2001)
3. Morita, M., Shinoda, Y.: Information Filtering Based on User Behavior Analysis
and Best Match Text Retrieval. Proceedings of the 17th Annual International
ACM-SIGIR Conference on Research and Development in Information Retrieval
(1994) 272–281
4. Nichols, D.: Implicit Rating and Filtering. Proceedings of the Fifth DELOS Work-
shop on Filtering and Collaborative Filtering (1997) 10–12
5. Oard, D., Kim, J.: Implicit Feedback for Recommender Systems. Proceedings of
the AAAI Workshop on Recommender Systems (1998) 80–82
6. Abiteboul, S.: On views and XML. Proceedings of ACM Symposium on Principles
of Database Systems (1999) 1–9
Using Effective Reading Speed
to Integrate Adaptivity into Web-Based Learning
Muan Hong Ng1, Wendy Hall1, Pat Maier2, and Ray Armstrong3
1Intelligence,Agents and Multimedia Research, University of Southampton,
SO17 1BJ Southampton, United Kingdom
{mhn99r,wh}@ecs.soton.ac.uk
2 Center of Learning and Teaching, University of Southampton,
SO17 1BJ Southampton, United Kingdom

[email protected]
3 Rheumatology Unit, Mail Point 78, Southampton General Hospital,
SO16 6YD, Southampton, United Kingdom

[email protected]
Abstract. It has always been difficult to determine to what extent a user has
read a page especially in the area of educational adaptive hypermedia systems.
We propose the use of an individual’s effective reading speed to estimate how
much of a page a user has read during their browsing activity. This method is
currently used to apply history-based link annotation in a medical web-based
learning application, JointZone. A validation test of this work has shown a
positive result in approximating user’s reading value when compared to
conventional methods.
1 Introduction
A difficult problem faced by many adaptive hypermedia developers is to identify

whether a page has been read. A generally unreliable solution is to use the viewing of
a page as an indicator that the page has been read and understood (e.g. InterBook [2]).
This method is obviously inefficient as the student might only be surfing through the
pages without paying much attention to the content. In MANIC [11], the system
assumes a high studied rating for a page if a user spends an optimal amount of time on
it. This optimal time that is generated, based on the length of the content, is static for
every user irrespective of their individual reading speed. However, Psychological
research has shown that there are “astonishing differences in the rate of reading
speed” among individuals [10]. It has also been found that two individuals reading at
the same speed can have different rates of comprehension [5]. Motivation for this
research comes from the users’ unpredictable browsing behavior on the web,
particularly the speed used in reading web pages. Research has shown that users
frequently skim read the web pages instead of reading them in detail [4]. Studies
have also shown that there is a general decline in the level of comprehension
(especially in the recall of specific details) with faster reading as compared to reading
at a normal speed [3, 7, 8]. Hence the reading speed, the time spent and the level of
understanding on each page varies among individuals. The work in this paper captures

Using Effective Reading Speed to Integrate Adaptivity into Web-Based Learning 429
this aspect of users’ differences in reading speed and uses this to calculate how well
they have read a web page based on the display time of the page.
2 Calculating the Effort Index

JointZone [13] is an interactive website for Rheumatology for both undergraduate
medical students and practicing doctors. It contains approximately 80 documents, 750
images and 30 interactive case studies. The user model in JointZone contains an effort
index (0-100%) for each page in the domain. This index is the system’s estimation of
the ‘degree of effort spent’ by a user on each page. The effort index is actually
calculated by comparing the display time of a web page with an individual’s optimal
reading time. The optimal reading time for a page is identified for each user
depending on the individual’s reading rate, comprehension rate and prior knowledge
of the domain. These user’s characteristics are obtained when they use the system for
the first time. Prior their first entry, a user will be asked to complete a knowledge test
(10 multiple choice questions based on the general concept of the domain) to give a
measurement of their prior knowledge (0-100%). Then each user has to take a
‘reading speed test’ where their reading rate is captured in number of words read per
minute. The reading time is recorded through a start and stop button on the web page.
This is followed by a comprehension test (six true or false questions on general ideas
of the text) to give a score of their comprehension rate (0-100%). The combined
measure of reading rate and comprehension gives the index of effective reading speed
[6]. This index is commonly used in the commercial product for speed-reading
courses [9, 12]. We have however, modified the equation by adding the factor of prior
knowledge (see Equation 1).
1
Effective reading speed = reading rate × ( comprehension rate + prior knowledge )(1)
2
total number of words in the page
Optimal reading time for a page = (2)
Effective reading speed
− (x − optimalTim e )
2
Effort Index Estimation, G ( x) = e 2σ 2 , where σ = 1.23 (3)
During a browsing session the optimal reading time for any page in the domain is
calculated in real time using Equation 2. The effort index on each page is then
estimated using a Gaussian function (see Equation 3) by comparing the actual time
spent on a page (x) with the optimal reading time. If ‘x’ approaches the optimal
reading time, a high effort index is assumed. However, one problem still exists, as it is
difficult to determine if a user has indeed read a page when it is displayed on the
screen. To tackle this, a heuristic cutoff point is used to give a zero effort index to
cases where display time falls below eight second or three times the optimal reading
time. On the other hand, the use of individual optimal reading time enable us to skew
the effort index based on a user’s prior knowledge since a good student who skims a
page will gain a higher effort index (with lower optimal reading time) as compared to
a poor student who spent the same amount of time on the page.
430 M. H. Ng et al.
3 Validation Test
A validation test was carried out to study the correlation between the effort index and
the users’ understanding of the content itself. Thirty subjects took part in the
evaluation. Prior to the experiment, all users are instructed to complete a series of
reading tests to capture their individual effective reading speed. The analysis for thirty
users shows a mean effective reading speed of 116 word/min (sd = 58.74), mean prior
knowledge of 30.67% (sd = 15.52) and mean comprehension rate of 63% (sd =
15.93). In the experiment, they were asked to read a page (length = 443 words) in the
domain followed by a performance test to examine how much they understand the
page. This test contains five multiple-choice questions, which asked users to recall
some important aspects in the text (main factual type) [3]. All users spent an average
of 3.19 minutes reading the page (sd = 1.12 minutes). The effort index for each user
was estimated by the system using Equation 3. The score of the performance test was
then compared with the effort index. The Pearson correlation test showed a significant
correlation between the effort index and the performance test (r = 0.521, p = 0.003).
The performance score was also compared with a separate calculation of effort index
using a standard optimal reading time for all users (2.25 minute using the average on-
screen reading rate of 200wpm). The Pearson correlation test shows a non-significant
correlation of 0.056 (p = 0.767). Hence, we can conclude that each user has a different
optimal reading time, which gives a relatively better approximation of the user’s
understanding of the domain.
Fig. 1. An example of the use of effort index in the history-based link annotation in a
personalized site map
4 History-Based Link Annotation
The effort index is used as a basis to form the history-based link annotation [1] in the
JointZone web-based medical learning. As shown in Fig. 1, this adaptive feature is
applied on a personalised site map to provide feedback to the user on the system’s
Using Effective Reading Speed to Integrate Adaptivity into Web-Based Learning 431
assumption of his/her reading value for each page. This site map acts as a navigational
support mechanism for the users and helps them to quickly gain an overview of the
domain on pages they have read or not read.

The individual effective reading speed used in calculating the effort index on web
pages gives us a relatively good approximation in judging how well a page has been
read. However, the cognitive issues in measuring the user’s understanding have yet to
be addressed in this work. Hence we cannot claim that the effort index completely
represents a user’s understanding of the text on the page. Nevertheless, the validation
test showed that our method gives a comparatively more accurate approximation than
existing methods. With this encouraging result, we will continue to expand the use of
history-based link annotation with the effort index. We will also undertake a set of
evaluations to specifically analyze this adaptive procedure within JointZone, to
discover if this feature aids the user.
Acknowledgements
Arthritis Research Campaign, UK, educational project grant A0549.
References
1. Brusilovsky, P.: Adaptive Hypermedia. Journal on User Modeling and User Adapted
Interaction, 11 (2001) 87-110
2. Brusilovsky, P., Eklund, J., and Schwarz, E.: Web-based Education For All: A Tool for
Developing Adaptive Courseware. In: Proceedings of Seventh International World Wide
Web Conference, 14-18 April, 30 (1-7) (1998) 291-300
3. Dyson, M. and Haselgrove, M.: The Effects of Reading Speed and Reading Patterns on
Our Understanding of Text Read From Screen. Journal of Research in Reading 23 (2000)
210-223
4. Horton, W., Taylor, L., Ignacio, A and Hoft, N. L.: The Web page Design Cookbook. John
Wiley, New York (1996)
5. Huey, E. B.: The Psychology and Pedagogy of Reading. The Macmillan Company (1908)
6. Jackson, M. D. and McClelland, J.L.: Processing Determinants of Reading Speed. Journal
of Experimental Psychology: General 108 (1979) 151-181
7. McConkie, G., Rayner, K. and Wilson, S.: Experimental Manipulation of Reading
Strategies. Journal of Educational Psychology 65 (1973) 1-8
8. Poulton, E. C.: Time for Reading and Memory. The British Journal of Psychology 49
(1958) 230-245
9. ReadingSoft.com found at http://www.readingsoft.com
10. Romanes, G. J.: Mental Evolution in Animals. Appleton & Co, New York (1885)
11. Stern, M. and Woolf, B. P.: Adaptive Content in an Online Lecture System. In:
Proceedings of the International Conference on Adaptive Hypermedia and Adaptive Web-
Based Systems, Trento Italy (2000)
12. TurboRead Speed Reading found at http://www.turboread.com
13. JointZone – A Study for Rheumatology found at
http://www.iam.ecs.soton.ac.uk/users/mhn99r/learn/userlogin/index.jsp
A Non-invasive Cooperative Student Model
Teresa Roselli, Enrica Gentile, Paola Plantamura, Veronica Rossano,

and Vittorio Saponaro
Dipartimento di Informatica, Università degli Studi di Bari

Via Orabona, 4 - 70126 Bari, Italy
[email protected]
Abstract. A crucial issue in building a user-centered Intelligent Educational

System is that of building and maintaining user models to enable the system to
tailor its behavior to the needs of the given individual. Various student
modeling techniques are described in the literature (overlay, perturbation,
differential), that represent the student’s knowledge in different ways. A
common problem affecting these techniques is how to obtain the data needed
to build the student model. Open models are one of the possible solutions but
these have some drawbacks. In the present article we propose a new open
model aiming to solve some of these, by minimizing the invasiveness of the
requirement for student involvement in the self-assessment process, together
with the sense of frustration such a requirement can arouse in students who are
unable to assess themselves and are therefore ignored by the system.
1 Introduction
One of the essential aims of teaching-learning processes is to enable students to
acquire the ability and knowledge needed to carry out successful problem solving in
the relative domain. The method adopted by Intelligent Educational Systems is to
induce the student to build a domain model, which is then compared with the system
knowledge base [1]. Various student modeling techniques have been reported in the
literature. These aim primarily to construct cognitive models of the learner’s domain
knowledge and can also include some meta-cognitive aspects [2], going beyond
simple user modeling to build explanations of students’ behavior not only by
observing their actions but also by engaging them in various learning situations.
In the overlay model proposed by Carr and Goldstein [3], the student’s knowledge
is treated as a subset of the system knowledge base and the model is built by
comparing the expert’s and the student’s responses. The system starts from the
assumption that all differences between the two are to be attributed to student
deficiencies in one or more abilities. The overlay model thus has the drawback that is
not equipped to deal with any student knowledge or opinions differing from those of
the expert. If the student adopts a problem-solving strategy that has not been
included in the system, the tutor will intervene to correct the “error”.
An alternative method is differential modeling, which compares the student’s
behavior, rather than his/her knowledge, with that of the expert in the same situation.
However, the student’s knowledge is still assumed to be a subset of the expert
knowledge, so that the model is again incomplete.

A Non-invasive Cooperative Student Model 433
A third method, the perturbation method thought up by Brown and Burton in 1978
[4], provides a similar student and expert model, in which differences between the
student’s and the system’s behavior are represented as perturbations.
In any case, the problem arises of how to acquire the data needed to build the
student model. A promising solution is offered by open student models. These
require the student to take an active part in the construction of the model, with the
aim of increasing the data obtained and hence the accuracy of the model. The fact of
giving students the opportunity to inspect and discuss their knowledge makes it
possible to cover wider cognitive aspects and make better forecasts of their future
performance, because the process of discussion of the model gives them the chance
to reflect on their understanding and experience of the domain, to revise and
reconsider their ideas and thus become more aware. This reflection stimulates a
better perception of the relationships and connections between different knowledge
parts and encourages active participation in the learning process [5,6].
However, one of the side effects of student involvement in the assessment process
is that students who are unable to make a self-assessment are ignored by the system.
The resulting sense of frustration could affect their performance, although this does
not, in fact, seem to occur, according to the experiments conducted by Beck [7].
However, the requirement for student involvement in management of their model
may be too invasive if too many opinions on their knowledge level are called for.
Many negative effects could result, such as distraction (while reflecting on acquired
knowledge, the new notions proposed could be neglected) and a feeling of insecurity
that could hinder the overall learning process.
The aim of the present work was to build a cooperative user-system student
model that could solve some of the problems reported in the literature with such
models. As regards the difficulty of the requirement for student involvement in the
assessment process being too invasive, causing annoyance and distraction, the
solution proposed is to merge the tutor’s communication of the expert assessment
with a request for optional self-assessment by the student. Another feature aims to
minimize students’ frustration if they are unable to carry out self-assessment and are
therefore ignored by the system. Use is made of graduated negative feedback
according to the student’s estimated self-assessment ability.
2 The Weighted Cooperative Student Model

The student’s calculation of his/her self-assessment ability is merged with the tutor’s
assessment, to obtain a combined (tutor-student) view. The differences are attributed
variable weights according to the expert estimate of the student’s self-assessment
ability. Another feature of the model is the fact that students are not obliged to make
a self-assessment. This prevents excessive invasiveness of the tutor, that might
systematically distract and condition the student by obliging him/her to make
continual judgments, is avoided [7].
After doing an exercise, the results are communicated and the option of making a
self-assessment is given. If the pupil does not wish to do so, or agrees with the
tutor’s assessment, he/she can go straight on to the next phase. Instead, in the case of
a different opinion, the process for conflict solving is activated and leads to an
integrated assessment that merges the student’s and tutor’s assessments, by
434 T. Roselli et al.
calculating the weighted mean of the two. The weights are calculated according to a
confidence function. The whole process is illustrated in Figure 1.
Fig. 1. Combined student-tutor assessment
Calculation of the combined assessment. The method for calculating the combined
assessment is based on a series of confidence coefficients (Fi) that attribute the
appropriate weights to the student’s (Vstudent) and tutor’s (Vtutor) opinions.
The starting confidence coefficient is F1 = 0.5, while the combined assessment
(Vfinal) is calculated by:
Vfinal = Fi⋅Vstudent + (1-Fi)⋅Vtutor (1)
There are three main factors involved in the calculation of the new confidence
coefficient: the previous coefficient, the student’s mean self-assessment ability and
the mean spread ratio between this and the tutor’s assessment.
Self Assessment Mean Capability Ratio. The sum of the equation below is defined
as the index of the student’s mean self-assessment ability:
SAMCR = mean(Vstudent – Vfinal)i (2)
th
where i = (1...N) indicates the i self-assessment made.
The domain of the SAMCR is [-2, 2]. At each step, this index represents the mean
self-assessment error.
Mean Spread Ratio of Student Errors. The mean spread ratio of student errors is
the sum:
MSRSE = mean |(Vstudent – Vfinal)i - SAMCR|i (3)
th
where i = (1...N) indicates the i self-assessment made.
This index gives an overall idea of how the student’s errors are spread over the
domain, i.e. how consistently the student correctly/incorrectly estimates, or
over/underestimates his/her performance.
Confidence function. The confidence function (f) enables calculation of the

confidence coefficients on the basis of the current SAMCR and MSRSE indexes, that
keep track of the student’s self-assessment ability, together with the current
th
confidence coefficient Fi. At the i step, the next confidence coefficient is calculated
as follows:
Fi+1 = f(SAMCRi , MSRSEi) ⋅Fi = ki⋅Fi (4)
A Non-invasive Cooperative Student Model 435
The lower the student’s self-assessment ability, the lower the ki value and the
lower the tutor’s confidence in the student. In addition, the student will receive
suggestions aiming to improve his/her self-assessment ability.
If the combined assessment is positive the student receives positive feedback; if
negative, the theoretical lesson for the ability in question is re-proposed and a new,
remedial, exercise is set to improve his/her ability. Further self-assessment is avoided
to prevent excessive invasiveness of the system.
The student model described has been implemented in an intelligent hypermedial

system, ‘Frazionando’, that targets primary school pupils. The knowledge domain
includes the study of fractions and their main properties. A pilot test was carried out
to verify the efficacy of the confidence function and confidence coefficients used by
the system and, in general, the reliability of the collaborative student model. Twelve
primary school students were selected on the basis of heterogeneous scholastic
performance and level of maturity, as assessed by their teachers. Among the positive
results obtained, the approach was judged useful and was particularly appreciated by
typically introvert students who have greater difficulty in communicating learning
problems to a human tutor. Another satisfactory result was that of reducing the sense
of frustration the system could arouse in students with less ability to make self-
assessment: thanks to the advice the system gave students about how to make their
self-assessments, they felt involved as participants in the overall assessment process.
For future developments of the model, further experiments on a larger student
sample must be made. We also aim to integrate the domain knowledge with data on
the typical errors made by students and to set up strategies for increasing the powers
of adaptation of the hypermedial system [8].
References
1. Roselli, T. : Artificial Intelligence can improve Hypermedial Instructional Technologies for
Learning. ACM Computing Surveys, Vol. 27, No 4 (1985)
2. Paiva, A., Self, J., Hartley, R.: Externalising learner models. Proceedings of World
Conference on Artificial Intelligence in Education. Washington DC. (1995) 509-516
3. Carr, B., Goldstein, I.: Overlays: a theory of modelling for computer aided instruction. (AI
MEMO 406) Cambridge. MA: M.I.T.. AI Laboratory (1977)
4. Brown, S., Burton, R.B.: Diagnostic Models for procedural bugs in basic mathematical
skills. Cognitive Science, 2 (1978)
5. Boud, D., Keogh, R., Walker, D.: What is reflection in learning. In Boud. Keogh.Walker
eds. Reflection: turning experience into learning Kogan Page. London. (1996) 7-17
6. Dimitrova, V., Self, J., Brna, P.: Involving the learner in diagnosis – Potentials and
problems. Computer Based Learning Unit, Leeds University, Leeds LS29JT, UK. (2000)
7. Beck, J., Stern, M., Woolf, B.P.: Cooperative Student Models. Computer Science
Department, University of Massachusetts. Proceedings of the AIED. (1997) 127-134
8. Brusilovsky, P.: Adaptive Hypermedia. User Modeling and User-Adapted Interaction 11.
(2001) 87-110
Category-Based Filtering in Recommender Systems
for Improved Performance in Dynamic Domains
Mikael Sollenborn and Peter Funk
Mälardalen University
Department of Computer Science and Engineering
Västerås, Sweden
{mikael.sollenborn,peter.funk}@idt.mdh.se
Abstract. In Recommender systems, collaborative filtering is the most

commonly used technique. Although often successful, collaborative filtering
encounters the latency problem in domains where items are frequently added,
as the users have to review new items before they can be recommended. In this
paper a novel approach to reduce the latency problem is proposed, based on
category-based filtering and user stereotypes.
1 Introduction
For personalization of web pages, Recommender systems are currently the most
common approach. Based on the information filtering technique known as
collaborative filtering [1,2], standard Recommender systems essentially function on a
peer review basis. When making recommendations, users with similar preferences are
identified, and their item ratings are used to propose items to one another.
In addition to collaborative filtering, personalized selections based on matching the
user’s previous selections with meta-data or content keywords for individual items -
known as content-based filtering - is also very common.
Traditional Recommender systems often encounter the latency problem [3], i.e.
new items cannot be used in collaborative recommendations before a substantial
amount of users have evaluated it, as the recommendations rely on other users
opinions. This problem is especially apparent in domains where new items are often
added and old items quickly get out of date. Content-based filtering may be a
solution, but runs the risk of only recommending items almost identical to the ones
the user has appreciated before [3]. As noted in [4], the most obvious solution to the
latency problem is to categorize the items in the system. In this paper we go one step
further and propose that for quickly changing domains, Recommender systems solely
based on categories may provide sufficient personalization.
2 Category-Based Filtering Approach

We refer to the personalization approach proposed in this paper as category-based
filtering. Its main characteristic is that selection of information is based on category

Category-Based Filtering in Recommender Systems 437
ratings instead of item ratings. To function, category-based filtering requires

categorization of every item, either manually or by an automated process.
The selection of items is based partly on individual user models, and partly on
collective user stereotypes. A user model represents the current knowledge about a
user’s reaction towards shown categories of items. The model consists of a preference
matrix containing the number of times the user has seen items belonging to each
category, and the frequency with which these items were clicked. A user stereotype,
in contrast, consists of collective category preference information for a group of
users, acquired through clustering. When the information in a user model is
insufficient for deciding which items to select, the user stereotype assigned to the user
is consulted to make assumptions about the user’s expected behavior.
The focus on categories reduces the latency problem, as new items can be
recommended as soon as the user’s attitude towards the item’s category is known.
The method is thus especially suited for domains with a constant flow of new
information (e.g. news, adverts), provided that efficient classification is possible.
2.1 Creation of User Stereotypes
User stereotypes are identified using offline hierarchical clustering of system users.
The variables determining cluster membership are as many as there are categories in
the system. Different values are assigned to the category attitudes negative, neutral,
and positive, and the Manhattan distance metric is used to measure distance between
clusters. Similar clusters are merged using the unweighted pair-group average
method.
The chosen method of capturing collective interests in user stereotypes is to utilize
what will be referred to as appreciation nets. Appreciation nets are graphs with nodes
and directed edges, where edges represent a probabilistic relationship. In Fig. 1 an
example of an appreciation net is given for four item categories. Here, the likelihood
that a person who likes hunting is also interested in motor sports is 60%, and 30% in
the opposite direction. Of the persons belonging to this population, 50% enjoy motor
sports, but only 20% appreciates hunting, as indicated in the category nodes.
0 .6 0 .2
M o to r
0 .3 S p o rts 0 .5 0 .9
H u n t in g 0 .2 0 .4 H e a lt h y
0 .2 0 .1 liv in g 0 . 1
0 .7
0 .4 0 .5
F is h in g
0 .7 0 .4
0 .4
Fig. 1. An appreciation net with four item categories
A user stereotype consists of an appreciation net, with all nodes connected to every
other node in both directions. When forming such a net, a joint distribution is made
from the ranked category preferences of every user belonging to the cluster. For each
category C the system stores the probability of a positive evaluation by any user
belonging to the group, as node values in the appreciation net. Secondly, the
438 M. Sollenborn and P. Funk
probability of a user appreciating C in case the user likes D, P(C|D), is preserved for
each category-to-category connection, stored as binary relationships between
category nodes in the appreciation net.
2.2 Classification of New Users
A new user must be assigned a user stereotype before personalized items can be
selected. Initially, the system doesn’t know enough about the user to decide which
stereotype is the most appropriate. Automatic classification is attempted by targeting
the user with information corresponding to differential probability values in the user
stereotypes appreciation nets. The goal is to determine which user stereotype
resembles the new user the most. Initially, it is assumed that the biggest cluster is an
appropriate starting point.
Category nodes in the appreciation net with high appreciation probabilities are
sought for, with these values being as unique as possible compared to the equivalent
category values in other clusters. The appropriateness of being chosen, F, is
calculated for every category node C by comparing it to the corresponding category
node Ci in every other user stereotype appreciation net, using
F = P(C ) ∗ ∑i
P(C ) − P(Ci ) . (1)
The categories with highest F are chosen for stereotype identification, i.e. a number
of items belonging to these categories will be shown to the user. After the user has
responded, a stereotype determination is performed. The categories involved in the
test are compared to the appreciation net in each user stereotype, putting emphasis on
similar categories with high probability values. The dissimilarity D is calculated for
every user stereotype, where Cui is category i in user model u, Csi is the
corresponding category i in user stereotype s, and M is an empirically chosen
modification rate (about 1.25), using
∑
D = P(Cui) − P(Csi) ∗ (1 − P(Csi) ÷ M ) .
i
(2)
The user stereotype that most closely resembles the initial behavior of the user is now
chosen for a second pick of categories using (1). The process continues until the
collected information indicates that the user can be classified safely.
2.3 Selection of Personalized Items
Personalized information is divided into appreciation-known and appreciation-

assumed. A new user is confronted with a lot of appreciation-assumed information,
but as the user model stabilizes, appreciation-known information gradually replaces
it.
Appreciation-known items are chosen only from categories the user has evaluated
enough times to be reasonably sure about his/her attitude towards it. Information is
Category-Based Filtering in Recommender Systems 439
selected by ranking the user preferences, picking items from categories that have been
positively evaluated. In this process, the system tries to balance the number of shown
items among the positive preferences, as well as sometimes picking sufficiently
evaluated categories with a less positive ranking to allow for re-evaluation.
When appreciation-assumed items are to be selected, the system chooses a
category node starting point in the appreciation net among the users positively ranked
preferences, and examines all connected category nodes. The category to select
information from is chosen randomly from a dynamically generated pie chart, where
each category not among the user’s positive preferences gets a slice size (choice
probability) calculated using equation 3. W is the connection weight, C the number of
selections of items belonging to this category, S the number of times shown to the
user, and L how many of the last H items in the category that has been selected by the
user. H is domain dependent; in our test evaluation the history length is ten items.
P = W * ((C + 1)/(S + 1) + (L+1)/H) (3)
Another form of appreciation-assumed selection, used in parallel with the method
above, works similarly to the automatic classification process by picking items from
categories in the appreciation net where the probability of a positive response is high.
The items selected by using each of these techniques are finally merged, and
presented to the current user.
3 Conclusions
In this paper we have presented an approach to Recommender systems for application

domains where items are frequently added. Provided that sufficient categorization is
possible, category-based filtering enables handling the latency problem.
Initial evaluations are showing expected good results, and large-scale testing is
currently ongoing.
References
1. Shardanad, U., Maes, P.: Social Information Filtering: Algorithms for Automating ‘Word
of Mouth’. Proceedings of the ACM Conference on Human Factors in Computing
Systems, ACM Press (1995) 210-217
2. Konstan, J.A., Miller, B.N., Maltz, D., Herlocker, J.L., et. al.: GroupLens: Applying
collaborative filtering to usenet news. Communications of the ACM 40 (1997) 3:77-87
3. Funakoshi, K., Ohguro, T.: A content-based collaborative recommender system with
detailed use of evaluations. In Proceedings of 4th International Conference on
Knowledge-Based Intelligent Engineering Systems and Allied Technologies, Volume 1
(2000) 253-256
4. Hayes, C., Cunningham, P., Smyth, B.: A Case-Based Reasoning View of Automated
Collaborative Filtering. In Proceedings of 4th International Conference on Case-Based
Reasoning, ICCBR2001 (2001) 243-248
Protecting the User from the Data:
Security and Privacy Aspects of Public Web Access
Adrian Spalka and Hanno Langweg
Department of Computer Science III, University of Bonn

Roemerstrasse 164, 53117 Bonn, Germany
[email protected]
Abstract. Observance of laws on the protection of minors is a serious problem

at places with Internet access. We examine protection mechanisms in view of
the Internet’s open and dynamic nature and claim that those based on deterrence
and restoration should be preferred to purely preventive ones. We then present a
strategy adequate for secondary schools. The technical part relies on adaptable
filter and logging components, which check and log requested web-pages. In a
feed-back loop a human auditor provides more information to the system, thus
increasing the effectiveness. On the other hand, separation of powers among
auditors ensures the users’ privacy.
1 Introduction
Schools use the Internet in the education of minors. But on the other hand, most of
them worry about misuse of the Internet. From the viewpoint of a school that makes
the Internet accessible to minors it is the users who need protection from harmful
material on the Internet. This situation is a departure from traditional computer
security paradigms and poses a range of new problems. We believe that measures to
protect minors from content should be adaptable to a school’s particular position.
As our empirical data reveal, at present left on their own schools with access to the
Internet pursue mainly two extreme approaches. The restrictive one confines access to
a small number of predetermined web sites. While suitable at elementary level
schools the pupils of which have little or no media competence, it only hides the
richness of information from older minors. The liberal approach allows unlimited
access to the Internet. If harmful material is discovered on the school’s computers, the
affair is most often innocuously swept under the carpet.
We have focused our attention on secondary schools with two aims in mind. The
Internet should be usable as an open environment and legal restrictions and the
school’s habits should be observed. Since preventive measures alone are inadequate
for this purpose, we favour a self-responsible and controlled access to the Internet,
depending on administrative, preventive and restorative methods. The result is a
system with both increasing technical precision and reliability and of educational
value. The system is implemented on a Microsoft Windows NT basis and the sources
can be freely obtained upon request to the authors.

Protecting the User from the Data: Security and Privacy Aspects of Public Web Access 441
2 Internet Access at Schools and Existing User Protection Efforts
Today many secondary schools in highly developed countries possess PCs; many of
them have the PCs connected in a local network and some operate an intranet. In the
first place one can think of applying the well-known protection methods accredited in
the industry, private and government institutions. Yet a look at the following list re-
veals that many of the conditions these methods rely on cannot be found at schools.
We can say of the pupils that they often have a competent understanding of
computers, are sometimes driven by a temptation to go over limits and have no
contractual obligations with the school. On the other hand we can say of many of the
teachers that they often have only little knowledge of computers, teach different
classes and that their acceptance of security measures degrades with increasing effort.
We see that security ergonomics must play an important role in the overall design.
At first glance we have a traditional threat scenario. The protection units are data,
programs and the hardware and they are all well known. Given that teachers are
assumed to behave in concordance with the rules, pupils are the only potential
intruders. The only protection goal is to protect data, programs and hardware from
adventurous or malicious-minded pupils – a task, which can be solved almost with
traditional means.
The situation changes drastically with the addition of Internet access. With this the
environment becomes open and dynamic. We now have a reversal in the protection
goal: it is the pupils who must be protected from the data. The problem is that there is
no technical way of completely specifying and locating the harmful content.
That on the Internet there is as much material proper to minors as harmful to them
is not new. Therefore, protection efforts are made at various levels. At the legal level
national laws can be enforced within the perimeter of their validity. Though useful,
the overall effectiveness of national legal measures is quite limited. As of yet
international legal agreement among most countries has been reached only on the
prohibition of child pornography – an essential but small part of harmful material.
Today’s technical systems rely solely on two filtering approaches, i.e. purely
preventive measures. The first one uses a web site’s address (e.g., NetNanny or Cyber
Patrol) and the second one a web page’s rating-label (e.g., the W3C PICS standard) to
determine if a requested page should be blocked or passed to the client.
3 A Comprehensive Protection System for Schools
Our approach is based on that the notion of protection of minors should not be
confused with the protection of young people from the world just because there are
dangers. Protection of minors is the protection from dangers in their education and
development and it can be best done by teaching them how to cope with dangers. Our
system as a whole is a combination of several types of measures: administrative,
preventive and restorative. Its effectiveness depends on an implementation of all of
them. It consists of four consecutive phases, which we subsequently describe in
detail: establishment of a school policy, automatic technical blocking and logging of
web accesses, semi-automatic post-processing, and sanctioning.
442 A. Spalka and H. Langweg
Phase I: The school’s security policy. The first phase is the establishment of a
written school law on the use of the Internet by pupils. It corresponds to a security
policy in the traditional sense. For the teachers it must comprise explicit statements on
their responsibilities, duties and prohibitions. They must accept it, approve of it and
act accordingly. For the pupils, the law must also comprise explicit statements on
their responsibilities and duties. Most important are the statements on the intended use
of the Internet, the automatic logging activities, the post-processing activities of
particular teachers to discover each improper use of the Internet and the types of
sanctions that can be applied upon improper use. We have prepared a skeleton policy,
which a school can adapt and extend according to its requirements.
Phase II: Automatic technical measures. Most schools have access to the Internet
through one or more dial-up lines. We assume that all these lines are connected to a
proxy, which our system can take control of. As a web server returns the requested
data, we intercept and process it in the following order. If the returned item’s web-
address can be derived from an entry in our white-list, it is immediately passed to the
client without any further action. If the returned item’s web-address can be derived
from an entry in the address-black-list of an enabled harmful category, the access is
logged and blocked, ie the client is notified of an improper access attempt. If the
returned item is a picture (GIF and JPEG at present) and its hash-value is found in our
picture-black-list of an enabled harmful category, the access is also logged and
blocked. Otherwise, the access is logged and the returned item is passed to the client.
The entries in the white-list are complete or partial web-addresses of content that is
considered appropriate without objections. Lists and categories are managed by the
school in phase III. We maintain two black-lists, one for addresses and one for hash-
values of pictures. An entry in the address-black-list comprises an address as in the
white-list and the categories it belongs to. We provide three classes: pornography,
racism and violence. A school is free to add more categories at its discretion.
A log-entry comprises all data to view the returned web-item later in the same way
as it was presented on the client’s computer and to identify client and time of access.
Phase III: Semi-automatic post-processing. The logged data are assessed and
evaluated. At first we use a classification program, which rates the logged web-
content’s degree of harmfulness with respect to existing categories. A list of weighted
keywords is associated with each category. The categories and their lists can be
changed by the school. The degree of possible harmfulness of a web-page is the
weighted sum of the keywords’ frequencies.
The next steps require manual intervention of two teachers. First, teacher A starts
the determination-function in a browser. He is shown a list of the logged web-
addresses and the computed rating in each category. He can order the list and can
selectively display a web-page or scan them like in a slide-show. It is now at his
discretion to decide if an access actually is a violation of the policy or not: If our
guess is right, he can mark this page and any pictures on it as harmful with respect to
the matching categories. The chosen part of the web-address and the pictures’ hash-
values are entered in black-lists. The page is marked for submission to identification.
If our guess is wrong, the page’s address can be entered in the white list.
From a purely functional perspective one teacher suffices to do both tasks: deter-
mination and identification. Possible objections with respect to the protection of both
Protecting the User from the Data: Security and Privacy Aspects of Public Web Access 443
the teachers’ and the pupils’ privacy led us to a separation of duties, which guarantees
the privacy of both parties.
Phase IV: Sanctioning. That decision is at the sole discretion of the school.
Advantages and disadvantages. Some drawbacks of our system are easily spotted.
Its effectiveness decreases as the volume of logged data increases. It is possibly
imprecise if not all logged data is scanned by a teacher and the system requires
regular manual intervention of teachers, at least 1-2 hours per week.
On the other hand, it offers a number of benefits. Improper use of the Internet is
detectable and the school has full discretion as regards determination of harmful
content and the decision on when a pupil’s behaviour should be counted as a
violation. Lists and assessment of content increase over time in concordance with the
school’s policy and can be extended in co-operation with other schools. Our system
uses ergonomic features; to the teachers it is like any other work on the Internet.
4 Conclusion
Schools have a legal obligation to protect pupils from harmful material on the Internet
but little technical assistance in accomplishing it. Owing to the global extent of the
Internet, legal regulations do not solve the problem in practice. We have examined
existing products and arrived at the conclusion that preventive measures are useful
and necessary but insufficient or inadequate for many reasons. Our approach gets its
strength from a combination of administrative, preventive and restorative methods. It
achieves the most important goals: it preserves the Internet’s open nature, offers pro-
tection from known harmful content and teaches how to deal with the unknown part.
References
1. Frishberg, M.: Protecting Kids, Here and There. http://www.wired.com/news/politics/
0,1283,36005,00.html, (2000)
2. Kobsa, A.: User Modeling, Privacy and Security. Invited talk held at the First International
Conference on Adaptive Hypermedia and Adaptive Web-Based Systems, (2000)
3. Kobsa, A.: Tailoring Privacy to User’s Needs. In: Bauer, M., Gmytrasiewicz, P.,
Vassileva, J. (Eds.). User Modeling 2001. 8th International Conference, UM 2001. Lecture
Notes in Artificial Intelligence 2109, (2001)
4. Net Nanny Software: Net Nanny 4 Product Description. http://www.netnanny.com/,
(2002)
5. SurfControl: Cyber Patrol Web Filter. http://www.surfcontrol.com/, (2002)
6. webwasher.com AG: Statement on protection of minors. http://www.webwasher.com/
en/corporate/minors.htm, (2002)
7. World Wide Web Consortium.: Platform for Internet Content Selection (PICS).
http://www.w3c.org/PICS, (1997)
Exploiting a Progressive Access Model to Information
Marlène Villanova-Oliver, Jérôme Gensel, and Hervé Martin
Laboratoire LSR – IMAG, BP 72, 38402 Saint Martin d’Hères cedex, France.
Tel.: (33) 4 76 82 72 80, Fax: (33) 4 76 82 72 87
{villanova,gensel,martin}@imag.fr
Abstract. We propose here to stratify the Information Space of a Web-based

Information System (WIS) by decomposing it into personalized sub-
Information Spaces. This stratification is described through a Progressive
Access Model (PAM) written in UML. The PAM gives WIS users, first, access
to some minimal and essential information, and then, allows them to navigate
through larger and/or smaller personalized Information Spaces. Together with
the PAM, we present a specific query language which allows to query a
stratification and to format replies according to the different levels of details of
the stratified user’s Information Space.
1 Introduction
The appropriateness of information a WIS delivers to its users turns out to be an acute
problem when designing such systems one ignores that 1) all users do not need the
same information, 2) users do not need all the available information all the time. The
first point often referred to as adaptability (see for instance [1]) can be addressed by
distinguishing users profiles. Recent approaches [1] announce a new generation of
web applications able to track down the behaviour of users and to dynamically adapt
information to users rights, needs, characteristics and material configurations (WAP,
browser, etc.), in terms of both content and presentation. The second point deals with
the delivery of information and the fact that users may occasionally need only some
parts of information. In order to protect users against a cognitive overload, we provide
them with a gradual organization for progressively accessing information. In [3] we
have defined two notions for organizing at a conceptual level the data of a WIS so
that users are offered a personalized and progressive access. The data model of the
WIS is stratified into different levels of detail which are exploited by progressive
access mechanisms. Once stratified, a data model constitutes an Information Space in
which a (group of) user(s) can navigate gradually from a level of detail to another.
The stratification is described through an extensible Progressive Access Model
(PAM), which has been implemented in KIWIS, a platform for the design and
automatic deployment of WIS [4].
In this paper, we propose first a UML description of the PAM for object-oriented
WIS. Then, we introduce a query language, called PAM QL, which allows to query a
stratified Information Space. Not only query expressions can concern the stratification
itself, but also levels of detail can be used in query expressions. The reply to a query,
whether this latter explicitly refers to the stratification or not, is formatted according
to the levels of detail defined by the stratification of the user’s Information Space.

Exploiting a Progressive Access Model to Information 445
2 The Progressive Access Model

The Progressive Access Model (PAM) stratifies the Data Model of the application
domain addressed by the WIS by introducing different levels of detail. In [3] we
describe how the PAM is coupled with both the Data Model and a model of the users,
called Generic User Model (GUM). The connections between these three models
allow a WIS designer to define as many adapted stratified views on the data model as
required for individuals or groups of users.
We give here a UML description of the PAM in the particular case of an object-
oriented Data Model (cf. Fig. 1). The class UserCategory connects the PAM to the
GUM. The class MaskableEntity (ME) corresponds to the set of information to be
stratified. It specializes in two classes called ClassDiagram and Class, since the
stratification can be performed at these different granularities. The class
Representation of Maskable Entity (RoME) denotes the different representations
available for a ME and which constitute its stratification. Each RoME of this set
corresponds to a level of detail at which the ME is presented more or less completely.
The sub-classes of RoME, called RoME_DataModel and RoME_Class, are
respectively defined for the two sub-classes of ME. The association structures
between the classes UserCategory, ME and RoME (or theirs sub-classes) is so that:
- 0 to n instance(s) of RoME can be linked to one couple of instances
(UserCategory, ME). No (0) instance means no stratification: the user is given no
progressive access to information, which is delivered in a whole. On the contrary, n
instances of RoME offer to the user n levels of representation for this ME.
- 1 to n instance(s) of UserCategory can be linked to one couple of instances (ME,
RoME), allowing several users to share the same stratification.
- one and only one instance of ME is linked to one couple of instances
(UserCategory, RoME), meaning that the stratification is proper to one ME.
UserCategory
1..* userCat
info rep
MaskableEntity RoME
1 *
structures
1..* 1..*
1..* belongsTo 1..* RoME_ClassDiagram RoME_Class Attribute
Class_Diagram Class
belongsTo
Fig 1. PAM extended description, fitting an object-oriented data model.
The class RoME_DataModel is linked to the class Class through the association
belongsTo. Linking a set of instances of Class to one instance of RoME_DataModel
means that these classes are accessible at the level of detail to which corresponds this
instance of RoME_DataModel. The class RoME_Class is linked to the class Attribute
through the association belongsTo. Linking a set of instances of Attribute to one
instance of RoME_Class means that these attributes are accessible at the level of
detail to which corresponds this instance of RoME_Class.
446 M. Villanova-Oliver, J. Gensel, and H. Martin
3 PAM QL: A Query Language for the PAM

Once the stratification has been created or modified by instantiating the PAM, the
PAM Query Language, which is based on a OQL-like formalism [5] can be used i) for
querying the stratification performed on an Information Space, ii) for querying the
Data Model and getting reply formatted to the stratification (i.e. according to levels of
detail). We give below examples of a query on stratification (left part) and of a query
which includes stratification features used in the formatted reply (right part):
GIVE STRATIFICATION SELECT ME.information, …
FROM maskable_entity ME FROM maskable_entity ME
[ FOR all_levels | level_interval [WHERE expression]
= [min, max] ] [ FOR all_levels | level_interval
= [min, max] ]
A query based on the GIVE STRATIFICATION element returns the different levels of
the stratification defined for the maskable entity ME (by default, from level one to the
current level). SELECT FROM queries (on the right part) search for the instances of ME,
whose presentation is limited to the field(s) of the information associated with ME, and
which respect the expression given in the optional clause WHERE. When ME is a
schema, information concerns its classes, while when ME is a class, information concerns
its attributes.
C ou rse
S tu d en t
Name
T each er C o ef follow s N a m e
h o urlyV o lu m e 1 ..* 1 ..* F irstn a m e
Name B irth D ate
1 ..* 1 ..* 1 ..* A d d ress
F irstn a m e T rain in g
S ervice tea ch es 1 ..*
Name 1 ..*
S a la ry
is_ reg istered
T em p ora ry P erm a n en t
RoM E 1
S a la ry M in S ervice
C o ntra ctT ype G ra d e RoM E 3 RoM E 2
Fig. 2. Example of a stratified schema (3 RoME are defined).
If * is used in the SELECT, information corresponds to the RoME of the current level.
In these two cases, the optional clause FOR allows either to expand the query to the
whole stratification (i.e. all the levels are presented) or to limit the list of the presented
levels to the given interval.
Some examples of such queries applied to the stratified Information Space of Fig. 2
are given in Table 1. Due to space limitation, we only present queries about the
schema although a similar approach exists for stratified classes.
Exploiting a Progressive Access Model to Information 447
Table 1. Examples of PAM QL queries applied to the stratified Information Space of Fig. 2
CURRENT LEVEL QUERY RESULT
Schema RoME 3 GIVE STRATIFICATION Level 1 : class Training, class Student
FROM schema Training Level 2 : class Course
Level 3 : class Teacher
Schema RoME 3 GIVE STRATIFICATION Level 1 : class Training, class Student
FROM schema Training Level 2 : class Course
FOR level_interval = [1, 2]
Schema RoME 2 SELECT * All instances of classes Training, Student and
FROM schema Training Course are listed, which corresponds to RoME
2. Each class, if it is stratified, is represented at
its lower level of detail
4 Conclusion
In this paper, we have presented a way to stratify the Information Space (data model)
of a Web-based Information System (WIS) in order to provide users with a
progressive access to information. This stratification is described by a Progressive
Access Model (PAM) written in UML. Users of a WIS can access, first, essential
information, through different levels of granularity, and then, more or less
information, depending on their interest, time or material configuration, etc. We have
also proposed PAM-QL a OQL-like query language which allows to query the WIS
about the stratification performed on the Information Space, but also aims at
formatting the data contained in a reply, according to the stratification (i.e. according
to levels of detail).
Our research is now directed towards dynamic adaptability techniques in order to
react more efficiently to end-users actions. The idea is to dynamically elaborate and
modify both the stratification and the navigation schema of a WIS, learning from the
user's behavior. The use of cookies technology to track information about users’
sessions in order to automatically adapt information, coupled with some meta-rules of
navigation, is one way we have started to explore.
References
1. Brusilovski P.: Methods and Techniques of Adaptive Hypermedia. User Modeling and User-
Adapted Interaction, Vol. 4, Kluwer academic publishers (1996) 21-45
nd
2. De Bra P.: Design Issues in Adaptive Hypermedia Application Development, 2 Workshop
on Adaptive Systems and User Modeling on the WWW, Proc. of WWW’99, (1999) 29-39
3. Villanova-Oliver M., Gensel J., Martin H.: Stratification of the Information Space in Web-
based Information Systems, ECIS’02, Gdansk, Poland, June 6-8 (2002)
4. Villanova-Oliver M., Gensel J., Martin H., Erb C.: Design and Generation of Adaptable Web
Information Systems with KIWIS, ICIT’02 : Coding & Computing. Session on Web &
Hypermedia Systems, Las Vegas, NV, US, April 8-10 (2002)
5. Cluet S.: Designing OQL: Allowing Objects to be Queried, Information Systems. 23(5)
(1998)
Adapting to Prior Knowledge of Learners
Stephan Weibelzahl and Gerhard Weber
Pedagogical University Freiburg, Germany

{weibelza,webergeh}@ph-freiburg.de
Abstract. Prior knowledge is an important factor that influences the

interaction with a hypertext and the learning gain. Our authoring sys-
tem NetCoach provides a way to assess the users’ prior knowledge and
to adapt the course in different ways. We describe how the adaptation
mechanism assesses the user’s knowledge with test items, infers the user’s
current learning state from this information, and finally adapts accord-
ingly. An evaluation study with an adaptive HTML course demonstrates
that this kind of adaptation might reduce the completion time, but re-
tains the learning gain.
1 Online Learning Courses and Prior Knowledge

Online learning courses are used by people that differ widely in prior knowledge
of the domain. Especially in further education and learning on demand settings
some learners will have a background of parts of the course while others are com-
plete beginners. However, regardless of the prior knowledge, everybody should
have the same knowledge after course completion. On the one hand users might
get bored if they have to work on topics that they are already familiar with. On
the other hand they are probably not able to estimate whether they do really
know everything on a topic of a course without having seen the chapters. Thus,
letting users decide on their own whether they have enough knowledge or not
might result in incomplete knowledge acquisition. Moreover, prior knowledge
has an impact on the learning gain. When constructing a hypertext authors
should considers the users’ prior knowledge [3]. It might be useful to adapt the
hypertext’s structure [2] or to provide different advisements [4].
In any case such adaptations require the assessment of prior knowledge. The
authoring system NetCoach [6] provides a mechanism both to assess the user’s
prior knowledge and to adapt the course accordingly.
2 Adaptation Mechanism
NetCoach is designed to enable authors to build adaptive online courses without
programming. While authors generate the content by filling in templates and
forms, the course functionality including user management, adaptation, commu-
nication facilities, and tutoring is provided by NetCoach.
Each NetCoach course implements several adaptive features, that require an
adequate assessment of the prior knowledge. Based on this information NetCoach
infers the user’s knowledge about each chapter and decides how to adapt.

Adapting to Prior Knowledge of Learners 449
2.1 Assessment of Prior Knowledge

Similar to a book, a NetCoach course is structured hierarchically with chapters
and subchapters. The knowledge about a subchapter or a set of subchapters is
assessed in so called test groups, i.e., a set of test items. Three types of test
items are evaluated on the fly automatically: multiple choice tests, forced choice
tests, and gap filling tests. Authors may define the difficulty of the test items
and a critical value of points that are required for completing the test. After the
learner has completed the chapter the items are presented in small groups (e.g.,
two in a row). Users achieve points for correct answers and loose points for false
answers until they reach a critical value.
In the beginning of a chapter that has subchapters with test groups a link to
a pretest is offered. In this case a random selection of the test groups’ difficult
items is presented.
In summary, the test items of test groups in subchapters which are usually
presented after the chapter are aggregated and now serve as pretest for the
superordinated chapter.
2.2 Inference of Current Learning State

Each item is evaluated as if it was answered in the test group. The only ex-
ception is that the result is weighted (default is 1.5). Afterwards, the current
learning state is updated, i.e, NetCoach computes on the fly whether a chapter
is solved, inferred, ready or not ready. See Weber, Kuhl and Weibelzahl [6] for
a detailed description of the overlay model and the different states of chapters
used in NetCoach. Such an inference mechanism requires information about re-
lations between chapters which is stored in a so called knowledge base. NetCoach
supports the two concept relations prerequisite and inference [6].
In summary, the inference of a user’s current learning state is done in two
steps. First, the items of the pretest are evaluated and points weighted by an
factor are assigned to test groups accordingly. Second, if a test group has been
solved because the critical value has been reached, this chapter is assumed to
be learned and other chapters might be inferred to be learned in correspondence
with the knowledge base.
2.3 Adaptation Decision

Finally, the learning state is used for two types of adaptations: adaptive anno-
tation and curriculum sequencing. First, links to other chapters are annotated
according to the state of this chapter (e.g., a green bullet indicates that the
chapter is ready to be visited).
Second, NetCoach suggests which page to visit next and thus guides users to
chapters that are ready and that are required to complete the course. Chapters
that are assumed to be learned are skipped. Thus, the pretests provide an easy
way to guide users to those chapters that they should learn. Other adaptation
techniques that use the same information would be possible (e.g., link hiding [1]
or changing the adaptation method [5]), but are not implemented in NetCoach.
450 S. Weibelzahl and G. Weber
(a) (b)
Fig. 1. (a) Duration of interaction (seconds). People who solved the pretest on one of
the three chapters required less time to complete these chapters (including the time to
complete the pretest). (b) Relative number of correct responses in the posttest. People
who solved the pretest on one of the three chapters gave equally or more correct
responses to test items in the post test. The standard deviations are indicated.
3 Evaluation
The pretest mechanism of NetCoach has been evaluated with the so called
HTML-Tutor. This online course introduces to publishing on the web and pro-
gramming HTML. We wanted to know whether the pretests can assess the prior
knowledge correctly and how the learning gain is influenced by the adaptation.
We observed a total of 140 users who accessed the public course from all over
the world. Two groups of users are distinguished: the first group (no pretest)
ignored the pretest and completed the chapters as usual, while the second group
(pretest presented) decided to answer the pretest. Consequently, most of them
were advised to omit at least some subchapters. In the end of the course users
completed a final test that included several test items on the pretest chapters. If
the pretest assessment was successful the second group should know as much as
the first group about the chapters, even though they did not read the contents.
We found that the pretest group completed the chapters much faster than
the standard group. For all three pretest that have been included in this analysis
the mean duration of interaction was lower (see Figure 1a). A 2-factor MANOVA
yielded significant differences between these groups (see Table 1). However, the
analysis of the posttest shows that the pretest group had at least as much know-
ledge on these chapters as the standard group. Their relative number of correct
responses was even higher for chapter 1. The remaining chapters did not differ
significantly (see Figure 1b).
In summary, the pretest group had at least as much knowledge about the
chapters although they spent less time for browsing these chapters. Note that
the users saved up to 80% of the interaction duration, but performed about 10%
to 20% better in the posttest. Especially for learning on demand settings when
people want to learn specific contents as efficiently as possible this could be an
important benefit.
Adapting to Prior Knowledge of Learners 451
Table 1. (a) Statistical results of a 2-factor MANOVA (dependent variables duration

and knowledge). (b) Statistical results of a 2-factor ANOVA (dependent variable du-
ration). For each factor the F-value (F), the degrees of freedom (df ef f ect, error ) and
the statistical significance (α) are reported (sample size N=140).
(a) (b)
factor F df α factor F df α
F1: presentation 15.84 2, 133 .000 F1: presentation 31.07 1, 134 .000
F2: chapter 3.15 4, 268 .015 F2: chapter 6.56 2, 134 .002
F1*F2 2.82 4, 268 .026 F1*F2 5.54 2, 134 .005
4 Conclusion
NetCoach offers easy creation of pretests by summarizing items from test groups.
The assumed state of each chapter is computed in dependence of the answers
in the pretest and the relations between chapters. Our results suggest that the
HTML-Tutor assesses the prior knowledge correctly. Despite the fact that people
were adaptively guided to omit those chapters that are assumed to be learned,
they were able to answer test items on the chapter’s contents even better than the
standard group. Thus, assessing knowledge with test items facilitates interesting
adaptation opportunities. Adapting to prior knowledge is an important approach
to increase the effectiveness and efficiency of learning courses and might even
increase the users’ satisfaction.
References
1. P. De Bra and L. Calvi. AHA! An open adaptive hypermedia architecture. The
New Review of Hypermedia and Multimedia, 4:115–139, 1998.
2. S. McDonald and R. J. Stevenson. Effects of text structure and prior knowledge
of the learner on navigation in hypertext. Human Factors, 40(1):18–27, 1998.
3. I. Park and M. J. Hannafin. Empirically-based guidelines for the design of inter-
active multimedia. Educational Technology Research & Development, 41(3):63–85,
1993.
4. E. C. Shin, D. L. Schallert, and W. C. Savenye. Effects of learner control, advise-
ment, and prior knowledge on young students’ learning in a hypertext environment.
Educational Technology Research & Development, 42(1):33–46, 1994.
5. M. Specht. Empirical evaluation of adaptive annotation in hypermedia. In
T. Ottmann and I. Tomek, editors, Proceedings of the 10th World Conference on
Educational Telecommunications, ED-MEDIA & ED-Telecom ’98, Freiburg, Ger-
many, pages 1327–1332, Charlottesville, VA, 1998. AACE.
6. G. Weber, H.-C. Kuhl, and S. Weibelzahl. Developing adaptive internet based
courses with the authoring system NetCoach. In S. Reich, M. Tzagarakis, and
P. de Bra, editors, Hypermedia: Openness, Structural Awareness, and Adaptivity,
pages 226–238, Berlin, 2001. Springer.
Monitoring Partial Updates
in Web Pages Using Relational Learning
Seiji Yamada and Yuki Nakai
CISS, IGSSE, Tokyo Institute of Technology

4259 Nagatsuta, Midori, Yokohama, 226-8502, Japan
[email protected]
Abstract. This paper describes an automatic monitoring system that

constantly checks partial updates in Web pages and notifies the user
about them. While one of the most important advantages of the WWW
is frequent updates of Web pages, we need to constantly check them out
and this task may take much cognitive load. Unfortunately applications
to automatically check such updates cannot deal with partial updates like
updates in a particular cell of a table in a Web page. Hence we developed
an automatic monitoring system that checks such partial updates. The
user can give a system regions in which he/she wants to know the updates
in a Web page as training examples, and the system is able to learn rules
to identify the partial updates by relational learning. We implemented
the system and some executed examples were presented.
1 Introduction
We currently obtain various information from the WWW and utilize them. While
one of the most important advantages of the WWW is its constant updates of
Web pages, we needs to frequently check the updates for acquiring the latest
information and this task forces much cognitive load on us. Thus, a number of
applications to automatically check and notify updates of Web pages have been
developed[1,2]. Unfortunately almost all of them notify the suer about updates
whenever any part of a Web page is updated, and most of such updates may not
useful to him/her.
Consider a weather report Web page and a user who has a plan to go to
a picnic on the next Sunday and is interested in the weather. He/she needs to
frequently check the next Sunday’s weather in the Web page. If a user employs
a Web update checking application, it notifies him/her all of updates including
other days’ weather changes thought such notifications are meaningless. Thus
partial update is defined as an update of a region in which a user is interested,
not of any part of a Web page. We argue that this partial update monitoring
is widely necessary in a lot of fields like stock market pages, the exchange rate
pages, etc.
We developed an automatic monitoring system PUM (Partial Update Moni-
toring) that constantly checks partial updates in Web pages and notifies the user
about them.

Monitoring Partial Updates in Web Pages Using Relational Learning 453
www
Updated
Web page
3-(b)
Identification of Fail
a monitor region
Unnecessay
Success 3-(b)-i
Notification
Information of Necessary
monitor region User
RI rules UC rules
4. Evaluate
rules
2-(b) Rule generation 2-(b) Rule generation

Indicate
Relational learning Relational learning update 3-(b)-i-A.
Indicate update
3-(b)-i-B.
Traning examples Training exmaples Good Evaluate update
region indentification update check
Evaluate update
Diff.
No good
Last
updated Give
2-(a) page monitor region
Genarating
training examples
1.Obtaining
a monitor
region
Fig. 1. System overview.
2 PUM: Partial Update Monitoring in a Web Page

2.1 System Overview
PUM is a system that identifies a region indicated by a user in a Web page,
checks for partial updates in the region and notifies the user about the updates
which he/she wants to know. Figure 1 shows overview of PUM , in which a dotted
line indicates interaction between a user and PUM .
Figure 1 also stands for the procedure of PUM . First of all, PUM obtains a
monitor region from a user. The user indicates a region in which he/she wants
to know the update by mouse highlight operation on interface of PUM (Fig. 2).
Next PUM extracts training examples for both of RI (region identification) and
UC (update check) from a region indicated by a user.
Then a relational learning system automatically acquires two kinds of rules
for region identification and update check. PUM utilizes RIPPER[3] as a re-
lational learning system. RIPPER acquires rules to classify examples into two
classes, and the learned rule is described with a symbolic representation.
After such rules are generated, PUM becomes able to identify a partial update
and to determine whether it is one which a user wants to know or not. It does so
454 S. Yamada and Y. Nakai
Fig. 2. Interface of PUM .
by using two kinds of rules (RI and UC). If PUM decides that an update is useful
to a user, it notifies the user about the update. Otherwise, PUM indicates the
updated Web page to a user and asks for his/her evaluation. PUM is implemented
using Visual C++ and Ruby on Windows2000.
Figure 2 shows the interface of PUM . The window consists of three sub-
windows: a Web browser window, a URL window and a training example window.
The Web browser window (lower right in Fig.2) shows a Web page in the same
way as a Web browser and a user can easily indicate a region by highlighting
it using the mouse. The URL window (upper in Fig.2) shows URLs of updated
pages. The training example window (lower left in Fig.2) indicates a table of
attributes and values of stored training examples.
2.2 Negative Examples for Region Identification
Since relational learning is a kind of inductive learning, negative examples play

an important role to avoid over-generalization. Thus PUM automatically gener-
ates negative example for region identification to improve learning efficiency.
We consider the neighborhood of an indicated region as near miss examples.
Hence PUM generates negative examples from four regions: left, right, upper and
lower regions to an indicated region.
Monitoring Partial Updates in Web Pages Using Relational Learning 455
Table 1. RI rules.
Eval. No. Class Condition

1 Good cIndex ˜’10/14(Sun)’, rNo ˜’7’.
2 Good rNo ˜’7’, cIndex ˜’Sun’.
3 Executed Examples
A typically successful example for PUM is on updates in a weather report Web
page shown in Fig.2. This page shows the weather forecast for the next seven days
in a table that scrolls horizontally. In this example, a user wants PUM to notify of
an update when the rain probability of Tochigi on Sunday (a highlighted cell in
Fig.2) decreases to less than 40%. Thus PUM needs to learn RI rules to identify
a cell indicating weather probability of Tochigi on Sunday and UC rules to check
whether the value of weather probability is less than 40. PUM successfully was
able to to extract the correct partial update after several evaluations by a user.
Table 1 shows the number of user’s evaluations and learnt RI rules at that
time. A rule consists of “Class” and “Condition”, and if the update satisfies
“Condition”, it is classified into the “Class”. A˜B in “Condition” means a con-
dition that B is included in an attribute A. An RI rule learnt from the first
evaluation identified a cell which is in the 7th-row and has ’10/14(Sun)’ as a
column index. This rule succeeded in identifying a region for four days; however
it failed on the fifth day. This is because a target region included ’10/21(Sun)’
instead of ’10/14(Sun)’ by scrolling. Then, PUM required a second user’s evalua-
tion and learned a new rule shown in Table 1. This second rule identifies a correct
cell using more general condition ’Sun’ as a column index, not ’10/14(Sun)’.
Additional successful examples were investigated in stock market, CD rank-
ing, exchange rate Web pages, etc.
4 Conclusion
We proposed a monitoring system PUM that constantly checks partial updates in
Web pages and notifies a user about them. The user can give a system regions for
which he/she wants to know the updates in a Web page as training examples; the
system can then learn rules to detect the partial updates by relational learning.
We implemented our system and presented some executed examples.
References
1. Web Secretary. (http://homemade.hypermart.net/websec/)
2. Saeyor, S., Ishizuka, M.: WebBeholder: A revolution in tracking and viewing
changes on the web by agent community. In: WebNet 1998. (1998)
3. Cohen, W.W.: Fast effective rule induction. In: Proceedings of the Twelfth Inter-
national Conference on Machine Learning. (1995) 115–123
Adaptation in the Web-Based Logic-ITA
David Abraham and Kalina Yacef
School of Information Technologies, University of Sydney, Australia

[email protected], [email protected]
Abstract. The Logic-ITA is an Intelligent Teaching Assistant system for the

teaching/learning of propositional logic. The system is tailored to two different
types of uses: for students, it is an autonomous Intelligent Tutoring System,
whilst for teachers, it includes functionality to set up learning levels, adjust pa-
rameters for progressing through these levels, monitor the class’ progress and
collect data. This paper presents some characteristics of the Web-based version
of the Logic-ITA. In particular, the web-based version allows a centralisation
on the server of the student models, which contain all the users’ individual
information related to their learning. This means that all the adaptation features
of the Logic Tutor are maintained, whilst providing (1) students with a large
degree of flexibility in terms of when and where they choose to use the tool,
and (2) teachers with the possibility to monitor accurately the results and statis-
tics.
1 Introduction
The aim of the Logic-ITA is to enhance the adaptation of the teaching in the context
of large classes. The ratio of one professor to several hundreds students necessarily
mean that the teaching strategy will be broad- leaving many students to experience
difficulties. Problem solving sessions (also called tutorial classes) help, but then again
the ratio is 1 to 20, and with a different person than the professor. The Logic-ITA
aims to be another teaching assistant, but one that interacts with all students at a
detailed level. It does this by providing students with an intelligent tutoring system,
which allows them to practice formal proofs, receive adaptive feedback and attempt
personalised exercises. And while doing all of this, the Logic-ITA is able to report the
student’s progress back to the teacher and tutors. The disproportional ratio between
teachers and students is broken down so the teacher can redirect his/her teaching style
to the class level, address specific difficulties. The Logic-ITA also addresses the
students’ feeling of being ‘lost in the mass’ because recurrent difficulties and
problems that are not resolved during the interaction with the system are brought to
the teacher’s attention.
As with any adaptive system, the availability of a user model is paramount. In a
learning context, it is called a student model. The data collected from the student’s
interactions with the system is necessary for adapting the teaching, and also for
providing feedback to the teaching team. However these models must be unique and
easily accessed by the teacher to ensure that use of the adaptation features is
maximized.

Adaptation in the Web-Based Logic-ITA 457
The Logic-ITA has been integrated in our Languages and Logic course, since
2001. We have now built a web-based version to help centralise the data while
making it more convenient for students for students. This paper highlights the
characteristics of the web-based Logic-ITA. The following section introduces the
reader to the tool. Then Sect. 3 describes the main adaptation features and finally
Sect. 4 explains the advantages of the web-based version.
2 Overview of the Logic-ITA
The Logic-ITA contains several tools: the Logic Tutor, the LT-Configurator, and the
LT-Analyser. The general architecture is outlined below.
Logic Student
Tutor LT-
model
Teacher interface
Analyser
Expert Pedagogical
module module
LT-
Configurator
Student interface
Fig. 1. General architecture of the Logic-ITA
The Logic Tutor is an Intelligent Tutoring System that allows students to

practice formal proofs, whilst providing them with context-sensitive feedback and
tailored exercises. A multimedia presentation of the Logic Tutor can be found at [1].
The system creates and maintains a Student model for each individual user. It
records all the attempted exercises along with the mistakes, and stores the student’s
level, performances, and student information. The student is able at any time to
browse through this data and compare his/her level and results with those of the class.
We believe that allowing students to scrutinise their student model enhances their
learning [4].
The Expert module contains the expertise of the system in propositional logic: it
contains the rules of equivalence and laws of inference, and it is able to apply them
validly. Proofs are not unique. Paths that the student follows to derive a conclusion
from given premises can vary. What really matters is that each step is validly derived
from the previous lines and that the conclusion is finally reached. The Expert module
does not solve the exercise but is able to check the correctness of students’ answers
dynamically as well as producing proper feedback when errors occur. This feedback
is provided using a bug library.
The Pedagogical module is the main module of the system. It contains the high-
level rules for sequencing the exercise training objectives, using the student model
data. It updates the student model, and then suggests exercises for the student to solve.
However it allows the student to select other exercises or even to create a new one,
thus leaving him/her in control of the learning.
The curriculum sequencing in the Logic Tutor is semi-dynamic. Through the
LT-Configurator, the teacher defines students’ learning levels and sketches the high-
458 D. Abraham and K. Yacef
level sequencing of the progression through the levels. The finer grain of training
objectives is decided on the fly with the data stored in the student model. The LT-
Configurator is a super-version of the Logic Tutor. Configuration settings are stored
in separate files.
Finally, the teacher can monitor the class’s results, levels and problems through
the LT-Analyser. These results are stored in a database and displayed graphically
according to the criteria entered by the teacher. For example, the teacher may look at
the performance results of the class on a particular exercise, or on all exercises
involving a particular rule, or the distribution of levels among a given group of
students, and so on.
The Logic Tutor and the LT-Configurator are written in Java, and use the Java
Swing library for the user interface. The LT-Analyser uses Microsoft Access and
Excel.
3 Student Model and Adaptation
The key to adaptation in any system is a user model, or student model in this learning
context. For each user of the system, the student model maintains up-to-date
information about the user that is relevant to the system: background, knowledge,
level, preferences and so on. In the Logic Tutor, the student model contains the user
id, user’s preferences, level, history of exercises and respective performances, number
of rules used correctly and incorrectly (broken down per rule), the number of mistakes
made (broken down per type of mistake). Below we describe how the Logic-ITA uses
this information to adapt to individual users.
Curriculum Sequencing. The educational material provided by the Logic Tutor

consists of a large set of exercises and simple HTML tutorial pages. The system is
able to choose exercises that are best suited to the student’s needs by using the student
model and the parameters set by the teacher in the LT-Configurator. High-level
parameters define the difficulty of the exercise and class of rules that can be used.
Within these constraints, the rules are given weights according to the student model
data. The exercise obtaining the higher weight will be selected for the student.
Exercises are indexed with the levels of difficulty and the rules they are likely to
involve.
First, the system calculates the level of difficulty that the student should face.
This is done using the previous performance results. At the end of each exercise, a
performance is calculated. There are discrete numbers, ranging from 1(the student has
solved the problem without making any mistake and using a minimal number of lines)
to 5 (the student has not solved the exercise and made mistakes). They are used to
assess how well the student is coping with the current level of difficulty. The range of
exercises that are considered for this evaluation is determined by the teacher in the
LT-Configurator.
Once this level is determined, rules are selected according to the student model.
For example, if a student is at level 2 and has made repeated mistakes with, say, the
Modus Tollens and Addition rules, an exercise using these two rules is more likely to
be selected next.
The student has the choice of either following or ignoring the suggestion made by
the system. This is similar to the mechanism used in ELM-ART[2]. In this sense, it
provides an adaptive learning support: the user is free to make a choice while still
seeing an opinion of an intelligent system [3].
Tailoring of Interface. Students are able to make a number of small changes to the
user interface. This allows the students to personalise their learning environment to
what they are most comfortable with. These personalisations are stored in the student
model.
Tailoring of the Human Teacher’s Interventions. One of the goals of the Logic-
ITA is to help the teacher adapt the content of lectures to the class’ current state of
learning, and attend to individuals’ problems. The student’s interactions results in
adaptation of the system to the student. Here, there is an additional level of
adaptation, which occurs at the teacher level: with the information provided, he/she is
in a better position to adapt his/her teaching material at the class level as well as the
individual level.
Using the LT-Analyser, the teacher can monitor the class’ progress by querying a
database containing users’ information. Some examples of queries are (for the whole
class, a subgroup or an individual):
- Which rule caused the most mistakes
- What were the 5 most common mistakes
- Which exercise caused the most mistakes
- What was the average number of mistake per student
- Trend in the number of mistakes in time, per rule and for all rules
- How many students are still at level 1?
Tested in 2001 at the class level, the LT-Analyser was found very helpful. It helped
the teacher to re-direct the content of the lectures, address the common
misconceptions and focus the revision lectures. In 2002 it is planned to use it at a
problem solving session level and individual level.
Tailoring of Feedback and Hints. When the student makes a mistake, the system
queries a bug library, which contains mistake patterns. Instead of giving a generic
feedback, the patterns are instantiated to the current mistake: the explanations given to
the student are contextualised with the current exercise and lines entered by the user.
For instance, if the student uses the ‘Simplification’ rule before the ‘Commutation’
rule to deduce the right hand side of a conjunction, the mistake module will use the
original formula to construct an explanation detailing how only the left hand side can
be deduced by Simplification.
Similarly, when hints are available, the user reads them in the context of the
current exercise. According to the evaluation survey, this feature contributes greatly
to the feeling of being individually assisted.
When no patterns can be found for the current mistake, the feedback states the rule
that has not been respected.
460 D. Abraham and K. Yacef
4 Why a Web-Based Version?
The original Logic Tutor was written as a Java application and so could easily be
downloaded and run on a local computer. However, this type of usage would lead to a
number of problems:
- The Intelligent Tutoring System will work to some extent, but not the Intelligent
Teacher-Assistant System. The teacher side of the Logic-ITA will not have
access to the student models – because they are stored locally on the students’
computers. This means that a significant part of the Logic-ITA cannot be
exploited.
- The ITS will work normally only if each user always use the same copy of the
Logic Tutor. If a student downloads the Logic Tutor onto two different machines,
or uses two different accounts, then there will be two different and incomplete
student models representing the same student. Therefore each system will have
an inaccurate model of the user, resulting in a slower progress in the curriculum,
inadequate feedback, duplication of exercises for the student, and of course
inaccuracy of the information provided to the student about his/her model.
Therefore we can see that to fully exploit the Logic-ITA, it is paramount that the
student models be unique and stored on a location that is both known and accessible
to the teacher side of the Logic-ITA.
Last year we released the Logic Tutor to the 450 students enrolled in our
Languages and Logic course on a Unix central account. Students could only use it
from their university undergraduate account and a script, run each night, collected the
relevant information from their home directory (they knew this was happening, of
course). Whilst this resolved the access of the student models, it was not very
satisfactory, because it created time and geographical constraints for the students.
Three weeks before the exams, we made the Logic Tutor fully available for
download. Constraints were released for student use but of course we no longer had
access to their student models.
Fortunately, WWW technology can provide a solution to this problem:
- centralised storage of student models on the server. Uniqueness of each student
model is guaranteed and the teacher has immediately access to up-to-date
information about the student’s learning progress and difficulties.
- time and geographical flexibility for the students. Students can study when and
from any computer they want, without carrying their student model with them or
creating duplicates.
- easy update of the software and databases. New exercises, new configurations
and software update are easier to manage.
- We have now completed the last stage of the implementation of the Logic-ITA
and have a web-based version ready for use in 2002. Transforming the Logic
Tutor java application into a java applet was not difficult. There are of course
issues specific to the web-based version:
- download times: the Logic Tutor is packaged as a 400Kb compressed Java
Archive (jar) file. This represents a significant download time: approximately 80
seconds with a 28800 bps line speed. However, once the file has downloaded, the
program’s response times are the same as the non-web version. Furthermore,
after the first download, many browsers will place the file in cache. This means
that subsequent download times are comparable with the boot time of the original
non-web-based Logic Tutor.
- security: accounts administration has been centralised to avoid users creating
multiple student models. University students are given an account corresponding
to their login, and external users need to request their user account.
- student models are kept private and cannot be accessed externally except by the
relevant student. All requests for a student model are forced to go through a
server program, which requires a valid user name and password pair before
granting access.
5 Conclusion
The Logic-ITA is an Intelligent Teacher-Assistant System requiring centralised

storage of the student models. We have tested a non-web-based version last year with
450 students enrolled in the Logic course. The user evaluation survey showed that,
whilst the learning benefits of the Logic Tutor were appreciated, the students would
have preferred the freedom of using it from both their Unix undergraduate accounts
and home.
The web-based version of the Logic-ITA is now ready for use this year and will
enable students to work from anywhere with Web access. It also enhances the
adaptation features and accuracy of the data of the Logic Tutor by ensuring unique
and centrally located student models.
References
1. Abraham, D., Crawford, L., Lesta, L., Merceron, A., Yacef, K. (2001) The Logic Tutor: A
multimedia presentation, Interactive Multimedia Electronic Journal of Computer-
Enhanced Learning, October 2001 issue (http://imej.wfu.edu/articles/2001/2/03/index.asp).
2. Brusilovsky, P., Schwarz, E., and Weber, G. (1996) ELM-ART: An intelligent tutoring
system on World Wide Web. In C. Frasson, G. Gauthier, & A.Lesgold (Eds.), Third
International Conference on Intelligent Tutoring Systems, ITS-96 (Lecture Notes in
Computer Science, Vol. 1086), Berlin: Springer Verlag, pp. 261-269.
3. Brusilovsky, P. (2001) Adaptive Educational Hypermedia (Invited talk). In: Proceedings
of Tenth International PEG conference, Tampere, Finland, June 23-26, 2001, pp. 8-12.
4. Kay, J. (2000). Invited keynote: Stereotypes, student models and scrutability. In the
proceedings of Intelligent Tutoring Systems, pp19-30
Collaborative Radio Community
Paolo Avesani, Paolo Massa, Michele Nori, and Angelo Susi
ITC-IRST,
Via Sommarive 18 - Loc. Pantè, I-38050 Povo, Trento, Italy
{avesani,massa,nori,susi}@irst.itc.it
Abstract. Recommender systems have been usually designed to support a single

user in a one-to-one relation between a human and a service provider. This paper
presents a collaborative radio community where the system delivers a personal-
ization service on the fly, on the basis of the group recommending, promoting a
shift from the one-to-one approach to a one-to-group scenario where the goal is
assisting people in forming communities.
Keywords: recommender systems, web radio, learning preferences, multicast

streaming, virtual community.
1 Introduction
Usually, recommender systems have been designed to support a single user in a one-to-
one relation between a human and a service provider. Although advices are generated
on the basis of the opinions of other users, the system doesn’t support a direct relation
between two users that play the different roles of recommender and recommended.
Moreover, current personalization systems distinguish between the recommendation
step and the use step (i.e. listening in the case of music).
The new technological landscape concerning connectivity has been exploited to
conceive user centered services as Smart Radio [6], where an entire radio channel is
devoted to a single user to deliver a fully personalized program. Very often in the past
this emphasis on personalization has penalized the advantages that could arise from
the interaction of a community of users. More recently a new awareness is developing
that considers helping people to help each other a new challenge for recommender
systems [7]. In this perspective we promote a shift from a one-to-one approach to a one-
to-group-of-many scenario where the goal is assisting people in forming communities.
At the technological level the “group-of-many” can be managed with a peer-group
approach that is receiving growing attention as new protocols become available e.g.
JXTA [5].
In this paper we present a collaborative radio community where, taking advantage
of a low band multicast streaming, the system delivers a personalization service on the
fly devoted to group recommending. The users are involved both as listeners and as
recommenders. The live interaction allows the users to elicitate their disagreement on
the radio program. A different preference can be formulated providing alternative order
relation among the sountracks. This kind of preferences are closely related to the current
theme of radio program: in this context, a theme plays the role of a potential new category
of music that should inform the selection of the contents.
Collaborative Radio Community 463
Learning how to summarize partial order relations associated to a given theme into a
global preference model allows the detection of new kind of non standard categories of
music, i.e. the theme. This feature extends our previous work on CoCoA [1] a Compilation
Compiler Advisor based on case-based reasoning, that supports the detection of the
genres of use (from which the name CoCoA-Radio comes).
In the following we describe how CoCoA-Radio works and show a brief overview
of the learning issues that arise when a community of users has to be recommended.
2 The Application
CoCoA-Radio is a thematic interactive community radio. It streams continuously radio

programs of a fixed number of MP3s (usually 20); every radio program is relative to a
given theme. The radio program is the same for every user of the community and can
be thought as an hits list for the current theme where the songs at the top level can
be considered more representative of this theme. Users can interact with CoCoA-Radio
using a web browser expressing their preferences related to what they think it is the
hits list for the scheduled theme. In this way they play the main roles of authoring,
recommending and listening.
Let us introduce the main concepts of the radio. A theme is just a mnemonic label
that refers to a common target feeling about music or a perspective to look at the music;
themes can range from traditional “pop” or “women rock” to more fuzzy and undefined
“hands moving”. A playlist is a user defined list of songs to be submitted to the radio.
A program is the current list of songs that CoCoA-Radio is playing; it is relative to a
theme and is a synthesis of the submitted playlists.
Let us now describe in details about how CoCoA-Radio works. It streams music
programs for given themes continuously. The user can contribute to the radio program
related to the next theme by submitting a playlist as a proposal. The system is in charge
of summarizing a radio program taking into account the contributions of the users.
The challenge is to schedule a radio program that best fits the scheduled theme and
consequently best satisfies the listeners feeling.
It is important to stress that the goal of the radio is not to satisfy the user requests but
the user expectation concerning the music for the current theme. We could argue that
there are two mutual expectations. First, given a theme the users are constrained to submit
playlists related to this theme. Second, the users can give their feedback, specifying the
relative order of couples of songs in the syntesized program, when the program doesn’t
satisfy their expectation on the music related to the given theme.
It is a goal of the system to adaptively modify its schedule in order to satisfy the
users’ expectation and consequently to minimize the users’ feedback.
With CoCoA-Radio, what can be seen as a minus (the fact that every user gets the
same music) becomes a plus; infact users can benefit of the presence of other users
because they are all together concurring to form the hit lists for a theme. In essence
CoCoA-Radio is a social application because users have a common goal and they are
supposed to work in a cooperative manner to achieve it. Moreover, in a certain way users
exploit their reciprocal musical knowledge. Of course this means that there should be a
minimum agreement among users; but this could be overcome by replicating channels
and clustering users on them depending on their musical preferences; this poses another
interesting challenge.
464 P. Avesani et al.
3 Multicast Streaming
Another advantage of streaming a single line of music for a community is the possible
bandwidth saving. CoCoA-Radio is intended for (but not limited to!) Intranet LAN use
and, for this reason, we used multicast IP to stream music. In this way, our streamer does
not establish a connection with every client but sends a single stream of packets on the
net in multicast; thus every machine that needs to play the music gets this packet. One
can imagine the great bandwidth saving with, for example, 100 users! The protocol used
for streaming is RTP1 (short for Realtime Transport Protocol). It is an Internet protocol
specifically designed for transmitting realtime data such as audio and video. Typically,
RTP runs on top of the UDP protocol and so it allows multicast transmission.
4 Learning Issues
In the following we describe some interesting open learning issues in the radio domain. In
general users refer to music using categories that don’t belong to the standard taxonomy
of music genres and genres of use are not stable during the time, so a possible goal is
to recognize which tracks can be classified under a given theme and how much a track
is representative for it. Moreover the synthesis of the program starting from a collection
of submitted playlists is a satisfiability problem that usually does not have solution due
to possible unsolvable conflicts between users’ playlists; in the context of the radio
community it could be reasonable to avoid this problem detecting the reliability of the
users and assigning less priority to playlists submitted by the less reliable of them. Both
these aspects are analyzed in the learning algorithm included in CoCoA-Radio described
in the next section.
5 Radio Program Synthesis

The CoCoA-Radio domain can be described as a set of songs to be ordered called
S = {s1 , . . . , sN }, a set of radio listeners called U = {u1 , . . . , uM }, a set of themes,
i.e. music categories C = {c1 , . . . , cL }, and a ranking function R (possibly induced from
c
a playlist P) that is proposed by a user ui on a given theme cj , Ruji : S × S → [0, 1]
cj
, where Rui (s1 , s2 ) = 1 is a strong recommendation that s1 should be ranked above
s2 , 0 the opposite; a value closed to 1/2 is interpreted as an abstention from making
recommendation.
Starting from the user preference functions Rui can be derived a preference function
M
of the form: P REF (s1 , s2 ) = i=1 wui Rui (s1 , s2 ) where wui are weights assigned
to each listener ui , determining the reliability of the user ui respect to the given music
theme c. A learning algorithm [2,3,4] can be designed to update incrementally the weight
values.
Learning is assumed to take place in a sequence of rounds. On each round, we assume
the learning algorithm is provided with a set of S t of songs to be ranked. A combined
preference function P REF t is computed and than used to produce a total ordering
function ρtcj said Program of songs S t . After producing ρtcj the radio receives feedback
from the listeners. The feedback at time t, F t , is a set of assertions “song s1 should be
preferred to song s2 ” so F t is a set of pairs (s1 , s2 ). The feedback allows to compute
1
http://www.ietf.org/rfc/rfc1899.txt
Collaborative Radio Community 465
the function Loss(R, F ) that describes the loss of a preference function R given by a
user respect to the users’ feedback. The loss is the major component in the incremental
adequacy of the weights wui given by the learning algorithm to every radio user during
the learning procedure.
At the end of the procedure for each theme cj we have an order for the instances in
S called ρcj . This order is a sequence of instances. It is possible to map it in the Radio
domain calling it Program, a sequence of S elements.
The CoCoA-Radio interaction can be summarized in three phases: An initial phase
of playlists submission, given the set of possible songs in the database and a theme, users
can submit their own playlist (an ordered list of songs) related to the given theme; A
phase of synthesis of a program on the basis of the submitted playlists; A final phase of
users’ feedback acquisition, users can give information about the synthesized order in
terms of songs pairs < s1 , s2 >; the program will be then recomputed on the basis of
the received feedback.
6 Future Work
At the time of writing we have developed a first version of CoCoA-Radio using a J2EE
software platform. The system has been deployed on a LAN with a few hundreds hosts
and an archive of 6000 mp3 tracks.
Our primary goal in the short term is to test our architecture on the field with a
restricted and controlled community of users, i.e. the researchers of our institute. We
are interested in assessing on a real interactive environment both the technological and
the collaborative architectures. The main focus of our experiment is the analysis of the
learning curve at run time; our goal is to assess the relation between the feedback from
users and the learning process. In this context, not only accuracy is important but also
how fast is the convergence of the process.
A secondary goal is related to the acquisition of a meaningful dataset taking advan-
tage of the contribution of the real users. This dataset could be exploited to evaluate
alternative working hypothesis off line.
Currently, we have two the planned enhancements to extend the application: the first
is the opportunity for the user to submit a proposal for the next theme, the second is the
replication of the CoCoA-Radio with an added service of recommendation to help the
user to choose what kind of channel to subscribe.
References
1. S. Aguzzoli, P. Avesani, and P. Massa. Compositional CBR via Collaborative Filtering. In
ICCBR’01 Workshop on CBR in Electronic Commerce, Vancouver - Canada, 2001.
2. William W. Cohen, Robert E. Schapire, and Yoram Singer. Learning to Order Things. In
Michael I. Jordan, Michael J. Kearns, and Sara A. Solla, editors, Advances in Neural Infor-
mation Processing Systems, volume 10. The MIT Press, 1998.
3. Y. Freud, R. Iyer, R. Schapire, and Y. Singer. An Efficient Boosting Algorithm for Combining
Preferences, 1998.
4. Y. Freund and R. Schapire. A Short Introduction to Boosting, 1999.
5. L. Gong. Project JXTA: A Technology Overview, 2001.
6. C. Hayes and P. Cunningham. Smart radio: Building music radio on the fly, 2000.
7. Loren Terveen and Will Hill. HCI In The New Millenium, chapter Beyond Recommender
Systems: Helping People Help Each Other. Addison-Wesley, 2001.
Improving Interactivity in e-Learning Systems
with Multi-agent Architecture
Ricardo Azambuja Silveira1 and Rosa Maria Vicari2
1 Universidade Federal do Rio Grande do Sul – UFRGS

Universidade Luterana do Brasil - ULBRA
[email protected]
2 Universidade Federal do Rio Grande do Sul – UFRGS
[email protected]
Abstract. Over the last years, many organizations started to use Distance
Teaching tools as instruments in employees’ qualification programs, creating
what we may call E-learning or Virtual Training in Human Resources
Development Programs. However, usually these organizations tend to use
technological resources already available, and do not shape their technological
platform into a pedagogical project. Recent advances in the field of Intelligent
Teaching Systems have proposed the use of Artificial Intelligence through
architectures based on agents’ societies. Teaching systems based on Multi-
Agent architectures make possible to support the development of more
interactive and adaptable systems. The objective of the paper is to discuss the
feasibility of implementing Distributed Intelligent Learning Environment –
DILE based on the Multi-Agents Architecture approach, aiming at the
achievement of human resources qualification through Virtual Training.
Besides, we present a proposal of an architecture named JADE - Java Agent
Framework for Distance Learning Environments.
1 Introduction
Computer Science, together with Psychology and Education, has been trying to refine
teaching computational tools towards personalized self-learning. Every day, new
approaches to the use of Computer and Education are bringing new perspectives to
this area. The evolution of Computer and Education became computational teaching
environments an excellent choice for Distance Learning, by bringing new vigor to this
field of science.
The projects of E-learning must take into consideration that there are different
classes of students: the non-cooperative, those who act in a passive way or even try to
frustrate the program’s objective; the cooperative, who follow orientations, but do not
necessarily know where to go; and the pro-active students, who know very well their
objective, and search for aid to relief the task burden. The teaching methodology
employed in each case is different and there must have a clear concern by the
technological environment on the profile of the student that will use the system. In
order to reach this goal, cognitive student’s modeling is required, and it must make a
clear specification of him or hers.

Improving Interactivity in e-Learning Systems with Multi-agent Architecture 467
The state of the art in the Intelligent Tutoring Systems and Intelligent Learning
Environments fields points to the use of Agent Society-Based Architectures [4,5]. The
fundamentals of the Multi-Agent systems have demonstrated to be very appropriate to
design tutoring systems, since the teaching-learning problem could be handled as a
cooperative approach. That is why the Intelligent Learning Environments, such as
JADE, are a class of teaching instruments much more advanced in the pedagogical
and organizational point of view, more adequate to the aims of Virtual Training in
organizations.
2 The JADE Project
The Java Agent framework for Distance learning Environments – JADE project [5,6]
proposes an infrastructure of project, development and implementation of Distribute
Intelligent Learning Environments – DILE, based on the approach of Multi-Agents
architecture towards Distance Education, for multiple domains. In this project we
implemented different versions of Eletrotutor prototype. Eletrotutor is a teaching
environment for Electrodynamics teaching, and in each version we refined JADE
architecture.
The environment we proposed contains a special agent responsible for each
teaching strategy developed, that is, for the domain knowledge retrieval over each
point to be presented to the student, for the task of proposing exercises and evaluating
proposals, examples and extra activities.
JADE architecture encompasses, therefore, a Multi-Agent environment composed
of an agent responsible for the system general control (Student’s Model), and a
Communication Manager and other agents (Pedagogical Agents), which are
responsible for tasks related to their teaching tactics, where each agent may have its
tasks specified according to its goal. All actions of student’s data accessing are taken
by the Student’s Model, thus when a pedagogical agent is required to update the
student’s historic, this agent sends to the Student Model data to be updated, as well as
any other change in the student’s state of teaching (see Fig. 1).
Communication between agents happens through a definition of a KQML-based
message set, implemented through communication resources of JAVA language
objects named RMI (Remote Method Invocation), used in the project [1].
The agent’s architecture is designed as robust and standardized as possible and that
enables reusing codes for different kinds of agents. The tasks performed in teaching
are decomposed and performed individually or in groups of agents. How the task will
be decomposed is defined by the content of messages exchanged between agents.
The cycle of agents’ execution, shown in Fig. 2, consists of the following steps:
• New messages processing: the task is decomposed;
• Determination of which rules are suitable in the current situation: analysis
of task and if necessary delegation of other agent(s) task;
• Execution of actions specified for such rules: task execution;
• Mental state update according to those rules: management of knowledge
about the world;
• Planning: module that must develop plans that reach goals specified by
agents intentions.
468 R. Azambuja Silveira and R.M. Vicari
Fig. 1. System Architecture: The Architecture of JADE system is composed of a set of agents:
(Pedagogic Agent) in charge of performing learning activities as examples, exercises, an others.
One special agent (Communication Agent) performs communication management among the
agents. There is an agent (Student Model Agent) responsible for student modeling and agents’
coordination. The Browser component (Remote Agent) performs the student interface and the
communication between the student and the system.
A g en t C y cle
B eliefs
N ew m essag e
M enDsa
esir
gemes
p rocessin g K n ow led g e
In ten tion base
B eh a vior al C og n itive
r ules settin g m od el
S k ills
P ar ticu lar an d
com m u n ication
action s
M en ta l m od el
u pd atin g
E xter n al
P la n n in g en viron m en t
Fig. 2. The agents’ cycle performs messages sending and receiving and specific task of the
agent, according to the knowledge base. As the agent receives a new KQML message it
processes the message according to its content, Appling the adequate behavioral rule.
According to this rules the message-receiving event can trigger some message sending, mental
model updating and some particular specific agent action.
2.1 The ELETROTUTOR Prototype
The Eletrotutor prototype was implemented as a test bed to evaluate JADE platform.
It is an Electrodynamics client-server intelligent learning environment designed
according to JADE architecture (available in http://www.inf.ufrgs.br/~rsilv)
Figures 3 and 4 show two snapshots of The Eletrotutor prototype:
Fig. 3. Eletrotutor’s snapshot 1 shows the main screen. The first button (Tutor) changes from
the autonomous mode to the tutorial mode. The second (Lições) invokes the lessons menu. The
third (Exercícios) invokes the exercises menu. The fourth (Exemplos) invokes the examples
menu. The fifth (Ajuda) call the help system. The sixth (Configurar) seventh (Sem figures) and
eighth (parar) change several interface configurations
The environment may be used in Tutorial, and Autonomous modes. In the first
mode, the student has total control over the study session, and may perform any
lesson, check any example or make any exercise. In the Tutorial mode, the system
undertakes the session control, defining the sequence of lessons, examples, and
exercises. For that end, the tutor makes use of a student’s cognitive diagnostic, taken
through the record of every action the student takes. Thus, teaching strategies observe
the student’s historic before taking the next actions. Teaching strategies are the
sequence of contents, examples and exercises that will be proposed to the student.
3 Conclusions
Distance Education systems based on the Internet does not have any time or space
constraint. Students can interact with the system anytime, anywhere. However, most
of Distance Education systems based on the web are not intelligent or adaptable.
470 R. Azambuja Silveira and R.M. Vicari
Researches have turned towards the use of adaptive www by using systems based on
Intelligent Tutors and Intelligent Agents architectures. All these issues have in
common a strong dependence on a sharp and robust student modeling. Through the
student model it is possible to provide customized teaching tactics, which reflect the
knowledge level of each student, his/her learning abilities and objectives. In this work
we intended to bring some important contributions, refining the efficacy of learning
environments and stressing the use of cooperative solving problem paradigm using
Multi-agent architecture. Further work will integrate the JADE implementation of
pedagogical agents with commercial or well-known academic learning environments
or frameworks and improve some issues like Agent Communication and Negotiation.
Fig. 4. Eletrotutor’s snapshot 2 shows an exercise. The system presents as many exercises as
the student want by clicking (Novo Exercício) button. This changes the instance of this kind of
exercise. By clicking the buttons (Tela1, Tela2, Tela3) the student invokes different kinds of
exercises for this lesson
References
1. BICA, Francine, Eletrotutor III: Uma abordagem multiagente para o Ensino a distância.
Porto Alegre: CPGCC da UFRGS, 1999. Master Dissertation.
2. CHEIKES, B. A. GIA: An Agent Based Architecture for Intelligent Tutoring Systems. In:
THE CIKM WORKSHOP ON INTELLIGENT INFORMATION AGENTS, 1995.
Proceedings, 1995
3. MATHOFF, J.;VAN HOE, R. APEALL: A Multi-agent approach to interactive learning
environments. In: EUROPEAN WORKSHOP ON MODELING AUTONOMOUS
AGENTS MAAMAW, 6., 1994. Proceedings Berlin: Springer-Verlag, 1996.
4. SILVEIRA, Ricardo Azambuja. Modelagem Orientada a Agentes Aplicada a Ambientes

Inteligentes Distribuídos de Ensino – JADE - Java Agent framework for Distance learning
Environments Porto Alegre: PPGC da UFRGS, 2000. Doctoral Thesis.
5. SILVEIRA, Ricardo Azambuja, VICARI, Rosa Maria. JADE - Java Agents for Distance
Education framework. In: DEC 2001, 2001, Austin. DEC 2001. CD-ROM, 2001
Adaptive Hypermedia Made Simple
with HTML/XML Style Sheet Selectors
François Bry and Michael Kraus
Institute for Computer Science, University of Munich

http://www.pms.informatik.uni-muenchen.de/
Abstract. A simple extension is proposed for enhancing HTML and

XML with adaptation. It consists in using the path selectors of style
sheet languages such as CSS and XSLT for expressing content and nav-
igation adaptation. The needed extensions to a path selector language
are minimal, a few additional constructs suffice. The processor of the
language can be kept almost unchanged, no new algorithms are needed.
Furthermore, it is proposed to use XML for expressing user model data
like browsing history, browsing environment (such as device, time, etc.),
and application data (such as user performances on exercises).
1 Introduction
In existing systems, extending HTML and XML with simple adaptive hyper-
media functionalities is done using a combination of cookies, ie client-side user
identification, server-side scripting languages like PHP [6], and URIs. This has
several drawbacks. Information about the user has to be stored and processed
on the server. Due to the nature of the Web’s HTTP protocol, this information
is limited as compared to the information (possibly) available on the client side:
For example, it is not possible to track navigation using the back and forward
buttons, navigation in different windows, or navigation on more than one server.
This prevents implementing non-trivial adaptive hypermedia systems.
In contrast, the approach outlined in this paper does not suffer from the
above-mentioned drawbacks, as it works on the client side. In common adaptive
hypermedia systems, the structure of the information, the information itself, and
the way of information acquisition together form a user model [2]. This paper
does not propose a specific user model, but a framework relying upon HTML and
XML that allows a simple implementation of user models. The main advantage
of this framework is to make adaptive hypermedia techniques available in the
Web context at low cost, ie with minimal changes of the existing standards.
This framework is described in more detail in [3]. It has been proposed to
the W3C Device Independence Working Group [4].

Adaptive Hypermedia Made Simple with HTML/XML Style Sheet Selectors 473
2 Browsing Context:
A Data Structure for Expressing User Models
HTML and XML have no means to express a user model. Therefore, a data
structure called browsing context is proposed [3], which allows a user model to
be stored by the browser, ie on the client side, to be accessed through style
sheets, and to be updated through Web applications using scripting languages
like Javascript [5]. These features make the data structure “browsing context”
convenient for an adaptive presentation of Web pages.
A browsing context consists of three components that can be distinguished
according to data acquisition: browsing history, browsing environment and ap-
plication data. “Browsing history” data are informations about the browsing
actions performed by the user in the past such as visiting Web pages, traversing
hyperlinks, opening and closing windows, etc. This information is automati-
cally generated by the browser and it is updated each time the user performs a
browsing action. “Browsing environment” data are informations about the de-
vice (hardware), browser (software), location, time, language, etc. Like browsing
history data, this information is automatically generated and updated by the
browser. “Application data” are informations specific to the Web application
being browsed by the user. In the case of an electronic tutor system, this can be
the user performances on exercises, like the numbers of correct answers.
Using style sheets and scripting languages in conjunction with a “browsing
context” offers the possibility to easily implement an adaptive hypermedia sys-
tem cf. [3]. For accessing the “browsing context” with style sheets and scripting
languages in a convenient manner, it is preferable to store it in XML format, eg
as proposed in [3]. Web browsers store an internal representation of the docu-
ment currently being displayed, eg as a DOM [1] tree. This document is referred
to in the following as naked document because it does not contain any browsing
context information. In a similar way as this naked document is stored by the
browser, a browsing context document [3] can also be stored by the browser.
Both the “naked document” and the “browsing context document” can be con-
sidered as the two parts of one (virtual) context enriched document stored within
the browser. The “context enriched document” takes over the role of the orig-
inal “naked document” within the browser, ie style sheets are applied to the
“context enriched document” instead of the “naked document”, scripting lan-
guages have access to the DOM tree of the “context enriched document” instead
of the DOM tree of the “naked document”, etc. Thus, the “context enriched
document” is a virtual document combining a “browsing context” (using which
adaptation is expressed) with a standard HTML or XML document. Note that
the materialization of this virtual document is not needed.
3 Implementing Adaptation Using Style Sheet Selectors
A simple extension to style sheet selectors makes it simple to implement adaptive

hypermedia functionalities with HTML and XML. The path expression of a style
474 F. Bry and M. Kraus
sheet selector is not to be matched against the original “naked document” tree,
but against the new context enriched document tree.
Typical Web style sheet languages like CSS and XSLT have constructs of two
kinds: style rules and selectors. Style rules define certain presentation parameters
for elements in the document tree (like fonts, colors and margins), and transfor-
mations of the document tree (like insertion and sorting of elements). Selectors
are path expressions that determine which style rule is applied to which element
in the document tree. Matching a path expression of a style sheet selector not
against the original “naked document” tree, but against the context enriched
document tree makes it possible to build path expressions that depend on the
content and structure of both, the “naked document” and the “browsing con-
text document”. Note that if the path expression of a selector contains no parts
referring to a browsing context, the semantics of a style rule remains unchanged.
Examples are given in [3].
4 Possible Extensions
Updating Application Data using Scripting Languages. Using style sheet selectors
to express content and navigation adaption is not sufficient for modeling certain
complex aspects of adaptive hypermedia systems. Still missing is the possibility
to store data in the “browsing context”, which then could be used by style
sheets as a source of adaptation. Scripting languages like Javascript can be used
to achieve this. In a similar way as Javascript code contained in Web pages can
change the (“naked”) document tree, Javascript code contained in Web pages
can change the content of a “browsing context”’s application data.
Modeling Locations. There are several different notions of location. (1) Loca-
tions can be informations about the country or region where the user is, like
Germany or France. This information is available in desktop computer systems
and does not change during a browsing session. (2) Locations can be informa-
tions about the geographical position of the user, expressed, eg as longitude and
latitude. This information is available in mobile devices like cellular phones or
PDAs with special positioning equipment, eg a GPS device. Geographical loca-
tion information can change during a browsing session. (3) Locations can also
be informations about virtual locations like home, car, office, meeting, etc. Infor-
mations about virtual locations can change during a browsing session. “Virtual
locations” are represented neither in current computer devices, nor within cur-
rent Web standards. All of these notions can be represented simultaneously as
browsing environment data in a “browsing context”.
5 Discussion and Concluding Remarks

The approach outlined here has both, advantages and limitations. First, the ap-
proach is quite simple. It introduces a wide range of adaptation features into
existing HTML and XML standards at the cost of very limited extensions to
Adaptive Hypermedia Made Simple with HTML/XML Style Sheet Selectors 475
these standards. The extensions to these standards are as follows: (1) Infor-
mation like browsing history and browsing environment data, most of which
is already stored by conventional browsers, is to be stored as a standardized
“browsing context” in an internal XML representation like DOM. (2) The style
sheet processor(s), eg those of CSS or XSLT, match the selector part of a style
rule not against the original “naked document”, but against the (virtual) “con-
text enriched document” (consisting of the “naked document” enriched with a
browsing context). The style sheet processor must recognize those selector com-
ponents referring to the “naked document” and those referring to the browsing
context. This is conveniently achieved using namespaces.
Apart from these, no further changes are needed, especially, no new algo-
rithms are needed. Only the processing of style sheet rules is extended, the
style sheet languages remain otherwise unchanged (because of the use of names-
paces cf. [3] section 3). This ensures upward compatibility with already existing
style sheets. Also, style sheets that make use of “browsing context” selectors are
downwards compatible with non-browsing context enabled browsers. With such
browsers the data can be accessed, only the adaptation features are missing.
Upwards and downwards compatibilities are essential for extensions to existing
Web standards. Thus, the approach proposed in this paper is a conservative
extension of the already existing and well-established Web standards.
The approach outlined in this paper is not specific to CSS or XSLT. It relies
only on path selectors, which play a central role in Web standards. The same
approach can easily be applied to other or future style sheet languages or to
other Web standards like XML query languages, as long as they build on path
selectors. Note also that this approach is stable against the changes from XPath
1.0 to XPath 2.0, which have introduced a considerably more complex type
system, a set of relational operators, and (certain kinds of) variables.
References
1. V. Apparao et al. Document Object Model (DOM) Level 1 Specification Version
1.0. W3C Recommendation, 1998. http://www.w3.org/TR/REC-DOM-Level-1 .
2. P. Brusilovsky. Methods and Techniques of Adaptive Hypermedia. User Modeling
and User-Adapted Interaction, 6(2-3):87–129, 1996.
http://www.contrib.andrew.cmu.edu/ plb/UMUAI.ps .
3. F. Bry and M. Kraus. Adaptive Hypermedia Made Simple Using HTML/XML
Style Sheet Selectors. Technical report, Inst. for Computer Science, University
of Munich, 2002. Full version of this paper. http://www.pms.informatik.uni-
muenchen.de/publikationen/#PMS-FB-2002-1 .
4. F. Bry and M. Kraus. Style Sheets for Context Adaptation. W3C Work-
shop on Delivery Context, 2002. http://www.pms.informatik.uni-muenchen.de/
publikationen/#PMS-FB-2002-3 .
5. Standard ECMA-262. ECMAScript Language Specification, 1999. ftp://ftp.ecma.
ch/ecma-st/Ecma-262.pdf .
6. PHP - Hypertext Preprocessor. http://www.php.net/ .
A Framework
for Educational Adaptive Hypermedia Applications
Félix Buendía1 and Paloma Díaz2
1Escuela Universitaria de Informática. Universidad Politécnica de Valencia.

46022-Valencia (Spain)
[email protected]
2Departamento de Informática.Universidad Carlos III de Madrid
28911 Leganes (Spain)
[email protected]
Abstract. Hypermedia technology has a large spread in the instructional field,

specially in Web-based educational environments. Adaptive hypermedia sys-
tems have been focused on such aspects as specifying user models to adapt the
educational contents and the navigation structure providing learning paths. The
work presented in this paper addresses a new aspect based on adapting the ac-
cess to the didactic structures, that encapsulate educational contents and learn-
ing paths, to support useful educational adaptive applications. It is presented a
framework to support the design process which is divided in two main parts:
the first one is focused on instructional issues such as the organization of edu-
cational contents and the learning tasks to access them; and the second one is
based on the translation of these instructional entities into hypermedia compo-
nents, for which the Labyrinth hypermedia model is used.
1 Introduction
Hypermedia technology has a large spread in the instructional field, specially in Web-
based educational environments. In this context, adaptive hypermedia systems (AHS)
have been focused on adapting the educational contents and the links that provide
learning paths but few attention has been paid to structural aspects dealing with di-
dactic requirements of instructional applications. Indeed, most AHSs mainly focus on
displaying sequences of contents in user adapted ways but they are not aware of a key
aspect in instructional applications: the structures in which educational contents are
organized to reach a specific learning goal or the activities that a student can perform
on these structures. This approach is supported by the thesis of authors like Merrill 1
who suggests that adequate instruction would require multiple types of knowledge
structures (KS) to be identified and made explicit to the learner. In 2 there is a pro-
posal to use these KS entities and extend them, becoming Didactic Structures (DS)
able to capture didactic relationships between the Instructional Objects (IO) used to
transmit some piece of knowledge. Another important aspect to be considered with
adaptation purposes is how learners access DS entities by means of Instructional

A Framework for Educational Adaptive Hypermedia Applications 477
Tasks (IT). In this paper we define a framework to deal both with instructional design
topics and, at the same time, with the design of computer-based applications. This
framework is based on the proposal presented in 2 which considers a triple model:
Domain, User and Adaptation. The remainder of the paper is organized as follows.
Section 2 describes some related work. Current definition of the Domain and User
model, and their translation to Labyrinth entities 4 are described in sect. 3. Section 4
goes deeper into the Adaptation model and, finally, section 5 presents some remark-
ing conclusions.
2 Related Work
Adaptive technologies have been traditionally applied in Web-based education in two

main areas: the adaptive presentation and the navigation support. The references to
structural aspects are restricted to rigid proposals such as the ‘curriculum sequencing’
technology 5. It is used in 6 where the adaptive functionality is based on following
the internal structure of the subject domain. A more advanced sequencing technique
is found in DCG 7 which allows the dynamic adaptation of Web-based courses. The
main idea of the DCG architecture is the explicit representation of the concept struc-
ture of the domain, separated from the teaching materials. Nevertheless, it is not in-
volved in the adaptation of the access to the underlying structures.
Moreover, proposals to develop adaptive courseware such as Interbook 8 or Net-
Coach 9 are mostly based on traditional course structures. For instance, Interbook is
based on the ‘electronic textbook’ metaphor and a sequence of learning goals as the
navigation guide but it does not include references to other internal structures that
could guide the learning process. More advanced techniques of knowledge-based
navigation support are described in KBNS 9 and in HyperTutor 11 but they do not
consider structure-based adaptation.
3 Domain and User Model
According to 2 an adaptive application uses a Domain model, describing how the

information content of the application is structured. In this case, it is based on a Di-
dactic model defined in 2 to design instructional applications. This model provides
static entities, the Instructional Objects (IO) and the Didactic Structures (DS), that
embody didactic information used to teach a specific topic. The functional model
entities are Instructional Tasks (IT) and Learning Scenarios (LS). Each DS has a set
of IT entities that represent its interface and they are dependant on the structure type.
For example, the instructor can define tasks such as Access or Identify. The first one
consists of doing a trip around the hierarchy tree at a given level (e.g. Europe states).
The Identify task requires a higher level because it involves setting up the relationship
between two o more hierarchy concepts (e.g. the geographical situation of a state).
The user must have a certain competency level (CL) before accessing an IT. An IT
also returns a competency value when the task is finished. LS represent the terms and
conditions that characterize the user learning. The adaptive functionality is focused on
478 F. Buendía and P. Díaz
the kinds of structures the users will be able to work with and the tasks they will exe-
cute on these structures. In order to ease their management from a computer-based
point of view, the Didactic model entities are specified using a Labyrinth notation.
This translation is described as follows (see Fig. 1). An instructional application is
assigned with an Instructional Hyperdocument whose student user set is initially
empty. It contains a group of DSs which are assigned to Labyrinth nodes. Each of
these nodes has an anchor list which locate the associated ITs. These anchors repre-
sent the DS interface and they are connected using links with User Hyperdocuments.
Each DS has a set of IO entities which contain knowledge objects represented as
Labyrinth contents.
Instructional hyperdocument
Didactic structures
Instructional instructional
User hyperdocument tasks objects
Root
Learning
scenarios
User profile Labyrinth Anchors

entities Links
Data user
Nodes
Contents
Fig. 1. Labyrinth representation of the Domain and User model entities.
4 Adaptation Model
Adaptation is based on the domain model (that is, DS and IT entities) and the user
model (that is, LS and CL entities). With this purpose the instructor must provide an
adaptation model consisting of rules enabling the user access to a DS given a certain
CL. These rules can be further specified by means of weighted graphs or a similar
notation (e.g. Petri nets) to represent the transitions between the different IT entities.
In the current case, the ability to include event-based specifications for each compo-
nent of a hypermedia application supported by Labyrinth [4] is used to specify the
conditions that activate a new IT as the next example shows:
Event with condition (‘LS_by_examples’ node) AND (‘Access_ task’ completeTaskAttribute)
triggers action [create_link_with (‘Identify_task’ node) in (‘Europe_parts’ node) ]
in which, a new link is set between a ‘learning by examples’ scenario and an Identify
task if an Access task has been already completed. Several events can be defined in an
analogous way and they can be assigned with different priority levels using the Laby-
rinth event value function. A given link is triggered (navigated) if it has a play event
whose priority allows it to be scheduled by the adaptation engine.
A Framework for Educational Adaptive Hypermedia Applications 479
5 Conclusions
This work presents a framework for designing educational hypermedia applications

that intends to set a bridge between instructional and hypermedia aspects. On the one
hand, it provides a model to specify Didactic Structures that eases the management
and access to educational contents. A XML-based notation is being developed to help
the instructor in such specification. On the other hand, the use of the Labyrinth hy-
permedia model permits the translation of the instructional entities to computer-based
specifications. This feature will ease the implementation and validation of educational
applications in hypermedia environments like the Web.
References
1. Merrill M. D., & ID2 Research Team. Instructional Transaction Theory: Instructional
Design based on Knowledge Objects. Educational Technology, 36(3), 1996 pp. 30-37.
2. Buendía, F.; Diaz, P. Benlloch J.V;, A Framework for the Instructional Design of Multi-
Structured Educational Applications, ED-MEDIA 2002, World Conference on Educa-
tional Multimedia, Hypermedia & Telecommunications, Denver, USA, 2002.
3. Wu, H., De Kort, E., De Bra, P.,Design Issues for General-Purpose Adaptive Hypermedia
Systems. Proceedings of the ACM Conference on Hypertext and Hypermedia, , Aarhus,
Denmark, August 2001, pp. 141-150.
4. P.Díaz, I. Aedo and F. Panetsos, Modeling the dynamic behavior of hypermedia applica-
tions. IEEE Transactions on Software Engineering vol 27 (6), June 2001, pp. 550-572.
5. Brusilovsky, P. Adaptive and Intelligent Technologies for Web-based Education. In C.
Rollinger and C. Peylo (eds.), Special Issue on Intelligent Systems and Teleteaching, Kün-
stliche Intelligenz, 4, 1999 pp.19-25.
6. Maria Grigoriadou, Kyparisia Papanikolaou, Harry Komilakis and George Magoulas
INSPIRE: An INtelligent System for Personalized Instruction in a Remote Environment.
Third Workshop on Adaptive Hypertext and Hypermedia, Sonthofen, Germany, July 13-
17, 2001
7. J. Vassileva, Dynamic Courseware Generation on the WWW, Proceedings of the work-
shop ‘Adaptive Systems and User Modeling on the World Wide Web’, Sixth Interna-
tional Conference on User Modeling, Chia Laguna, Sardinia, 2-5 June 1997.
8. Brusilovsky, P., Eklund, J., and Schwarz, E. (1998) Web-based education for all: A tool
for developing adaptive courseware. Computer Networks and ISDN Systems (Proceedings
of Seventh International World Wide Web Conference, 14-18 April 1998,, pp.291-300.
9. Gerhard Weber, Hans-Christian Kuhl and Stephan Weibelzahl Developing Adaptive
Internet Based Courses with the Authoring System NetCoach Third Workshop on Adap-
tive Hypertext and Hypermedia, Sonthofen, Germany, July 13-17, 2001
10. Eklund, J. and Sawers, J., Customising Web-based course delivery in WEST® with navi-
gation support, in: Proceedings of WebNet’96, World Conference of the Web Society,
San Francisco, CA, October 15–19, 1996, pp. 534-535.
11. Pérez, T., Lopistéguy, P., Gutiérrez, J., and Usandizaga, I. (1995b). HyperTutor: From
hypermedia to intelligent adaptive hypermedia. In Maurer, H., ed., Proceedings of ED-
MEDIA’95, World Conference on Educational Multimedia and Hypermedia. Graz, Aus-
tria: AACE. pp.529-534.
METIORE: A Publications Reference
for the Adaptive Hypermedia Community
David Bueno1, Ricardo Conejo1, Cristina Carmona1, and Amos A. David2

1 Department of Languages and Computer Science, University of Málaga,
29071, Málaga, Spain.
{bueno,conejo,cristina}@lcc.uma.es
2 LORIA, BP 239, 54506 Vandoeuvre, France
[email protected]
Abstract. The Web is one of the most powerful sources of information on any
topic. However looking for scientific literature is a difficult task. Prior
knowledge of link sites is necessary and if you are lucky they point to
conferences proceedings available on-line. In fact the case the user is not able to
make queries about the available documents and must check them one by one
using general purpose search engines. In this paper we propose our system
METIORE as a source of information for the Adaptive Hypermedia
community. The idea is to put together all the publications on this research area
and provide an adaptive tool to find papers or people working in the field.
METIORE is a Personalized Information Retrieval system that keeps a user
model based on objectives.
1 Introduction
The medium most used by the research community to find relevant work is the Web.
The search engines such as Altavista or Google are improving their way of retrieving
information. These engines index Web documents, but many of the scientific
publications are in postcript or pdf formats which are not indexed by the engines.
Also, because of the large quantity of Web pages they index, many of the results that
can be obtained are not relevant. It is therefore necessary to provide information
resources and associated tools where researchers can just find scientific contents
without the usual noise introduced by typical search engines.
The idea of having a search engine exclusively for scientific information has
already been developed. FermiVista [3] is an index of articles and prepublications
available on the Web or FTP servers of Universities and Research Centers. It contains
documents in the domain of mathematics, computer science and physics.
Unfortunately this index has not been updated since 1998.
One of the best specific indexes is the CiteSeer [4]. It locates scientific articles on
the web using search engines like AltaVista and allows for the extraction of
information of the citations in publications. The main feature is the creation of an
Autonomous Citation Indexing [5]. Even though this system can be applied to any
field of research, it is specific for Computer Science literature.

METIORE: A Publications Reference for the Adaptive Hypermedia Community 481
In this paper we present the application of our Retrieval System METIORE

oriented to the implementation of an index/search engine for the Adaptive
Hypermedia (AH) publications. In Sect. 2 the generalities of METIORE will be
presented. The adaptation to the Web to make the AH index is explained in Sect. 3.
Finally the conclusions and future work are presented in Sect. 4.
2 METIORE
METIORE is the acronym of Multimedia coopErative InformaTIOn Retrieval

SystEm. It is a multipurpose IRS that can be applied to different kinds of database. It
has been used in three different areas. For the first area, METIORE_STREEMS is an
IRS for managing multimedia information on trees authorized for reforestation by the
European Union (EU). The project was sponsored under the EU project
LEONARDO. For the second area, METIORE_LORIA is used for managing the
database of publications of the computer science laboratory research center, LORIA,
Nancy France. Some evaluations of METIORE have been done using this database
[1]. In the third area, METIORE_REVUES is used for the access and analysis of a
collection of journals called ‘Relations Publiques Informations’.
The ability of METIORE to manage different types of data is in part due to the use
of XML to import the different databases. From these sources it generates files with
inverted lists and cluster information to provide a fast access. The core of METIORE
is independent from the data; it only needs to know the parameters that will be used
for the personalization. So if it is working with publications it may use author, title,
keywords, etc. But if it is working with data on trees, the parameters may be height,
altitude, scientific name, etc.
METIORE has a powerful search interface that allows the user to make simple or
complex queries and also perform global analysis of the database. It is also possible to
explore the database without knowing anything about its content. This last possibility
is very useful as a starting point in the process of information retrieval. The following
are examples of the types of query that can be processed using METIORE on the
database of publications:
- Documents about Script languages published after 1998
- List of all authors with their number of publications
- Authors who have written with X and how many times (this shows people working
together)
- Names of authors and keywords that they use
- The yearly frequency of the use of the keyword ‘expert systems’ (this provides a
means of observing the evolution of a research topic in terms its frequency of use
in publications)
The most relevant feature of METIORE is the personalization of the response. Our
approach on personalization is based on the concept of objective. For us the objective
is the expression of the user’s information needs formulated in natural language. The
objective is currently used to group the set of queries, concepts and decisions that the
user makes on the system having his objective in mind. Our hypothesis is that
grouping the user’s interactions into objectives will help the user to find information
482 D. Bueno et al.
in his history and help the system to build a specific model of the user from one or
more sessions.
This concept is very important to us since we believe that every user has a
minimum knowledge of his information needs before attempting to use an IRS. This
knowledge can of course be improved with the use of the system and consequently
the user’s ability to express his information needs through the system’s interaction.
For us, the queries do not necessarily express all the information needs of the user but
rather his approach towards solving the problem of his information needs.
For each user objective the system constructs a model. In order to give users a
personalized response, after querying the system, they should evaluate some
documents of the response. This provides the system with the user’s preferences for
this objective and thus help the user in future queries to have the most relevant
solutions first. This is different from many retrieval systems that gives a list of
solutions (for example 400 documents) that are sorted only according to the user’s
query. METIORE sorts the solutions using the query and the user model in order to
select the best solutions that are presented at the top of the list.
Another interesting feature of METIORE is that the user has available his history
of evaluated documents. This history is organized by objectives, for each of which the
evaluated documents are sorted by their relevance to the user’s objective. With the
history the user can review his past solutions, modify his evaluations and also look for
related documents.
3 METIORE on the Web
To make METIORE more accessible, it has been necessary to make it available on the
Web. The best option has been to keep the core of METIORE to maintain all its
functionality and used as a server. The user will have a web interface that is
connected to a Java Servlet which will send transactions to the METIORE daemon
through a socket connection. A specific protocol has been created for the
communication between the applications. The main programming language of
METIORE is incrTcl. Fig. 1 shows the architecture for METIORE on Web.
Serv er
User
W eb C lient
Profile
AH WWW M E T IO R E
M E T IO R E Servlet C lient
daem on
AH Sock et
D atabase connection
Fig. 1. The architecture to make METIORE accessible through Web
For Adaptive Hypermedia literature and in general for scientific publications on

the Web we have designed a database that contains all the information needed to
represent a publication (including its url). It also includes the information on authors
and conferences/journals with links to their homepages. This makes METIORE_AH a
METIORE: A Publications Reference for the Adaptive Hypermedia Community 483
source of data where papers, conferences, editorials and people can be easily found
and contacted.
All the publications of the Adaptive Hypermedia Home Page [2] and related
conferences are included in the database. We have realized that most of the papers
don’t have associated keywords. In order to obtain the keywords (more precisely the
root of keywords) and have them as a source of information in queries, the full
documents are treated in the following way: The paper is parsed using a stop words
list and the porter algorithm [6]. Then the frequency of words is obtained and the
most relevant are chosen to represent the article. A web page is also provided to let
researchers include their publications in METIORE. When all these documents are
inserted, it will be possible to use the characteristics of METIORE to obtain relevant
publications for the users through the Web.
In this paper we have presented the application of METIORE for managing the
Adaptive Hypermedia publications on the Web. We hope to integrate some other
functions that are available in similar systems such as CiteSeer. One of the functions
is the processing of the references to have some kind of citation index, or to
automatically include papers to the database. However, METIORE offers other
possibilities that can’t be found in other applications, like the complex data analysis,
the objectives of retrieval that can be maintained in different sessions, or the active
history to review relevant publications. Others features such as having the list of the
most relevant articles or authors according to the interactions of users with METIORE
will be available soon. With the data obtained through the user’s interactions, new
evaluations of METIORE will be made. Our objective is to apply METIORE to other
areas of research after evaluating this application.
References
1. Bueno, D., & David, A. A. (2001). ‘METIORE: A Personalized Information Retrieval

System ‘. User Modeling 2001
2. De Bra, P. (1999) Adaptive Hypertext and Hypermedia [Web Page].
http://wwwis.win.tue.nl/ah/.
3. Fermigier, S. (1996) FermiVista [Web Page]. http://fermivista.math.jussieu.fr/.
4. Lawrence, S. (1997) CiteSeer Research Index [Web Page]. http://citeseer.nj.nec.com/.
5. Lawrence, S., Lee Giles, C., & Bollacker, K. ‘Digital Libraries and Autonomous Citation
Indexing’. IEEE Computer, 32(6), 67-71
6. Porter, M.F., (1980). ‘An algorithm for suffix stripping’, Program, 14 (3),130-137
User-Driven Adaptation of e-Documents
1 2 3 2
P. Carrara , D. Fogli , G. Fresta , and P. Mussio
1
Istituto per le Tecnologie Informatiche Multimediali del CNR, Milan, Italy
[email protected]
2
Dipartimento di Elettronica per l’Automazione, Università di Brescia, Brescia, Italy
{fogli,mussio}@ing.unibs.it
3
Istituto di Scienza e Tecnologie dell'Informazione "A. Faedo” del CNR, Pisa, Italy
[email protected]
Abstract. This paper proposes a new strategy for designing e-documents adapt-
able to user’s aims and habits. This strategy is based on the creation of a cas-
cade of interaction environments requiring the local adaptation of the e-
documents. The strategy is supported by a recently introduced model of Human
Computer Interaction, and made effective by the development of BANCO, an
environment allowing Web modification by the users at the client side.
1 Introduction
The increasing diffusion of end-user computing based on the interaction with hyper-
media systems and applications poses several new problems: the environment with
which users are interacting is a new type of document, an electronic document (e-
document), which evolves the traditional documentation styles, opening to their users
new communication possibilities as well as creating new dangerous situations. E-
documents are less persistent than traditional ones, because they only exist and are
perceivable if an electronic machinery maintains them in existence. They can be man-
aged and adapted more easily than paper ones; they can be active and reactive to
users’ requests; they appear to the users as a whole, but they can result from merging
of different elements which may be distributed in different, geographically remote
repositories. This paper reports on a strategy for designing e-documents adaptable to
user’s aims and habits [1], based on the user generation of a cascade of environments.
The strategy is based on a recently introduced model of Human Computer Interaction
[2] and has been made effective by the development of BANCO (Browsing Adaptive
Network for Changing user Operativity) [3], a feasibility prototype for the Web, de-
signed to explore the interaction with e-documents.
2 Bootstrapping Workshops for e-Documents Adaptation

In usual practice, designers develop the tools to manage e-documents focusing on the
computational and management aspects and not on the solution of problems. This

User-Driven Adaptation of e-Documents 485
design approach forces end-users to translate their problems and solutions in com-
puter oriented notations, generally not amenable to their reasoning, but often mislead-
ing. The approach we present reverses this situation: the designer has to produce an e-
document in a form understandable by the users, prone to their habits and adaptable
to the user situation. To this aim, it is important to distinguish the various kinds of
users and their roles. In general, users belong to specific user communities, playing
distinct roles: Figures 1 and 2, for example, show e-documents designed for an appli-
cation aimed at monitoring glacier evolution. In Fig. 1 the intended user is a photo
interpreter, while in Fig. 2 the e-document should be used by environmental agency
clerks, to monitor the situation and prepare reports based on data analyzed and ar-
chived by other experts. These e-documents are called application workshops: they
are devoted to end-users who perform a given task to obtain some final product.
The two e-documents of the figures are designed for two communities, whose no-
tations share a common subset of symbols. These two communities may be consid-
ered members of the larger community of experts interested in glaciological data.
However, they also use different tools to perform their tasks: a photo interpreter, for
example, obtains the spectral characterization of the glacier areas, the clerk derives
reports on the environmental situation. The use of two different application work-
shops lessens their cognitive load in choosing tools, avoids possible errors, mistakes
and misunderstandings and makes the navigation in the workshop more simple and
reliable. In our view, the creation of the application workshops has to be done by an
expert in the management of environmental data, who knows notation and needs of
both the communities, and is able to plan the documentation to be produced. This
expert, here called ‘the manager’, needs a workshop to create new application work-
shops. This workshop, called manager workshop, is an instance of a different type of
workshops, called system workshops, in which tools are available for creating a work-
shop for a community of users. System workshops can be recursively defined for
experts who create the system workshops for managers.
Fig. 1. Interaction with B-Glacier, the workshop for a photo-interpreter: 1a) the selection of an
ablation area; 1b) the results of the computation fired by the previous selection.
486 P. Carrara et al.
In this way a cascade of system workshops can be created. At each level, an expert
uses a system workshop to create the tools and notations for a more specialized ex-
pert, who will contribute in creating a final application workshop customized to spe-
cific end-users and type of task. On the whole, the approach defines workshops to
support task achievement by a community of users. The number of levels between the
top computer scientist workshop and the bottom application workshops and the num-
ber of the workshops at each level are determined by the work organization in the
community.
Fig. 2. Interaction with B-Monitoring, the workshop for a clerk: the selection of an ablation
area (1a) in this context fires the simple retrieval of data relative to that area (1b).
3 A Model of the Human - e-Document Interaction Process
The strategy for designing workshops is supported by a model of human-e-document

interaction [4], which informally evolves the PCL (Pictorial Computing Laboratory)
model presented in [2] for WIMP (Windows, Icons, Menus, Pointers) environments
[5]. In this approach, Human-Computer Interaction is modeled as a cyclic process in
which the user and the interactive system communicate by materializing and inter-
preting a sequence of messages at successive time stamps, the user by cognitive crite-
ria, the computer by programmed criteria [2]. Users interpret the whole images repre-
sented on the screen display, formed by text, graphs, pictures, icons, etc. by recogniz-
ing characteristic structures (css), i.e. sets of image pixels which they perceive as
functional or perceptual units. The cs recognition results into the association of a
meaning with a structure. Once interpreted the message, the user decides what to do
User-Driven Adaptation of e-Documents 487
next and manifests his/her intention by an activity performed operating on the input
devices of the system. On the other hand, the system associates each graphical entity,
which is in turn a set of pixels, i.e. a cs, with a program , and computes the response
to user activity materializing the results on the screen, so that they can be perceived
and interpreted by the user. In principle, this cycle is repeated until the user decides
that the process has to be finished.
4 Workshop Implementation with BANCO
The tool by which the cascade of workshops can be implemented is BANCO [3]. It
exploits a novel architecture on the Web, in which a user can locally customise
his/her environment still maintaining the consistency of data and tools with those of
the Web. On the user side, BANCO, enriched by a kernel of scripts and a library of
customisation files, is accessed by a XML/SVG-complaint browser. The messages
exchanged with the server are in eXtended Markup Language (XML), and convey
also the specification necessary for their interpretation by the browser, in Scalable
Vector Graphics (SVG) format [6]. SVG is used to specify the set of css and their
organization into the layout of the user interface as well as the links to computational
tools. The programs representing the meaning of the css are written in Jscript. They
materialise data according to customisation rules. Changing the library of rules they
can be adapted to different user habits. The workshops are implemented by several
instances of BANCO, each one specialized to its users and tasks through the defini-
tion of the set of its css, computational meanings and links between them, which are
interpreted by a uniform mechanism.
BANCO is an e-document written not to be read, but able to transform other un-
readable e-documents into a form readable by users of a specific community.
References
1. Brusilowsky, P.: Methods and techniques of adaptive hypermedia, User Modelling and
User Adapted Interaction 6(2-3) (1996), 87-129
2. Bottoni, P., Costabile, M. F., Mussio, P.: Specification and Dialog Control of Visual
Interaction, ACM TOPLAS 21(6) (1999) 1077-1136
3. Carrara, P., Fresta, G., Mussio, P.: SVG: More than a markup language for vector graph-
ics, Euroweb 2001, Pisa (Italy), (2001), 245-257
4. Carrara, P., Fogli, D., Fresta, G., Mussio, P. (2002) Toward overcoming culture, skill and
situation hurdles in human-computer interaction. appear on Journal Universal Access in
Information Society
5. Dix, A., Finlay, J., Abowd, G., Beale, R.: Human Computer Interaction, Prentice Hall,
London, (1998)
6. W3C: WWW Consortium, 2001. http: //www.w3c.org/
Server Independent Personalization

for Effective Searching in the World Wide Web
Lillian N. Cassel1 , Ursula Wolz2 , and Robert E. Beck3

1
Villanova University; Villanova, PA USA
[email protected]
2
The College of New Jersey; Trenton, NJ USA
[email protected]
3
Villanova University; Villanova, PA USA
[email protected]
1 Introduction
Effective Web searching involves two components: (1) a search tool must know
about and characterize as many Web pages as possible and (2) it must match
a user query to the most suitable resource. The essential problem addressed by
personalization systems relates to the second component of Web searching. How
can the information that a search service holds about resources available on the
Web be matched to the diverse needs of the global user base? In this paper we
introduce an alternative to the usual model for user personalization.
2 WHAT
The Web Host Access Tool (WHAT) project is a research effort aimed at im-
proved access to relevant information on the Web for users at all levels of ex-
pertise in a highly personalized manner. We focus on effective presentation and
management of search results to users after assisting in the construction of a
search query. The expected effect of this project is to reply to a user search
query with a set of resources that (1) has been obtained from a variety of Web
search tools, (2) has been pre-evaluated for suitability to the user’s request, and
(3) has been formatted for effective presentation of the results to the user.
A fundamental premise is that the user wants control of the search and
the search results, but the user should not be burdened with the details of
obtaining the results. We posit the need for a highly individualized search context
that includes knowledge of prior search experiences, general knowledge of search
and search engines as well as explicit information provided by the user. The
user should be sheltered from the details of formats, query construction rules,
and search strategies of particular search services. In particular, if a user does
repeated searches within a particular domain, the experience of those searches

This work is partially funded by the United States National Science Foundation
Awards EIA-0079770, EIA-0130798

Server Independent Personalization for Searching in the World Wide Web 489
should remain accessible and inform the new search. Ideally, the user’s changing
expectations within a search context (such as discussed by Vakkari and Hakala
[9]) should drive the selection of results presented by the search tool. The burden
for providing this history should be undertaken by the search system.
A distinguishing characteristic of our work is that the added functionality is
provided by a program running entirely on the user’s own local system, namely
one under the user’s control. Access to the system while traveling and while
using other computers is expected and the mechanism is beyond the scope of
this paper. This placement has implications for privacy, for ethical treatment of
the user’s requirements, and for system effectiveness and efficiency.
The major focus of our work is to aggregate search results from numerous
sources, merge these results with what we know about our user’s preferences
and responses to prior searches, and present a view of the merged search results
consistent with these preferences. We keep our personalization procedures in-
dependent of the actual structure of the knowledge representation used by the
information source. Thus, our approach is not tied to any particular search envi-
ronment and can be considered a generalization of the problem of personalization
of search support.
As with a search through a general-purpose search engine, our user poses a
query through query terms. Unlike users of a general search engine interface, the
WHAT user is able to specify a context for the search. The context is defined
by its title and a list of keywords that help to specify its meaning. The set of
results is retrieved, based on both the query terms and the context keywords.
The user is presented with an ordered list of results and selects those most
relevant to the current interest. User feedback concerning the appropriateness
of the results becomes part of the context definition and influences responses
to future queries. Changing the context adapts the user’s setting to the user’s
preferences and focus.
2.1 System Design: WHAT and the WHAT Observer
The Web Host Access Tools (WHAT) system, implemented in Java, includes
a query constructor, a context manager (and database), and a response filter.
Undergraduate research assistants have presented posters on iterations of the
system since 1998. [1,2,3]
The query constructor interacts with the context manager and the user inter-
face. From the context manager, the query constructor learns of terms that have
previously been included in similar searches that can assist the current search.
It combines the user search terms with weighted keywords associated with the
context to form search strings to send to a set of search services. In collabora-
tion with the user, the context manager accumulates and analyses prior search
topics enhancing the search context. The response filter, in addition to eliminat-
ing duplicate results, examines prior results in this context for clues about user
preferences. The response filter also captures user impressions about results and
updates the context history for use in future searches. Current work builds on
490 L.N. Cassel, U. Wolz, and R.E. Beck
known techniques [4] for judging the relevance of a page to a search query to
predict the user response.
While a number of projects are attempting to link a given query to a special-
ized search engine [8], WHAT has the advantage of an explicit specification of a
context for the query. Specification of the search context gives important infor-
mation for routing the search to specialized search engines, as well as filtering
results that come from general search engines.
We seek to minimize the client’s need to interact with a server, both for
performance reasons and to protect the user’s privacy. At the same time, we do
not want to burden individual client systems with tasks better suited to a central
facility. Tasks well suited to a server facility include recognition of changes at the
search sites that require software updates, learning about new search tools and
integrating them into the system, determining performance characteristics that
affect choice of which tools to use for a given type of query, perhaps determining
time-of-day factors for making a good choice of search tool. Gregg Davis at
Villanova made initial steps in identifying types of seach engine failures that can
easily be recognized and corrected in a client system and those better suited to
intervention from a WHAT system server. [5]
2.2 WHAT Observer

System assessment requires evaluating the “goodness” of a response as impacted
by the various systems we have implemented and plan to enhance. As we devel-
oped specific evaluation questions we saw the need for the WHAT Observer to
be used exclusively as a research tool. [7] This continues the work of Hartson,
et al [6] on remote interface evaluation, but extends it to allow the evaluation of
the impact of underlying analysis systems (such as the context manager). The
WHAT Observer allows us to selectively observe the interaction between the
user and the WHAT system. Data are automatically stored in a server-based
database. User surveys elicited through web forms can augment the database.
2.3 Goodness Metrics

Our initial “goodness” metric is determined by the order of query responses. We
posit that in any search, the “best” responses should appear before “worse” re-
sponses in the presentation order. When a user initiates a query through WHAT,
three ordered lists result that can be captured by the Observer: (1) The web
search services return an ordering determined by their metric. (2) The WHAT
context manager reorders the initial list and presents it to the user. (3) The user
implicitly reorders the list when assigning “yes”, “no”, or “maybe” tags to some
items and ignoring others. The tags create a tripartite grouping of the examined
items. We presume in this analysis that examination is systematic and complete
up to a point in the initial list. Items that are unexamined come after such a
point and are consigned to the “no” group.
In a perfect response, the contents of each group would exactly match the
groups assigned by the user. We can analyze the migration of items from WHAT’s
Server Independent Personalization for Searching in the World Wide Web 491
response list to the user groupings. Little migration suggests a “good” ordering.
Note that there is no significance within the grouping. An item’s position in the
group is derived entirely from its position in the original list. An exact definition
of “little migration” awaits analysis of our initial test results. Furthermore, it is
dependent upon whether the user thought the search was “successful.” Conse-
quently, we anticipate that the degree of migration as a metric will not be fixed
for all users, but will be dependent upon the user and the scenario in which the
search occurred.
The migration metric also provides insight into the efficiency of personaliza-
tion. A system that expends significant resources in time and space may not be
cost-effective if there are only small improvements in minimizing migration.
3 Summary
The WHAT project, which builds on existing Web search tools, places resources
oriented to user-specific services on the user’s system. This allows specific per-
sonalization to an individual user and addresses the important issues of privacy
and user trust. The resulting system includes an assessment component and early
results indicate significant improvement over general Web search techniques.
References
1. Jonathan Anderson and Jason Dobies. The web host access tools (what) project.
SIGCSE 2001 Charlotte, NC, 2001.
2. Brice Behringer, Mark Nikolsky, and Michael Sipper. A gui for web host access
tools. Technical report, The College of New Jersey, Ewing, NJ, 1998.
3. Greg Bronevetsky. The brains of what: A data structure for internet searches.
SIGCSE 99, New Orleans, LA page 378, 1999.
4. Chris Buckley, Gerald Salton, and James Allan. The effect of adding relevance infor-
mation in a relevance feedback. In Proceedings of the 17th ACM-SIGIR Conference
on Research and development in information retrieval, pages 292–300, 1994.
5. Greg R. Davis. Web host access tool updating protocol. Technical report, Villanova
University Department of Computing Sciences, 1999. Independent Study Project
under the direction of Dr. Lillian Cassel.
6. Rex Hartson, José C. Castillo, John Kelso, and Wayne C. Neale. Remote evaluation:
the network as an extension of the usability laboratory. In Conference proceedings
on Human factors in computing systems, pages 228–235, 1996.
7. Jared Klett. Web host access tools observer. SIGCSE 99, New Orleans, LA, page
379, 1999.
8. Atsushi Sugiura and Oren Etzioni. Query routing for web search engines: Archi-
tectures and experiments. In Ninth International World Wide Web Conference
(WWW9), Amsterdam, The Netherlands, May 2000.
9. Pertti Vakkari and Nanna Hakala. Changes in relevance criteria and problem stages
in task performance. Journal of Documentation, 56(5):540–562, September 2000.
Preventing Misleading Presentations of XML
Documents: Some Initial Proposals
Alison Cawsey, Euan Dempster, Diana Bental, Daniel Pacey, Howard Williams,
Lachlan MacKinnon, and David Marwick
Department of Computing and Electrical Engineering,

Heriot-Watt University, Riccarton,
Edinburgh EH14 4AS, UK; Tel: +44 131 451 3410; Fax: +44 131 451 3431
[email protected]
Abstract. It is now straightforward to develop a range of different stylesheets

to present XML documents in different ways, for example to create personal-
ised presentations. If the XML document is available on the World Wide Web,
then (subject to copyright) anyone can create their own stylesheet to present
that document in new ways. This has the potential to allow improved "added
value" services, such as personalised news feeders. But the power of stylesheet
languages such as XSLT means that the document may be substantially trans-
formed, with sections deleted or re-ordered. This re-structuring may result in a
misleading and even dangerous presentation. This paper presents some propos-
als for putting some control in the hands of document authors, to allow them to
indicate allowable transformations, and to provide a limited validation mecha-
nism to verify that a transformed document meets the requirements of the au-
thor.
1 Introduction
Anyone can create their own stylesheets to present XML documents that are available
on the World Wide Web. It is straightforward in JavaScript, for example, to specify
that a stylesheet at one location should be applied to an XML document taken from
another location. This has great power - for example, it allows those providing infor-
mation services to "pull in" material from different sites and present it in a way that
suits their users. Similarly one can develop services to present material in the opti-
mum manner for various display devices, or to meet the needs of users with disabili-
ties. Abstracting services could display certain sections with the amount presented
tuned to the readers’ requirements.
Already there are a number of systems that use XML in personalisation. GUIDE 3
is an online tourist guide where a user profile is supplied by the tourist; SETA 1 is a
web shopping assistant that personalises information on products. More personalised
systems are reviewed in 2. Currently, the development of the resources and the per-
sonalisation services are typically managed by a single group, hosted on one server.
But as more and more resources are designed in this way we will see new services
emerging which use the data in the XML resources provided by others, but present

Preventing Misleading Presentations of XML Documents: Some Initial Proposals 493
this information in new ways. However, the power of stylesheet languages such as
XSLT 5 mean that the document may be substantially transformed in the process of
presenting it in these new ways. Sections may be omitted, re-ordered, or generally re-
structured. This may result in misleading or dangerous presentations, as important
information is omitted, or re-ordering changes the document semantics. This problem
is very familiar to the document summarisation community 6. Simple sentence ex-
traction systems can result in confusing summaries, as the juxtaposition of two sen-
tences originally taken from separate sections of a document results in incorrect inter-
pretation.
This paper presents an initial analysis of this problem, and some limited proposals
to enable authors of XML documents to specify just what transformations are allow-
able, and to provide a way of validating transformed documents. We start by looking
at some possible problems that may arise in transforming documents, taking medical
documents as our case study. We then suggest an annotation scheme that would allow
authors to specify allowed transformations, avoiding the problems identified. Finally
we consider how or whether that scheme may be enforced, and suggest the use of a
simple validation service, where the transformed document can be validated against
the annotations.
2 The Problem of Misleading Transformations
We have looked at a range of document types in order to determine some of the likely
problems that may arise through transformation, and how an author could state con-
straints. Here we will focus on an example of a drug catalogue (available as a publicly
accessible XML file). This catalogue might include name, price, side-effects and use
of a range of drugs. Using stylesheets an intermediary provider could:
• Present just part of the catalogue – maybe just the drugs that relate to the user’s
condition.
• Omit part of the drug information (e.g., the side effects).
• Re-order the catalogue, perhaps putting relevant drugs at the top, or ordering by
price.
• Combining the catalogue with another, perhaps adding the supplier name into the
table.
• Include some side-effects of drugs, but only as space allowed.
• Add or omit disclaimers, or recommendations to consult professionals.
It should be clear that some of these transformations are inadvisable and potentially
harmful, whereas others may play a useful role. As a result the provider of the cata-
logue may want to specify certain constraints on the transformation, such as:
• If a drug is included, then its side effects should always be listed as given.
• A disclaimer of liability should always be included.
There will be many other constraints of different forms. Suppose the catalogue also
provided some information on the advantages and disadvantages of a drug. These
could be omitted, but if the advantages are listed, so should the advantages. It may be
further required to list the advantages and disadvantages in the order given. And there
494 A. Cawsey et al.
are cases where the juxtaposition of elements is significant. For example, adding in a
paragraph into a paper might destroy its coherence.
From this analysis we suggest that authors may want to provide constraints of the
following forms.
1. Element is required to be included.
2. The ordering within Element must be retained.
3. If Element1 is included, so must be Element2.
4. Element1 and Element2 must not be split.
In most cases general rules such as this applying to all drugs may suffice. But there
may be special cases where the author may wish to supply additional constraints.
• If presenting DRUG-X, ensure that the additional information is included.
So we want to be able to describe whether rules apply to all elements of a given
type (e.g., for all drug data always include any side-effects), or just to specific ele-
ments listed (e.g, for this drug, ensure the additional information is included).
3 A Proposed Annotation Scheme
Based on the discussion above, we need a way of specifying, for a given document or
document type, which sections can be re-ordered, deleted, and which sections are
dependent on others (so the inclusion of one demands the other). If every (significant)
element in the document has an associated ID, then we can indicate these constraints
using some simple statements based on describing relationships in groups of ele-
ments. We specify for groups whether the group is mandatory, must retain order, and
whether elements can be split. The following two examples illustrate this scheme:
<group order="no" maysplit="no" mandatory="yes">
<item required="yes">Image1<item>
<item required="yes">Legend1<item>
</group>
Image1 and Legend1 are mandatory, but they can be re-ordered. Extra material
cannot be inserted between them – perhaps that would destroy the connection be-
tween image and legend
<group order="yes" maysplit="no" mandatory="no">
<item required="yes">advantages1<item>
<item required="yes">disadvantages1<item>
</group>
The advantages and disadvantages are optional, but if the group is included then
both items are required.
The above scheme allows rules to be applied to individual elements (through their
IDs). However it is easy to expand this approach to allow rules that apply to all ele-
ments of a given type. We could simply supply default values for an element type
(e.g., side-effect) which can be overridden in a specific case. This annotation scheme
allows authors to indicate the transformations permitted, in a machine-processable
format, but leaves open the problem of how to verify that these constraints are met.
Although we cannot avoid malicious action, we can provide some simple tools that
would allow a co-operative information provider to check that their transformed
document meets the constraints requested by the source provider.
Preventing Misleading Presentations of XML Documents: Some Initial Proposals 495
4 Validation
If the information provider is transforming the document into a completely different

document type (as is indeed currently normal) then the problem is difficult - at best, if
they retained the element IDs, we could use simple tree traversal to verify orderings,
etc. However, things become a little easier if we assume that the presentation of the
document will be done in two stages: transformation (into the same document type)
and formatting (into formatting objects 5, or into XHTML). We will therefore look
briefly at the specific problem of validating a transformed document against con-
straints, given that it is in the same document type as the original.
To a limited extent, we could merely ensure that the DTD or Schema is as "tight"
as possible and require that the transformed document is a valid instance of this
document type. Schemas (or DTDs) allow us to state ordering constraints and com-
pulsory elements). But we also want to be able to specify rules-- IF there is an ele-
ment matching this constraint THEN do this check. We also want to be able to assert
constraints that just apply to one document instance, or one particular element (e.g.,
for THIS paragraph, don't omit the copyright). Such specific constraints are not easily
handled within the document type.
What is needed is a schema language that allows you to flexibly specify rules -- IF
there is an element matching this constraint THEN do this check. At first glance
Schematron 7 seems to provide a solution. Schematron is a schema language based on
such rules. It allows you to identify patterns in the document, and, if these patterns
occur, check for the existence of some other element. An example rule is given be-
low, which checks and reports on obligatory child elements for a drug element.
<pattern name="Drug checker">
<rule context="drug">
<assert test=”side-effect”> Side effect is missing
</assert>
<report> Side effect is present </report>
<assert test=”@producer”> Producer attribute is missing
</assert>
</rule>
</pattern>
The above would apply to all drugs. We could specify more rules for more specific
elements in a similar manner, with, for example, context=drug[@ID=”d2].
Schematron is very good for specifying fairly simple rules such as these, but more
complex cases are harder to handle. We have also looked at using XSLT as the lan-
guage for the validator, providing significantly more flexibility. We can specify tem-
plates that result in error messages being written out if constraints are violated:
<xsl:template match="drug">
<xsl:choose>
<xsl:when test= “side-effect”> OK </xsl:when>
<xsl:otherwise> No side effect listed. </xsl:otherwise>
</xsl:choose>
</xsl:template>
We have developed an XSLT stylesheet which validates documents against con-

straints expressed in the format indicated in the previous section. This could be used
496 A. Cawsey et al.
in an online validation service, allowing cooperative providers who use their re-
sources to verify that transformed documents meet the source providers' constraints
(perhaps adding some icon if it does).
5 Conclusion
We have presented an initial analysis of a likely problem with misleading presenta-

tions, if authors “lose control” of the styles applied to their document. We suggest a
simple annotation scheme allowing authors to specify allowed transformations, and
explore how simple validators could be written to check that constraints are obeyed.
A simple validation procedure such as this cannot address the problem of malicious
adaptation (which requires a social/legal solution), but allows information providers
to describe some basic constraints on transformation, which can be checked by coop-
erative secondary providers.
Acknowledgements
This work has been funded by the Engineering and Physical Science Research Coun-
cil, grant reference GR/N22229/01 Dynamic Information Presentation, and we grate-
fully acknowledge their support.
References
1. Ardissono, L. & Goy, A., Tailoring the interaction with users in electronic shops. in Kay,
J., (ed), User Modelling: Proceedings of the Seventh International Conference, UM99,
Springer, (1999).
2. D Bental, L MacKinnon, H Williams, D Marwick, D Pacey, E Dempster and A Cawsey,
Dynamic Information Presentation through Web-based Personalisation and Adaptation -
An Initial Review, In Joint Proccedings of HCI 2001 and IHM 2001, A Blandford, J Van-
derdonckt and P Gray (Eds), pp 485-500, Springer, (2001).
3. Cheverst, K., Davies, N., Mitchell, K. & Smith, P., Providing tailored (context-aware)
information to city visitors. in Brusilovsky, P., Stock, O. & Strapparava, C., (eds), Adap-
tive Hypermedia and Adaptive Web-Based Systems, 73 -85. Springer, (2000).
4. Extensible Stylesheet Language (XSL), Version 1.0, W3C Recommendation 15 October
2001, (2001), http://www.w3.org/TR/xsl/
nd
5. Kay, M., XSLT 2 Edition - Programmer's Reference, Wrox Press Ltd, Birmingham, UK,
(2001).
6. Mani, I., & Maybury, M., (eds) Advances in Automatic Text Summarization, MIT Press,
(1999).
7. Van der Vlist E., XML Content Management System Using XSLT, Schematron,
and Ant, O'Reilly Open Source Convention, XTech 2001: Cutting Edge XML.
(2001).
Adaptivity Conditions Evaluation for the User
of Hypermedia Presentations Built with AHA!
1 2
Alessandra Cini and José Valdeni de Lima
1
Fundação Universidade de Caxias do Sul, Rua Francisco Getúlio Vargas, 1133,
Caxias do Sul - RS, Brazil, Phone: + 55 (54) 218 2278
[email protected]
2
Universidade Federal do Rio Grande do Sul, Av. Bento Gonçalves, 9500 - Bloco IV / 226,
Porto Alegre - RS, Brazil, Phone: +55 (51) 33 16 61 61
[email protected]
Abstract. The importance of adaptive hypermedia propelled the development

of several authoring tools. Even though most tools are efficient for their
designed task, they do not guarantee that the adaptivity is used in a high degree,
in order to satisfy users with distinct goals, preferences, background and/or
needs, as usual on the Web. In this paper we will present a system in
development built to evaluate the adaptivity degree of presentations built into
the AHA! authoring tool. The author, after build his presentation in AHA!,
submits it to our system for evaluation. The results obtained from the evaluation
can aid the author to know if the built presentation it is according to the
proposed goals or the adaptivity degree should be increased.
1 Introduction
The importance of adaptive hypermedia area in recent years and the difficulties of
adaptive presentations authoring propelled the development of several authoring tools
for adaptive hypermedia.
The several authoring tools available, such as AHA! [1] (Adaptive Hypermedia
Architecture), developed at Eindhoven University of Technology and InterBook [2],
developed at Carnegie Mellon University, have countless resources to promote the
adaptation and to facilitate the construction of a presentation.
The production of an adaptive presentation must consider the peculiarities of the
content that is being approached and the user’s individual goals, preferences,
background and/or needs. Considering that most of the presentations are available via
the Web, the user’s profile can be quite distinct, requiring the presentations to have a
high adaptivity degree to satisfy such users.
2 Importance of Adaptivity Degree Evaluation of a Presentation
The use of an authoring tool for adaptive hypermedia doesn't assure that a
presentation is built with high adaptivity degree. As a consequence, presentations can

498 A. Cini and J. Valdeni de Lima
be built with insufficient adaptive resources. If we consider the fact that, in a

presentation, the author can define a display requirement for only one page and that
this requirement can be fulfilled by accessing another single page, the adaptivity
degree can be very low. Still, we cannot say that such presentation is not adaptive.
Starting from the perception of this limitation, we are building a system able to
measure the presentation adaptivity degree. We did opt for evaluating presentations
built in the AHA! system, because the current release of InterBook is only available
for Apple Macintosh computers. Due to the differences in the presentations syntax,
we didn't yet get to evaluate presentations built in any other authoring tool.
We believe that the larger the number of concepts participating in the adaptive
process, the larger will be the adaptivity degree and the larger will be the amount of
users that can use the presentation in a personalized way.
It is important that the user interacts with the presentation in an adaptable way,
reporting as much as possible about what he already knows about each concept,
updating or even correcting the values of his user model [3]. The larger the number of
concepts which may be directly changed by the user, the more freedom will be
available to adapt the presentation for real needs. Thus we believe we will have a
presentation with a greater adaptivity degree.
The existence of an adaptive global navigational map helps the user to understand
the overall hyperspace structure and his absolute position on it [4], which can make
the navigation through the presentation more pleasant.
We believe the more fragments and conditional links a page of the presentation
has, the more opportunities it has of being adapted to the user. It is not desirable that
users with different needs and requirements receive a page in the same way.
3 AHA! Functions and the Authoring Process
The AHA! architecture is based in the elements defined in the AHAM reference
model [5], that are: domain model, user model, adaptation model and adaptive engine.
The domain model describes the application domain in terms of fragments, page
concepts and abstract concepts. A page concept is a concept represented by a page
while an abstract concept is a concept that is not a page. Each page is a XML file that
contains fragments of HTML text that are conditionally included, and with hypertext
links that, in a similar way, can also be conditionally shown. The fragments are
content parts presented in the pages [1].
The adaptive definitions the author inserts in an AHA! presentation are part of the
user model and the adaptation model and include value generation relationship rules
for concepts and requirement relationship rules for concepts [6]. The first ones
indicate which other concepts are also updated when a given concept has its value
updated in the user model, while the last ones indicate which conditions must be true
so that a certain concept can be presented. These relationships rules are stored in two
XML files named, respectively, genlistfile and reqlistfile. The conditional expressions
are also part of the adaptation model and are presented directly in the XML pages of
the presentation, indicating the desirability of fragments.
Adaptivity Conditions Evaluation for Hypermedia Presentations Built with AHA! 499
4 Characteristics and Operation

of the Presentation Evaluation System
Our presentation evaluation system, developed in Java, analyzes XML pages of a

presentation built in AHA! and generates some statistical information. The
information is used to measure the adaptivity degree (high, medium or low) of a
presentation from several points of view.
In this system, the presentation evaluation starts by reading the properties file of
the presentation. In a presentation built in AHA!, the properties file contains basic
information. For example, the physical name of the reqlistfile and genlistfile files and
the physical name of the documentslistfile file which contains a list of all pages of the
presentation.
An analysis of the reqlistfile and genlistfile files allow important information to be
obtained which can be used to evaluate the adaptivity degree of a presentation. The
following information is generated by the presentation evaluation system: total
amount of course pages; amount and name of page concepts that update values in
other concepts; amount and name of page concepts or menu items (separately) that
have restrictions for their display; amount and name of page concepts (separately) that
can have their values in the user model updated directly by the user; amount and name
of the concepts that are restrictions for other concepts; amount and name of pages of
the presentation that have fragments and/or links that are conditionally displayed; and
amount and name of the menu items that are updated by rules.
From the obtained information, it is possible to measure several adaptivity aspects.
The following ones are considered: a) the adaptivity degree of the user model in the
generation of updates is measured by the percentile of pages that update other
concepts in relation to the total pages of the presentation. When we have a very small
amount of pages that update other concepts, it means that most of the pages don't
influence the personalization of the presentation; b) the adaptivity degree in
restrictions of the adaptation model is measured by the percentile of pages that have
restriction for their presentation in relation to the total pages of the presentation. If
few pages have display restrictions, the presentation tends to be very similar for users
with different characteristics; c) the user adaptable behavior degree in the
presentation is measured by the percentile of concepts which can be altered directly
by the user and are used as requirements in restrictions for other concepts in relation
to the total number of concepts which can be altered directly by the user; d) the
content adaptation degree in the pages is measured by the percentile of pages that
have conditional fragments in relation to the total of pages of the presentation. The
larger the amount of pages with conditional fragments, the greater the possibility of
the page to be adapted to each user; e) the adaptive navigation degree in the pages is
measured by the percentile of pages that have conditional links in relation to the total
pages of the presentation. The larger the amount of pages with conditional links, the
greater the possibility of the page to be adapted to each user; f) the existence of an
adaptive navigational map is another adaptivity aspect measured, which indicates if
there exist menu items updated by rules. The information is presented with only
Boolean values.
For the first five adaptivity aspects presented, if the percentile obtained is lower
than 20% the adaptivity degree is considered low, if it is from 20% to 40% it is
considered medium and if it is larger than 40% is considered high.
500 A. Cini and J. Valdeni de Lima
The percentiles above defined to measure the adaptivity degree for each of the
analyzed aspects were established by preliminary analyses. A form of evaluation of
the results obtained with the system would be to build a presentation with two version
with different adaptivity degrees, one classified by our system as a high degree and
the other as a low adaptivity degree. The two presentations would use the same
content pages, just the information of the reqlistfile and genlistfile files would be
different, as well as the fragments and conditional links of the pages. Different user
groups could use the two presentations and, later on, the groups would be evaluated to
verify the satisfaction degree with the presentation and the amount of acquired
information. This could confirm that presentations indicated by our system as having
a high degree of adaptivity are really better.
Studies still need to be done to establish new tests that can be applied to the
adaptive presentations built in AHA! to evaluate other adaptivity aspects, as support
to multiple languages, media adaptation, etc.
References
1. De Bra, Paul; Aerts, Ad; Houben, Geert-Jan; Wu, Hongjing: Making General-Purpose
Adaptive Hypermedia Work. Proceedings of the WebNet Conference. San Antonio, TX
(2000) 117-123.
2. Brusilovsky, P., Eklund, J., and Schwarz, E.: Web-based education for all: A tool for
developing adaptive courseware. Proceedings of Seventh International World Wide Web
Conference. Brisbane, Australia (1998) 291-300.
3. De Bra, Paul; Ruiter, Jan-Peter: AHA! Adaptive Hypermedia for All. Proceedings of the
WebNet Conference. Orlando, FL (2001) 262-268.
4. Brusilovsky, Peter: Methods and Techniques of Adaptive Hypermedia. Adaptive Hypertext
and Hypermedia, Klumer Academic Publishers. The Netherlands (1998) 1-43.
5. De Bra, Paul; Aerts, Ad; Houben, Geert-Jan; WU, Hongjing: AHAM: A Dexter-based
Reference Model for Adaptive Hypermedia. Proceedings of ACM Hypertext’99. Darmstadt,
Germany (1999) 147-156.
6. Wu, Hongjing; Houben, Geert-Jan; De Bra, Paul: Supporting User Adaptation in Adaptive
Hypermedia Applications. Proceedings InfWet2000. Rotterdam, the Netherlands (2000).
Development of Adaptive Web Sites
with Usability and Accessibility Features
Marta Fernández de Arriba and José A. López Brugos
Computer Science Department. University of Oviedo, Viesques Campus, Gijón. Spain.

Abstract. This paper discusses a development system of Web sites

adapted to the preferences of each user. A markup language (AWUML -
Adaptive Web site Usable Markup Language) based on XML technology
is defined for use in the specification of the page contents, which guaran-
tees the fulfillment of usability and accessibility requirements. Each user
visualizes his own version of the web site resulting from the application
of his profile, as much in presentation aspects as in content topics. The
separation of document contents from its presentation makes mainte-
nance of the web site easier. The use of this system facilitates access to
the network, independent of the disabilities or the technical equipment
limitations of the end user.
1 Introduction
The way to achieve success and user loyalty consists in developing adaptive web
sites that take into consideration their preferences, limitations and necessities.
Furthermore, most web sites present deficiencies in their orientation to a
specific hardware/software configuration. This deficiency makes it difficult for
some of the potential users to browse the web site, sometimes even blocking
access to certain resources.
Another problem is that nowadays most of the developed web sites mix the
content definition of the site with its presentation, thus the consequent increase
in cost and effort in its maintenance (this is especially important due to the
continuous updating that the web sites undergo).
In this paper, a development structure of web sites adapted to the needs of
users while fulfilling usability and accessibility requirements is defined. Thus,
the system architecture will be defined as well as a markup language(AWUML:
Adaptive Web site Usable Markup Language) using XML technology [5].
2 Web Site Structure
A web site developed using the system defined in this paper is made up of a
set of pages whose contents are defined in a series of XML files. The XML files
follow the syntax specified for the AWUML markup language.

502 M. Fernández de Arriba and J.A. López Brugos
The web site relies on a user identification module to recover the user’s
preferences of presentation as well as of contents. These preferences are stored
in the corresponding user profile.
Once identified, the user will receive the requested web site page in his
browser, which is generated by applying (filtered according to the corresponding
parameters of the profile of this user) the translation file to the contents using
XSLT format [6]. The XSLT file is generated by the presentation filter according
to the related parameters of the profile, as presented in Fig. 1.
Usuario1 UsuarioN
Internet
Presentation
filter Adapted
Web site
XSLT
Identification User Translator

profile
AWUML
Content
Web Site filter
Content
Fig. 1. Adaptive web site structure.
The presentation features selected by the user as well as the web site content
preferences are stored in the user profile.
Each web site has a list of topics (using a hierarchical structure) related to
the page information presented. Every page must belong to one of these topics.
All content elements and links can belong to a subcategory of the main topic of
the page (by default, they belong to the same topic as the page).
Each user sets up the preferred topics list to be shown to him in his profile.
When selecting a topic, all parent topics are automatically selected. The user
will also be able to indicate all the children at the time he selects a topic.
3 AWUML: Adaptive Web Site Usable Markup Language

The page contents of a web site developed using AWUML are presented in Fig. 2.
The web site is made up of a set of pages written in XML following the AWUML
syntax, forming a tree design, and beginning on a node root that is the home
page.
Adaptive Web Sites with Usability and Accessibility Features 503
Home Page
appearance.xml
Page 1 ... Page n
DTD files
Page 1.1 ... Page 1.n
Fig. 2. Web Site Contents in AWUML.
All the web site pages must have an ’appearance.xml’ file that defines the
common presentation features of all its pages.
The structure and the contents of the web site pages (using AWUML syntax)
are defined in two DTD files (one for the home page and the other for the rest).
Each page of the web site developed using AWUML is an XML document
that uses the syntax defined by the language according to the following structure:
1. Reference.
The XML version is specified, as well as the DTD file associated to the page,
the XML file that specifies the common aspect and the translation module.
2. Root element.
Head. This includes the reference to the appearance file, the title of the
page, the topic to which it belongs and the location within the web site.
Sections. The different links are specified from the current page to the
sections at an inferior level. In each section, a link, a description, the
topic and a quick access key appear.
Contents. A series of multimedia elements is specified. They define the
contents of the page. Thus, textual contents, images, sounds and links
can appear with their corresponding associated attributes.
Returns. Links to the home page of the web site, to the previous page or
to the beginning of the present page are specified.
4 Document Translations
From the specification of a web site written in AWUML, the server shall translate
the corresponding XML pages using the style sheet translate.xsl.
The translated XML pages are obtained from the web site content pages
applying them to a filter according to the content preferences stored in the
profile of the user who requests the pages.
The translation file (translate.xsl) is generated online, as determined by the
presentation preferences in the user profile. Therefore, a page will reach each
end user adapted to his preferences and necessities as a result of his request, so
that the same page of the web site could be shown in as many final versions as
different user profiles exist.
504 M. Fernández de Arriba and J.A. López Brugos
5 Usability and Accessibility Requirements Achieved

The development of web sites using strictly AWUML language allows us to assure
the fulfilment of a set of usability and accessibility requirements defined by the
standards of the Web Accessibility Initiative [7], the recommendations published
in Spain by AENOR (Spanish Agency for Standardisation and Certification)[1],
heuristic usability [3] and usability designing guidelines [4] and [2].
Among the requirements achieved, the following should be emphasised:
– Provide equivalent alternatives to auditory and visual contents.
– Use style sheets to control layout and presentation.
– Identify the primary natural language of a document.
– Ensure user control of time-sensitive content changes.
– Design for device-independence and provide keyboard shortcuts to links.
– Provide clear navigation mechanisms.
– Ensure that documents are clear and simple.
6 Conclusions
In this paper, a system for the development of adaptive web sites is presented,
considering usability and accessibility problems.
The definition of a complete markup language is presented, using the capac-
ities provided by XML, which allows the development of web sites that fulfil
accessibility and usability requirements, facilitating the development and main-
tenance of these sites (from the point of view of the contents of the pages as well
as of their presentation). This separation of contents from presentation allows
the generation of versions different from the final documents (even in different
languages) that the user receives in his browser by merely generating a transla-
tion module (translate.xsl) on-line, based on the user profile.
The possibility of multimedia element configurations that are sent to the end
client allows for the sending of optimized information through the network, either
for adapting to the limitations or preferences of the user or to fulfill requirements
of the client software (as in the case of WAP technology).
The use of this language allows the standardisation of the interfaces in the
Network, enabling anyone’s access to and use of its contents (including disabled
people).
References
1. AENOR: UNE 139.802 EX (1998). URL: http://acceso2.uv.es/aenor/accengl.htm
2. Nielsen, J.: Usabilidad. Diseño de sitios Web, Ed. Prentice-Hall (2001)
3. Nielsen, J.: Heuristic Evaluation. URL:http://www.useit.com/papers/heuristic
4. Pearrow, M.: Web Site Usability HandBook, Ed. Charles River Media (2000)
5. W3C: Extensible Markup Language (XML) URL:http://www.w3.org/xml
6. W3C: XSL Transformations (XSLT). Version 1.0. W3C Recommendation 16 Novem-
ber 1999. URL:http://www.w3.org/TR/WD-xslt
7. W3C: Web Content Accessibility Guidelines 1.0. W3C Recommendation 5 May
1999. URL:http://www.w3.org/TR/WAI-WEBCONTENT/
An Adaptive e-Commerce System Definition
1 2 1
Francisco José García , Fabio Paternò , and Ana Belén Gil
1
Departamento de Informática y Automática – Facultad de Ciencias
University of Salamanca – Spain
{fgarcia, abg}@usal.es
2
CNUCE - C.N.R.
Pisa – Italy
[email protected]
Abstract. The emergence and growing popularity of electronic commerce and

more specifically of catalogue-based sales, has made it a necessity to define
flexible and adaptive e-commerce systems and architectures. Our work
describes an e-commerce system that is suitable for small and medium
enterprises, where several enterprises share a common e-commerce site, which
automatically adapts its interface to offer end-users the products organised in
electronic catalogues (e-catalogues). To be precise, in this paper we present the
overall architecture of the e-commerce system, an architecture that is based on
an agent-oriented technology. We also describe the actual state our system,
which is called e-CoUSAL, and implements an adaptive agent in the server side
of the architecture.
Keywords. Adaptive agent; E-commerce; Adaptivity in E-commerce; Adaptive
interface; E-catalogue; XML.
1 Introduction
Inspired by the growing popularity of new technologies and owing to the increasing
use of the Internet, several kinds of systems with worldwide scope have appeared
recently. Perhaps the major application area of Internet development is precisely e-
commerce. An especially popular e-commerce segment is the one based on product
catalogues, also known as electronic catalogues or e-catalogues. An e-catalogue can
be defined as the electronic presentation of information about the products and/or
services of an organisation. While other applications can provide similar services, e-
catalogues provide a range and effectiveness of service that exceeds the capability of
any competing application, such as physical or CD catalogues. The interactive
possibilities of e-catalogues eliminate physical storage and make continuous updating
possible and efficient [1].
The entry of an enterprise into the e-commerce world involves strategic decisions
[2] that are not cost and risk free, this is often an obstacle for the integration of an
organisation into the virtual commerce community. The impediments are more
important in small organizations or business, the so-called Small and Medium
Enterprises (SME), in which the amount of investment in technology solutions cannot
be very large.

506 F.J. García, F. Paternò, and A.B. Gil
In this work the overall e-CoUSAL architecture [3] is presented from an agent-
based perspective, and we also describe the first prototype of this e-commerce system;
it implements an adaptive agent in the server side of the architecture. Thus, the
remainder of this paper is organised as follows: Section 2 explains the proposed
agent-based e-commerce architecture with their components, Section 3 examines the
adaptive components of this architecture that are actually implemented. Finally,
Section 4 closes the paper, presenting our conclusions and some ideas on further work
needed.
2 An Agent-Based e-Commerce Architecture
2.1 Components of the e-CoUSAL e-Commerce Architecture
In a schematic way, we show in Fig. 1 the major components of the proposed

architecture. As we stated above, the main commercial policy is based on e-catalogue
shopping, supported by two main components: the e-commerce web server and the
visual catalogue-designer tool. The e-commerce server is the central element, which
interconnects the different parts involved in a typical commerce environment but
more dinamically. The e-catalogue is the element by means of which the end-user
views the seller’s information and interacts with it.
sh o p p in g
SM E F in a l c lie n t
com m erce
m a n a g e m en t
e - c a ta lo g u e p a ym e n t sy s te m
d e sig n
E -co m m e rce
se r v e r
p u b lic a tio n
V isu a l c a ta lo g u e - d e sig n e r to o l B ank
Fig. 1. Components of the e-CoUSAL e-commerce architecture
The SME becomes the main actor of its own virtual business approach as an active
element within the commercial process. It is the responsible for the inclusion and
management of its own contents in the e-commerce site through the use of a
specialised software tool. The designer tool permits the definition, publication and
update of an e-catalogue, and also the configuration of a web server architecture that
allows end-users to have access to this e-catalogue.
The relationships between the SME and the e-commerce server, through the
catalogue-designer tool and also through the server management services, represent a
An Adaptive e-Commerce System Definition 507
B2B dimension in this e-commerce model. Moreover, the server has to provide the
end-users with the commercial services needed for browsing the e-catalogue and for
purchasing; these functionalities in the server site define a B2C dimension of this
model. Thus, the overall architectural model defined above presents a B2B/B2C
hybrid e-commerce model [2].
2.2 Agential View of the Defined Architecture
In a multi-agent e-commerce environment it is necessary to organise agents into

different categories, depending on their functionality and competencies. Several
different forms of agents for e-business systems are distinguished in [5].
Fig. 2. Agents in e-CoUSAL architecture: ovals represent agents and arrows represent
communication between them or between external entities, as end-users. Bold shapes or lines
mean that the agent is actually completely or partially implemented; dotted shapes or lines
mean future development in our architecture.
According to this classification, we have identified the agents and relationships

needed to support our e-commerce architecture proposal, in Fig. 2. We present the
agent-based architecture in coarse granularity and in high abstraction levels, because
we are defining the architectural layer of the system, relating the agents to the main
components presented in Fig. 1.
The application agent is the visual designer tool used by a SME to create e-
catalogues.
The broker agent is in charge of receiving the e-catalogues, expressed in XML
format, validating and storing information in the proper internal database. The tool
sends these catalogues through the system-level agent.
The negotiation agent is still in its first stages of development. This kind of agent
will receive the business components of each SME in the same way that the broker
agent receives the e-catalogues. This property will allow SMEs personalize their
business policies (discount, payments and so on) through these components.
508 F.J. García, F. Paternò, and A.B. Gil
General business activity agents are a set of agents that manage the typical e-
commerce services of the site: shopping-cart management, selling certificates…
The authentication agent is a security agent type that is in charge of identifying the
end-user, which makes it possible to adapt its interfaces to the user’s shopping tastes.
The personal agent is the responsible of customising the interaction with the user
into the e-catalogues and therefore it is implemented as an adaptive agent.
Finally, a planning agent is needed. The presence of heterogeneous problems to be
faced and the fact that many tasks could be carried out at the same time invite to
design the multi-agent architecture outlined above. Its responsibilities include
publishing the e-catalogues that are sent by the SME, managing the e-catalogues
database, accepting the orders of the e-commerce service agents, and giving to the
personal agent the proper data to generate customised information pages.
3 Personalizing Capabilities in e-Commerce Server Site
The e-commerce site offers its clients efficient access and shopping management for
the different products that are published in the server. Customizing capabilities of the
site are very important to tailor all the kind of customrer preferences and they are
actually supported by the personal agent. Concretely, the implementation of this
personal agent is based on an adaptive agent-pattern derived from the adaptive agent
model proposed in [4], including a meta-level layer that gives each agent the ability to
take appropriate decisions about control or adapt the specific attributes of the system
over time to new circumstances. The user interacts with the e-commerce server that
presents the customized pages to offer relevant information to the client. The
hypermedia pages that represent the e-catalogues are generated on the fly, getting the
contents from the e-catalogues that were sent by the SMEs, which is the knowledge
layer.
anyPage : JSP Page :L a n g u a g e M a n a g e r se ssio n : H T T P S e ss io n

:C lie n t
re q u e s t g e tA ttrib u te L a n g u a g e ()
LanguageM anager
g e tT a g s ()
It searchs in the XM L
file the tags for the
requested page in the
chosen langu age
Fig. 3. DTD of the idiomatic XML file
Other interesting part of this adaptive process is the internalisation capability of the
server site. In this system there are not pre-compiled pages because all of them are
generated to support the chosen language by the end-user. To perform the dynamic
An Adaptive e-Commerce System Definition 509
generation of the language-independent pages there exists a database where the

elements that have language dependencies are stored. This mechanism is XML-based.
The idiomatic XML-file is hierarchically organised, and it has as many nodes as
languages supported by the e-commerce site. Each language includes the contents
shown by the system, these contents are classified by the elements that use them (JSP
pages in this case). Each JSP page has an object, called LanguageManager, which is
in charge of extracting the necessary elements for the requested page from the XML
file. The selected elements are expressed in the same language as the language of the
session. This scenario is depicted in Fig. 3.
4 Conclusions and Further Work
As it’s defined in [6], adaptable systems are systems that allow one to modify some of
their parameters and then adapt their behaviour accordingly. If the system adapts to
the user automatically it is called adaptive. According to this definition, in this paper
we have introduced an adaptive system for e-commerce proposal, a system that is
especially suitable for SME, called e-CoUSAL, an agent-based architecture for the e-
commerce system, and two main components are identified: a visual catalogue-
designer tool to generate e-catalogues and an e-commerce server site that stores the
generated catalogues.
Various types of adaptation can be supported by a system. Currently the server site
implements an adaptive agent that performs the interaction with the user, adapting the
content, the presentation and the navigation properties of the dynamically generated
information-pages from the e-catalogues stored in the server.
Further work is needed to implement the whole architecture, in which there are two
more adaptive agents: the negotiation agent and the customisation one. The definition
and implementation of these two agents are very important to achieve a flexible and
adaptive e-commerce system, because we have now an adaptive system in the
interface area, but with the new agents we will have customisation profiles that allow
a more usable and flexible system for the end-user, and, on the other side, the
enterprises.
References
1. Baron, J. P., Shaw, M. J., Bailey, A. D. Jr.: Web-based E-catalog Systems in B2B
Procurement. Communications of the ACM, 43(5):93-100. (2000)
2. García, F. J.: A B2B/B2C Hybrid E-Commerce Model. Submitted to SIGecom Exchanges,
Newsletter of the ACM Special Interest Group on E-commerce. (2001).
3. García, F. J., Moreno, Mª N., Hernández, J. A.: e-CoUSAL: An E-Commerce Architecture
for Small and Medium Enterprises. In Advances in Business Solutions. Catedral. (In
Press). (2002)
4. Guessoum, Z., Quenault, M., Durand, R.: An Adaptive Agent Model. In Proceedings of
AIB’S. (York, March 20-24, 2001). (2001)
5. Papazoglou, M. P.: Agent-Oriented Technology in Support of E-Business.
Communications of the ACM, 44(4):71-77. (2001)
6. Paterno, F., Mancini, C.: Designing Web User Interfaces Adaptable to Different Types of
Use. In Proceedings of Museums and the Web. (New Orleans, LA, USA). (1999)
Web Site Personalization Using User Profile Information
Mohit Goel1 and Sudeshna Sarkar2
1 IndianNavy, INSMA, Headquarters Western Naval Command,

Mumbai, India 400001
[email protected]
2
Dept. Of Computer Science Engineering, IIT, Kharagpur
West Bengal, India 721302
[email protected]
Abstract. In this paper we discuss a technique for web site personalization.

Connectivity analysis has been shown to be useful in identifying high quality
web pages within a topic or domain specific graph of hyper linked documents.
We have implemented a system that creates a view of a subset of a web site
most relevant to a given user. This sort of personalization is useful for filtering
a useful subset of a site so that the user gets a low volume of quality informa-
tion. The essence of our approach is to augment a previous connectivity analy-
sis with content analysis. We present an agent which assists the user when he
browses and distills a personalized sub graph of the website based on his user
profile.
1 Introduction
The Web is full of information and resources [1]. In this project we present a personal
agent that assists a user when he browses. Adaptive web pages help the user in find-
ing relevant information by tailoring their content and layout specific to the visiting
user.
Another motivation to develop such a system is for users with poor Internet con-
nectivity. Search engines often return top-level home pages, which do not have much
of content, and invariably a user needs to browse deeper to extract the relevant
information. Take the case of a school in rural India where Internet access can only
me made feasible through dial-up connection restricted to one or two hours in a day,
and the number of computers available are limited. In such scenarios, online browsing
is not feasible, and yet we will like to make available to the teachers and students of
such schools the power of the Internet. We can use our system to transmit offline
requests from the users to such sites, and build a system where a limited view of the
site can be packaged and sent to the user in off-line mode. Our approach is unique as
it attempts for ‘on the fly’ personalization which is different for each user. We model
a Web site as a graph. Each node of the graph as an individual URL and the links
embedded in the HTML are the edges between the nodes.

Web Site Personalization Using User Profile Information 511
2 Motivation of Our Work
The motivation for this algorithm stems from the fact that a user may like to have a
condensed or wrapped up view of the site or a collection of documents depending
upon the constraints or limitations he is dealing with. A typical constraint may be on
the bandwidth of Internet connection a user has. People who have poor Internet con-
nectivity often find online browsing of documents impractical due to slow downloads
and frequent dropping of connections. Many a times they have a window of good
connectivity available to them, which is not sufficient for online browsing. In these
cases, which are very relevant to rural areas of India, we can download the sites on
the client and view them offline. In this case a user may give her profile or a query
and she expects the result to be downloaded on her disk, which is connected. The
problem with a condensed user view of the site is that we get too many broken links
and the site is not navigable using the hyper links.
In our algorithm we take care of this problem explicitly and connect the sub graph
using a backward breadth first traversal. Another scenario is when a user desires ease
of navigation and he still wants the freedom to surf all the pages, which are there in
the site. This requires a slightly different adaptation where the good links are high-
lighted or annotated and uninteresting links are maintained as in the original graph.
3 System Architecture
The algorithm computes two scores for each document: a hub score and an authority
score. Documents that have high authority scores are expected to have relevant con-
tent, whereas documents with high hub scores are expected to contain links to rele-
vant content. We wish to get the high content pages and for that we need nodes with
good authority score. For reaching these nodes we need to traverse the sub graph
through the nodes with high hub score, which act as good junctions in the sub graph.
The algorithm works in two distinct phases. Phase 1 of the algorithm is termed as
preprocess which generates the profile of the web site or the document collection. The
algorithm first parses HTML to text but maintains information regarding various
Meta tags so as to enable us to give weights to HEAD, TITLE etc. For each node in
the site we store the pointer to the TFIDF value, which is a vector representation in
our system. Hub and Authority scores are calculated using a simplified version of
Kleinberg’s algorithm [4] for finding hubs and authorities.
We define a good node based on four heuristics. We define a good hub score and
check the hub score of each node against this score and in case it is more than the
good hub score, it is considered as a good node irrespective of the fact whether the
node itself does not contain relevant information for the user. The heuristic is that
good hubs in any case act as focal points for navigation and in general improve the
ease of browsing for a given website.
Similarly we define a good similarity measure and check the cosine similarity [8]
score of user profile with each node in a site and in case the similarity is more than
512 M. Goel and S. Sarkar
this measure, it is obvious that this particular node is very relevant to the user’s inter-
est and is therefore included as a good node.
Our system identifies the relevant content pages in a site. The system then attempts
to connect these pages together with the home page to make available to the user. The
problem is that the pages selected thus may not be all connected from the home page.
Sometimes, it happens that even if a page is reachable from the home page, its depth
is very high. The system first identifies non-reachable nodes, and then attempts to
find a set of hub pages that are each ancestors of a set of these unreachable pages. By
including these extra pages we make all the pages reachable. In case where we do not
find appropriate link pages, we separately display these pages. We plan to use
unsupervised clustering to group these unreachable pages.
One of our experimental sites was http://www.Bharat-rakshak.com. This site is an

unofficial website for Indian Armed Forces. The front page of the site is broken into
sections corresponding to the main organization of the site: Indian Navy, Indian
Army, Indian Air Force, Indian Space, Indian Missiles, Famous battles, Recruitment
information and general information etc. For this experimentation we assumed that a
user is interested in Indian Navy and in past has shown a keen interest in the Navy
section of the site. This is represented as a directory, which consists these files. Once
the system is executed, a user profile is generated using the various files in this direc-
tory and a similar profile vector is generated for each node of the site.
Observation on Pages Selected as Good Nodes. The majority of the pages given
high ranking were from the sub graph, which contained nodes dealing with Navy. All
the starting nodes of various arms were also selected as they had many outgoing links
giving them a higher hub score. A lot of pages from the ‘Missile’ and ‘ Air Force’
section were also selected. On closer scrutiny it was found that these pages had re-
lated information and matching keywords with the ‘Navy’ pages such as information
on Naval version of missiles, about maritime patrolling by Air Force (A joint opera-
tion carried out by Navy and Air Force) etc.
Observation on Pages Not Selected as Good Nodes. The pages, which were not
selected as Good Nodes, were found mostly from the unrelated pages and pages hav-
ing more images and less content (Although some of these pages were dealing with
Navy). This was primarily due to low cosine similarity of user profile with these
pages.
Web Site Personalization Using User Profile Information 513
5 Conclusion
We have developed a system that adapts a website based on the user profile. We also
carried out experimentation on various websites with different user profiles and found
the results in conformance to most of our heuristics about deciding a good node for
the user. This agent can be run at the user’s machine, the web server or at the proxy
server based on the requirement and connection bandwidth, web server load etc. The
client side implementation of our system can be integrated with a dial up connection
where a user gives an initial set of query words and the search is further refined using
an auxiliary set of query words (User Profile in this case). This filtering will get us a
subset of the original website or domain which then be connected as explained in
previous chapters and a coherent browsable personalized sub graph can be either
mailed or send via normal HTTP protocol to the user.
At present the system displays all the nodes in a frame and also when shortcuts are
generated. The system can also be further enhanced using supervised and unsuper-
vised clustering methods, which can cluster the results displayed in the adaptive
frame.
Using machine-learning algorithms for automatically processing web server logs
and generating user profiles can further augment the system. At present we feed the
user profile explicitly by giving a collection of documents. This user profile can be
generated by the system by data mining on web logs for the server side implementa-
tion of this system.
References
1. Etzioni, O., Perkowitz, M., Adaptive Web Sites: an AI Challenge, In Proceedings of the
Fifteenth International Joint Conference on Artificial Intelligence, 1997.
2. Etzioni, O., Perkowitz, M., Adaptive Web Sites: Automatically Synthesizing Web Pages,
AAI, 1998.
3. Lieberman, H. (1995). Letizia: An agent that assists web browsing. In International Joint
Conference of Artificial Intelligence, Montreal, Aug.1995.
th
4. Kleinberg, J. 1998. ‘Authoritative sources in a hyper linked environment.’ Proc. Of 9
ACM-CIAM Symposium on Discrete Algorithms.
5. Kramer, Joseph et al. ‘A User-Centered Design Approach to Personalization’. CACM.
August 2000. Vol. 43 No. 8. pp. 45-48. ACM: 2000.
6. Balabanovic, M., Shaham, Y.. 1995, Learning Information Retrieval Agents: Experiments
with Automated Web Browsing. Proceedings of the AAAI Spring Symposium Series on
Information Gathering from Heterogeneous, Distributed Environments: 13-18.
7. Porter, M.F. (1980) An algorithm for suffix stripping. Program; Automated Library, and
Information Systems, 14(3), 130-137.
8. Salton, G. (1992). The state of retrieval system evaluation. Information processing &
Management,8(4): 441.
Who Do You Want to Be Today?
Web Personae for Personalised Information Access
J.P. McGowan, Nicholas Kushmerick, and Barry Smyth
Smart Media Institute, Department of Computer Science, University College Dublin

{jp.mcgowan,nick,barry.smyth}@ucd.ie
Abstract. Personalised context sensitivity is the Holy Grail of web information

retrieval. As a first step towards this goal, we present the Web Personae person-
alised search and browsing system. We use well-known information retrieval
techniques to develop and track user models. Web Personae differ from previ-
ous approaches in that we model users with multiple profiles, each correspond-
ing to a distinct topic or domain. Such functionality is essential in heterogene-
ous environments such as the Web. We introduce Web Personae, describe an
algorithm for learning such models from browsing data, and discuss applica-
tions and evaluation methods.
1 Introduction
Despite recent advances in Web information retrieval technologies (e.g. [1,2]), Web
search services will find it increasingly difficult to return relevant and valuable results
unless they deploy mechanisms for delivering personalised context-sensitive results.
As a step towards this goal, we introduce Web Personae, a simple method for devel-
oping web user models, and describe several applications that use Web Personae to
deliver personalised context-sensitive search results.
Web Personae are designed to address a long standing issue in personalised infor-
mation filtering: people often have multiple information needs, and attempting to
model a user with a single monolithic profile can lead to poor retrieval accuracy.
Consider the following scenario. Michelle has a variety of interests: she is a medical
doctor, and enjoys golf, plays computer games, and regularly visits the theatre. Given
a list of Michelle’s favourite web pages or other aspects of her browsing history that
reflect these interests, a Web Personae system should automatically discover that
Michelle can be modelled by distinct personae such as Golf, Games, Theatre and
Medical. Furthermore, given these models and a sample of her current browsing be-
haviour, the Web Personae system should recognise which persona is currently active,
and personalise her information access. For example, if Michelle browses on pages
with words such as ‘green’ and ‘tee’, then the Web Personae system should recognise
that her active topic is Golf.
Web Personae enable a variety of personalised information access applications.
First, as a user surfs, several kinds of adaptive hypertext applications could dynami-
cally transform the HTML of web pages. The most obvious application is web page
recommendation: based on recent pages, we can recommend either links on the cur-
rent page, or new pages entirely. For example, if a user is interested in Golf, then,
when they visit a generic sports site, the application could highlight links specifically

Who Do You Want to Be Today? Web Personae for Personalised Information Access 515
relevant to golf. Similarly, if the user then goes to a search engine, it would be possi-
ble to highlight the results returned for their query that more closely match their per-
sonal needs – based on their estimated web persona. A more sophisticated method
would involve a full re-ranking of the results based on the current persona, so that
relevant results are ranked higher. For example, the application could apply a modi-
fied version of the PageRank [1] or HITS [2] topology-based algorithms that weights
pages according the similarity to the current persona, before calculating the topology-
based ordering. Finally, another application is query expansion based on high-weight
terms in the current persona. These applications demonstrate a variety of ways in
which Web Personae can deliver customised context-sensitive information access.
Several of these have been implemented.
2 Web Personae Construction and Recognition
The three main components of the Web Personae system are the Constructor, which
learns the personae, the Recogniser, which estimates which persona is currently ac-
tive, and the Application, which uses this to provide personalised context-sensitive
hypertext adaptation.
Personae Construction
The Constructor component uses hierarchical clustering techniques over web page
content. This content is initially provided in the form of a list of frequently visited
pages, browsing history, or bookmarks. Given these URLs for a given user, the Con-
structor clusters the web pages using the standard TF-IDF cosine similarity metric. In
order to automatically discover a user’s distinct personae, the clustering process is
halted when the ratio of intra-cluster similarity to inter-cluster similarity has reached a
maximum. The clusters learned by this process are assumed to represent the user’s
several Web Personae. Below we describe some experiments we intend to run to
evaluate this assumption. As well as this offline functionality, the Constructor has a
greedy, incremental mode in which the personae are modified to track preference
changes. However, due to noise and transient browsing behaviour we have found that
this online mode poses several thorny user-interface challenges.
Personae Recognition
Once the personae have been identified in the offline stage, we must utilise these
when the user is online. An essential design requirement is that the user should not
have to explicitly indicate which persona is currently active; the system should be
able to infer the current persona based on user actions. Furthermore, this inference
must be made rapidly as the user surfs from page to page.
The Recogniser component uses a simple and efficient similarity estimate. We
convert the centroids of the personae clusters and the current pages into term vectors
that captures only the word frequencies, without taking IDF into account. We then
select the Web Persona that has the largest cosine similarity with the current docu-
ment. This gives us a quick persona recognition system, which has worked well in
preliminary experiments.
516 J.P. McGowan, N. Kushmerick, and B. Smyth
3 Related Work
Document clustering has mainly been used in information retrieval for improving the
effectiveness and efficiency of the retrieval process. We utilise automatic clustering to
reveal different domains of user interest. The documents are represented in a vector
space [3], then we use hierarchical agglomerative techniques to produce a cluster tree.
Web document clustering has been extensively researched in recent years [4,5,6,7].
Applications range from bookmark organising to recommendation systems.
Many systems have been developed which assist web browsing. Letizia [8] learns
the interests of a user by observing their browsing behaviour - it can then recommend
links to follow - i.e. it models the browsing process, rather than explicitly modelling
the user, as our system does. WebWatcher [9] takes some user interests as an initial
input, then updates these interests based on pages they visit. The system then recom-
mends pages, based on these interests and the previous browsing behaviour of other
users with similar interests.
Various systems have been developed which utilise user models for personalisa-
tion. WebMate [6] is an agent that assists browsing and searching. It represents differ-
ent domains of user interest using multiple term vectors - it updates these incremen-
tally when users give positive feedback for visited pages. However, it does not cluster
the vectors to produce a 'persona' as in our system. WebACE [5] constructs a custom-
ised user profile by recording information about the documents the user browses. It
then clusters these documents, using novel clustering techniques, and uses these to
generate queries to search for similar documents.
Personal View Agent [10] tracks, learns and manages user interests. Beginning
with a fixed palette of categories, the system follows the user, detecting their domains
of interest. This ‘personal view’ takes the form of a tree and corresponds closely to
our notion of web personae. This view can be updated - i.e. it can adapt to changing
user interests using a 'personal view maintainer', which can split and merge categories
in the personal view.
4 Applications and Evaluation
We have introduced the notion of Web Personae, discussed how they enable personal-
ised context-sensitive information access, and described how they can be automati-
cally learned and recognised from browsing behaviour using standard information
retrieval techniques.
We are currently conducting an empirical evaluation of the learning and recogni-
tion components. Preliminary tests indicate good performance but we are currently
designing more sophisticated evaluations. One experiment involves looking at logs
for servers that provide a local search facility. Using a user's accesses to the server
obtained from the web logs, we build their personae. We then look at click-through
data from local searches run by that user to estimate how effective the Web Personae
system would have been at re-ranking these search results.
We have built some applications for the system, such as the search result reorderer
(see Fig. 1), and are currently building more, such as the page recommender service
and web query expansion service discussed earlier. Our main application will be a
personalised search service, based on both result-persona similarity, and the more
Who Do You Want to Be Today? Web Personae for Personalised Information Access 517
sophisticated system that does a full re-ranking based on link analysis techniques.
(Acknowledgements: This research was funded in part by grant N00014-00-1-0021
from the US Office of Naval Research.)
Fig. 1. Search Reorderer Application - background window indicates active persona,

foreground shows result of reranking based on this active persona. Details and a
demonstration of the system are online at http://www.smi.ucd.ie/WebPersonae.
References
1. S. Brin and L. Page. The anatomy of a large-scale hypertextual web search engine.
In Proc. 7th International World Wide Web Conference, 1998.
2. D. Gibson, J. Kleinberg & P. Raghavan. Inferring Web communities from link topology.
In Proc. ACM Conf. Hypertext & Hypermedia, 1998.
3. G. Salton & M. McGill. Introduction to Modern Information Retrieval. McGraw-Hill,
1983
4. D. Cutting, D. Karger, J .Pederson, J.Tukey. Scatter/Gather: A Cluster-Based Approach to
Browsing Large Document Collections. In Proc. SIGIR, 1992.
5. E. Han, et al. WebACE: A web agent for document categorization and exploration. In
nd
Proc. 2 Int. Conf. on Autonomous Agents, 1998
nd
6. L. Chen, K.Sycara. WebMate: A personal agent for browsing and searching. In Proc. 2
Int. Conf. on Autonomous Agents, 1998
7. Y. Maarek, I. Ben Shaul. Automatically Organising Bookmarks per Contents. In Proc.
WWW5, 1996.
8. H.Lieberman. Letizia: An agent that assists web browsing. In Proc. IJCAI-95, Montreal,
Canada
9. T. Joachims, D. Freitag & T. Mitchell. WebWatcher: A tour guide for the World Wide
Web, In Proc. IJCAI-97, Nagoya, Japan
10. Chien Chin Chen, Meng Cheng Chen, Yeali Sun. A Web Document Personalisation User
Model and System. In Proc. User Modelling, 2001.
Adaptive Navigation Path Previewing
for Learning on the Web
Akihiro Kashihara, Shinobu Hasegawa, and Jun’ichi Toyoda
The Institute of Scientific and Industrial Research, Osaka University,

8-1, Mihogaoka, Ibaraki, Osaka 567-0047,
[email protected]
http://www.ai.sanken.osaka-u.ac.jp/thome2/kasihara/index-e.html
Abstract. The main issue addressed in this paper is how to help learners plan a
navigation path in existing web-based learning resources, which is an important
process of self-directed learning in hyperspace. Our approach to this issue is to
provide learners with the adaptive preview of a sequence of web pages as navi-
gation path. Following the idea of path previewing, we have developed an as-
sistant system. The system displays an overview of a web page selected by
learners from a hyperspace map, by extracting information from the HTML
document file related to the navigation path-planning context. It also enables
learners to transform a sequence of previewed pages into a navigation path
plan.
1 Introduction
Existing web-based learning resources provide learners with hyperspace where they
can navigate through the web pages in a self-directed way to learn domain con-
cepts/knowledge. The self-directed navigation involves making a sequence of the
pages navigated, which is called navigation path [2]. It also involves constructing
knowledge from the contents embedded in the navigated pages, which would enhance
learning [4].
On the other hand, learners often fail in making a navigation path [4]. The self-
directed learning in hyperspace requires them to comprehend the contents of web
pages they have visited, and concurrently to monitor their own navigation process
such as planning the navigation path to be followed, which can be viewed as meta-
cognitive activities in self-directed learning [1,2,3]. Navigation monitoring holds the
key to success in self-directed learning. However, it is difficult to maintain navigation
monitoring since the learners would focus on comprehending the contents of the
visited pages [2].
This paper proposes navigation path planning with adaptive previewing of the
navigation path. The key point of this idea is to provide learners with an adaptive
preview of a sequence of web pages. The path previewing enables them to plan the
navigation path in a learner-centered way before navigating the web pages.
P. De Bra, P. Brusilovsky, and R. Conejo (Eds.): AH 2002, LNCS 2347, pp –.

Adaptive Navigation Path Previewing for Learning on the Web 519
We have developed an assistant system, which consists of path previewer, page

previewer, and hyperspace map [3]. The page previewer extracts information attached
to some HTML tags in a web page, which learners select in the hyperspace map, from
the HTML document. The information to be extracted depends on the topic on which
the learners focus in planning the navigation path. The path previewer makes a se-
quence of previewed pages, and displays it as navigation path preview. These facili-
ties help learners plan which page to visit without visiting real web pages in hyper-
space.
2 Navigation Path Planning
Self-directed learning with navigation path planning is done in two spaces, which are
the space for planning the navigation path based on a learning goal and hyperspace
for executing the plan. In the path planning space, learners preview web pages, and
plan which page to visit and the sequence of pages to be visited so that the learning
goal can be achieved. In the hyperspace, they are expected to navigate web pages
according to the path planned. The path planning and navigation are repeated during
learning in hyperspace.
Figure 1 shows a user interface of the assistant system where learners can preview
a navigation path in a learner-centered way. The system is composed of spatial map,
page previewer, and path previewer. The spatial map represents the hyperspace of a
web-based learning resource selected by learners as a network of nodes correspond-
ing to web pages. It is automatically generated and displayed in the map window
when they select the learning resource. Nodes in the spatial map are tagged with page
titles indicated by title tags in the HTML document files.
By double-clicking any node in the spatial map, learners can have an overview of
the web page corresponding to the clicked node in the page preview window, which
is generated by the page previewer. Assuming that such information is located within
representative HTML tags such as Title, Headings, etc., the page previewer can ex-
tract words, sentences, images or links indicated by these tags [3].
When the learners decide the starting point of the navigation path, they can trigger
the path previewer, which changes the page preview window into the path preview
window. The path previewer constructs a sequence of previewed pages starting with
the current page, which represents a navigation path. The path preview window has a
link list, which includes anchors of the links the current page contains. Selecting any
one from the list, they can have an overview of the page, to which the selected link
points, next to the preview of the current page, and then put the previewed page into
the sequence, making a navigation path. The page preview is adapted to the context
of planning the navigation path. The page preview generation consists of the follow-
ing two processes:
(1) Identifying the topic on which learners focus in planning, and
(2) Identifying information to be previewed according to the focal topic.
520 A. Kashihara, S. Hasegawa, and J. Toyoda
(b) Page preview window
(a) Spatial map window
(c) Path preview window
Link list
Fig. 1. User Interface for Navigation Path Planning.
The path previewer first identifies the focal topic from keywords included in the
anchor that learners select from the link list of the current page for previewing the
next page. Next, by using the Heading tags of the HTML document of the page to
which the selected anchor points, the page previewer divides the document into sec-
tions, which are indicated by the tags. Among these sections, it identifies a section
whose Heading tags include the focal topic, as the section to be previewed. If no
Heading tag includes the focal topic, it finds in which section keywords representing
the focal topic appear most frequently and selects this as the section to be previewed.
The page previewer next extracts information to be previewed to generate the
adaptive page preview. It first extracts not only information attached to the Heading
Adaptive Navigation Path Previewing for Learning on the Web 521
tag of the identified section, but also information attached to Title tags, and the ances-
tor/ descendant Heading tags. This enables the page preview to preserve the hierar-
chical structure of topics embedded in the page. The page previewer then extracts key
information from the identified section, that is attached to the representative HTML
tags, such as FontSize/Color/Face, Img, and A href tags.
Let us illustrate this with a simple example. In this example, a learner whose learn-
ing goal is to explore factors reducing the reliability of computer networks first tries
to plan a navigation path for achieving the goal from the web page Purpose and Ap-
proach of Reliability Design. The starting page in the navigation path is generated as
shown in Fig. 1(b). Next, he/she selects the anchor Unauthorized Acts from the link
list in the path previewer to have a page preview of the web page Information Secu-
rity of Network Systems, to which the anchor points. In this case, he/she seems to
focus on unauthorized acts. The path previewer accordingly identifies the focal topic
with Unauthorized Acts, and then generates the adaptive preview of the page as
shown in Fig. 1 (c), which includes proper information for the focal topic.
3 Conclusions
This paper has described navigation path planning system with an adaptive preview-
ing function of the navigation path in a hyperspace of web-based learning resources.
The distinction between navigation path planning and navigation allows learners to
raise their awareness of monitoring their navigation process, and to focus on compre-
hending the contents of the learning resources in hyperspace. Since the navigation
path planned also gives learners an overview of the contents to be learned before
navigating the hyperspace, their learning can be improved.
References
1. Cunninghan, D.J., Duffy, T.M., and Knuth, R.A. The Textbook of the Future, in
McKnight, C., Dillon, A., and Richardson, J. (eds): HYPERTEXT A Psychological Per-
spective (1993), Ellis HorwoodLimited, 19-49.
2. Hammond, N. Learning with Hypertext: Problems, Principles and Prospects, in
McKnight, C., Dillon, A., and Richardson, J. (eds): HYPERTEXT A Psychological Per-
spective (1993), Ellis HorwoodLimited, 51-69.
3. Suzuki, R., Hasegawa, S., Kashihara, A., and Toyoda, J. A Navigation Path Planning
Assistant for Web-based Learning, Proc. of World Conference on Educational Multime-
dia, Hypermedia & Telecommunications (2001), 851-1856.
4. Thuering, M., Hannemann, J., and Haake, J.M. Hypermedia and Cognition: Designing for
Comprehension. Communication of the ACM, 38, 8 (1995), ACM Press, 57-66.
A Case-Based Recommender System
Using Implicit Rating Techniques
Youngji Kim1 , Sooho Ok2 , and Yongtae Woo1

1
{yjkim@cdcs,ytwoo@sarim}.changwon.ac.kr
2
Kosin University, Yeongdo-gu Pusan 606-701, Republic of Korea,
[email protected]
Abstract. We propose a new case-based recommender system using im-

plicit rating information. We present intra-attribute and inter-attribute
weight derived from past interests of a user stored in the access logs,
and a new similarity function to estimate similarities between new items
set and the user profile. To verify the efficiency of our system, we have
performed experimental comparisons between the proposed model and
the collaborative filtering technique by mean absolute error(MAE) and
receiver operating characteristic(ROC). The results show that the pro-
posed model is more efficient than the traditional collaborative filtering
technique.
1 Introduction
A recommender system is an automated information filtering system to rec-
ommend personalized items that are identified by user’s preferences. There are
several approaches to develop a recommender system using collaborative filter-
ing(CF), content-based, and case-based reasoning techniques. First, CF predicts
a user’s preference based on the similarity between the rating pattern of the user
and other users, but it has several problems such as early rater, sparsity, and
gray sheep. Content-based method uses information about the item itself to make
suggestions, but the method does not account for community endorsements [1].
Some researchers have exploited hybrid CF and content-based recommendation
models to complement the disadvantages of the both [2]. On the other hand,
case-based reasoning technique regards items in which a user is interested as the
cases and recommends the most similar cases compared to the past historical
interests when a new case presents to a user [3]. Individual item preferences can
be supplied explicitly by users or collected implicitly by analyzing access logs.
Recently, several implicit rating techniques have been proposed to collect infor-
mation automatically from the user’s behavior patterns in the access logs and
to predict an individual preferences [4].
In this paper, we propose a personalized recommender system based on case-
based reasoning approach using implicit rating technique. We present a new
implicit rating method based on intra-attribute and inter-attribute weights de-
rived from past preferences of a user stored in the access logs. The proposed

A Case-Based Recommender System Using Implicit Rating Techniques 523
system has been tested using the access logs of the users in a recruiting site.
The results show that the proposed system is more efficient than the typical CF
technique in [2].
2 The Proposed Recommender System
Conceptual Architecture. In a traditional case-based reasoning system, when

a new problem appears, the system retrieves the most similar case(s), reuses the
case(s) to solve the problem, revises the proposed solution if necessary, and re-
tains the new solution as a part of a new case [5]. We have adopted the principles
of case-based reasoning for our system. In order to recommend new items to a
user, the system suggests the recommendation of the most similar ones using
the past preference information stored in the user profile, and the result of the
recommendation is updated in the user profile for learning the new case. Figure 1
shows the conceptual diagram of the proposed system.
G
User New
Similarity
Profile Items
Recommendation
Profile Item
Database Database
User Access
Implicit Rating
Access
Logs
Fig. 1. Conceptual diagram of the proposed system
User Profile and Implicit Rating. We define a user profile as P = {u, A, W ,

D}, where u is the personal information about the user; A is a set of attributes
of an item; W is a set of intra-attribute weights; and D is a set of inter-attributes
weights among attributes. The weights W and D are automatically computed
by analyzing each user’s access logs.
Intra-attribute Weight. We assume that an item has n attributes A1 ,A2 ,· · ·,
An , and each attribute Ai has m attribute values ai1 ,ai2 ,· · ·,aim . Let Wi =
{wi1 , wi2 , · · · , wim } be the intra-attribute weight corresponding to an attribute
Ai , i = 1, 2, · · · , n. We define intra-attribute weight wij corresponding to aij by
kij
wij = m , i = 1, 2, · · · , n, j = 1, 2, · · · , m, (1)
p=1 kip
where kij is the access count to aij of the user.
524 Y. Kim, S. Ok, and Y. Woo
Inter-attribute Weight. Inter-attribute weight is a set of statistical coefficients

D = {d1 , d2 , · · · , dn } corresponding to A. Let xij be the probability of access to
aij among ai1 ,ai2 ,· · · ,aim , then inter-attribute weight di of Ai is defined as
follows:
m
1

di = yi − xij , yi = max (xij ), i = 1, 2, · · · , n (2)
m j=1 1≤j≤m
For example, a user looking for a job position in a recruiting site. Here, we
have assumed that the major factors to select a job are the academic major, job
type, and location attributes. If intra-attribute weights in the user profile are
given as follows: major={database:0.8, network:0.2}, job type={professor:0.6,
researcher:0.3, Post-Doc.:0.1}, and location={Seoul:0.2, Daejeon:0.3, Pusan:0.5}
by equation 1. Thus, the inter-attribute weights will be 0.30, 0.27, and 0.17
respectively by equation 2. In case of di equals to zero, the attribute is ignored.
Therefore, we can ensure that the major attribute is the most important one for
the user in this case.
Similarity Computation and Recommendation. When new items are pre-
sented to the system, the similarity between the new items and each user’s pref-
erences is evaluated by calculating intra-attribute and inter-attribute weights for
each user’s profile. The results of the recommendation are presented in the sim-
ilarity order. A new item set I contains explicit information about the specific
field, and is defined as I = {A }, where A is a set of attributes as such attribute
Ai = {ai1 , ai2 , · · · , aik }, and they have k attribute values. The domains of at-
tributes in user profile are the same as the domains of attribute Ai in new item
set correspondingly. We define a new similarity function between a user profile
P and a new item set I as follows:
n

Similarity(P, I) = (f (aij , aij ) × di × Wi ), (3)
i=1
where f (aij , aij ) is a filtering function which returns to 1 if aij and aij match
one another.
In the above example, three new job positions as job1 = <major: database,
job type: researcher, location: Seoul>, job2 = <major: network, job type: pro-
fessor, location: Pusan>, and job3 = <major: algorithm, job type: professor,
location: Pusan> are presented. Then, the similarities between the user profile
and three new jobs are evaluated as follows: job1: 0.355, job2: 0.307, and job3:
0.247. Therefore, the system recommends these jobs as job1, job2, and job3 in
order to the user.
To verify the effectiveness of the proposed system, we have tested the users of a
recruiting site using mean absolute error(MAE) and receiver operating charac-
teristic(ROC) [6]. MAE and ROC are commonly used to measure the statistical
A Case-Based Recommender System Using Implicit Rating Techniques 525
Table 1. Experimental results
MAE ROC-sensitivity ROC-specificity ROC-accuracy ROC-error rate

CF 0.923 0.667 0.210 0.701 0.299
Proposed 0.818 0.900 0.424 0.809 0.191
accuracy and decision-support one of recommendations respectively. First of all,

we have gathered the explicit preferences grade from 1 through 5 for 1,030 con-
tents and 824 users. We have then collected 14,263 implicit ratings from the
access logs for them. In the experiment, 1,484 recommendations have been used
as test set by the proposed method and CF algorithm in [2]. In general, the
lower values of MAE, ROC specificity and ROC error rate, or the higher values
of ROC sensitivity and ROC accuracy means the better results of recommen-
dation. Table 1 shows the comparisons of the MAE and ROC(by sensitivity,
specificity, accuracy, and error rate) values between the proposed method and
CF. Experimental results show that the proposed method is more efficient than
CF one.
4 Conclusion
In this paper, we proposed a personalized recommender system to recommend
specific items to a user based on case-based reasoning approach using implicit
ratings. We presented two weights, intra-attribute and inter-attribute weight,
related to attributes of items. These weights are evaluated implicitly from access
logs of each user. A new similarity function using those coefficients have proposed
to estimate similarities between new item set and the user profile. MAE and ROC
values in the experimental results show that the proposed system makes more
efficient recommendation than traditional CF technique.
Acknowledgement
This work was supported by grant No. 2001-111-2 from the University funda-
mental research program of the Ministry of Information & Communication in
Republic of Korea.
References
1. Popescul, A., Unger, L., Pennock, D., Lawrence, S.: Probabilistic Models for
Unified Collaborative and Content-Based Recommendation in Sparse-Data En-
vironments. Proceedings of the Seventeens Conference on Uncertainty in Artificial
Intelligence(UAI-2001), Morgan Kaufmann, San Francisco (2001) 437–444
2. Claypool, M., Gokhale, A., Miranda, T., Murnikov, P., Netes, D., Sartin, M.: Com-
bining Content-Based and Collaborative Filters in an Online Newspaper. Proceed-
ings of ACM SIGIR Workshop on Recommender Systems, August 19 (1999)
526 Y. Kim, S. Ok, and Y. Woo
3. Bradley, K., Rafter, R., Smyth, B.: Case-based user profiling for content personal-
isation. Proceedings of the International Conference on Adaptive Hypermedia and
Adaptive Web-based Systems, Trento, Italy, August (2000) 62–72
4. Oard, D., Kim, J.: Implicit Feedback for Recommender Systems. Proceeding of
AAA Wrokshop on Recommender Systems, Wisconsin, USA, July (1998) 81–83
5. Watson, I.: Applying Case-based Reasoning: Techniques for Enterprise Systems.
Morgan Kaufmann Publishers, Inc. (1997)
6. Good, N., Schafer, B., Konstan, J., Borchers, A., Sarwar, B., Herlocker, J., Riedle,
J.: Combining Collaborative Filtering with Personal Agents for Better Recommen-
dations, Proceedings of the AAAI conference. (1999) 439–446
IMAP - Intelligent Multimedia Authoring Tools
for Electronic Publishing
Sarit Kraus1 , Alexander Kröner2 , and Lea Tsaban1

1
Department of Mathematics and Computer Science
Bar-Ilan University, Ramat-Gan, 52900, Israel
{sarit,lea}@macs.biu.ac.il
2
German Research Center for Artificial Intelligence (DFKI) GmbH
Stuhlsatzenhausweg 3, 66123 Saarbrücken, Germany
[email protected]
Abstract. IMAP provides software tools that support the authoring of

electronic presentation by helping the author in multimedia content se-
lection and layout design. IMAP consists of a Content Manager and a
Layout Manager. In this paper we present the usage of IMAP tools for
the authoring of on-line newspapers. The content selection is based both
on the author’s specifications and the user’s interests, where the interre-
lations between objects play an important role in the evaluation of the
set of objects. Layout management relies on a set of layout requirements
taken from a layout profile and an author-defined style sheet. As was
demonstrated in our experiments, integrating of both techniques yields
interesting newspapers, whose layout can be customized by the reader.
1 Introduction
The long-term goal that we address in IMAP is the development of software

tools for content and layout management that will simplify and accelerate the
creation of electronic presentation by automating the complex, time-consuming
tasks of multimedia content selection and layout design. In this contribution, we
focus on the authoring of personalized online newspapers.
Personalization of presentations has become a main issue in recent years
(e.g., [1,2]). Part of this research focuses on electronic access to news (e.g., [5]).
However, most of these works focus on identifying the user’s fields of interest,
i.e. the construction of a user profile. Then the articles are scored using the user
profile and the highest scored documents are included in the newspaper. We focus
on the interest of the user in the overall newspaper while enabling the author
to influence the content of the newspaper via constraints, and maintaining the
satisfaction of the user.
An appropriate layout for the personalized content selection should reflect
the user’s personal interests as well. From the manifold approaches to automated

This research was supported by a Grant from the G.I.F., the German-Israeli Foun-
dation for Scientific Research and Development.

528 S. Kraus, A. Kröner and L. Tsaban
layout (e.g., [8,4,9]), we chose constraint techniques, which had already turned
out to be a valuable means for creating Web pages (cf. [3,6]). We show how
to extend such an approach so that the layout reflects constraints imposed by
the given browsing environment, and may further be customized by the user
according to his preferences.
2 Content Management
The input of the Content Management (CM) system includes: (i) A set of can-
didate documents for the newspaper. (ii) A set of constraints on the selection
of documents provided by the author (editor) of the newspaper, e.g., “At least
a certain percent/number of the documents should be from a certain subject
and sub subject.” (iii) A reader profile: The profile of preferences of the reader.
(iv) A number K: The number of documents the newspaper should include. The
output of the CM is a list of K documents for the newspaper, ranked according
to the level of the reader’s interest.
The evaluation of a set of documents depends on the estimation of the agent’s
interest in the documents and the measure of the satisfaction of the set of doc-
uments of the author’s constraints. We score each document according to the
reader profile, using the correlation between the two. The similarity between a
reader profile and a document profile is calculated by referring to the keywords
that appear both in the reader’s profile and the document’s profile (see [7]).
Given a set of K documents, and a set of constraints we compute the satisfac-
tion measure of each of the constraints by the documents. The score assigned
to a constraint increases as the set of chosen documents comes closer to fully
satisfying the constraint. We use the following function to predict the reader’s
satisfaction from a setof documents: Eval(constraints, K documents) = P ar ·

i wi + (1 − P ar) · j cj · vj where: (i) P ar is a number between 0 and 1,
indicating the level of importance we give to the reader’s preferences. This pa-
rameter is determined by the author. (ii) i goes through the K documents in the
set considered for the newspaper. (iii) wi is the normalized relevancy measure
of the i’th document according to the reader profile. (iv) j goes through all the
constraints. (v) cj is the weight of the j’th constraint (vi) vj is in the interval
[0, 1], according to the level of satisfaction of the j’th constraint.
We asked 31 readers to use the software we developed, with 1225 documents
we downloaded from CNN (see http://www.cnn.com/). We built reader profiles
for the readers by explicitly asking about their interests. By checking the similar-
ity between each pair of readers we verified that the readers were not too similar.
Each reader received three different newspapers, one with no constraints, and
the other two with two different sets of constraints.
The first set: (i) At least 1 document from the subject “World” and the
sub subject “Middle East” should appear. Importance: 6 (ii) No more than 20%
of the documents will be from each subject. Importance: 6 (iii) Each document
that appears will not be lower than a relevancy of 0.004. Importance: 1 (i) At
least one document from each subject shall appear. Importance: 10
IMAP - Intelligent Multimedia Authoring Tools for Electronic Publishing 529
The second set: (i) At least 1 document from the subject “World” and the
sub subject “Middle East” should appear. Importance: 10 (ii) At least 10% of
the documents should be from the subject “US” and the sub subject “News”.
Importance: 8 (iii) The documents that appear will not be lower than a rele-
vancy of 0.004. Importance: 1 (iv) The documents that appear will not exceed
a similarity of 0.7. Importance: 10
For each newspaper, we asked the readers to give each document a score from
1 to 5, where: 1 means “not interesting at all” and 5 means “very interesting”.
We also asked the reader to answer the following questions about the newspaper
in the same manner: (i) How interesting did you find the newspaper? (ii) How
varied did you find the newspaper? (iii) How good did you find the newspaper?
(according to the way the reader defines “a good newspaper”).
The answers of the readers show that the first set of constraints is better than
the second set. Our results show that with respect to all the measurements, the
first set of constraints is significantly better than the second set and also better
than the “empty” set. We also checked which portion of the documents received
a score higher than 3 from the reader, i.e. “interesting” or “very interesting”.
We found that the documents of the newspaper with the good set of constraints
received significantly higher scores than the documents of the other two news-
papers. Significance was tested using the one-tailed paired t-test with p = 0.05.
3 Layout Management
The Layout Management (LM) aims at achieving a layout for the selected con-
tent, which meets a set of layout requirements imposed by an author-defined
style sheet as well as by a layout profile, which describes the reader’s technical
requirements and preferences.
In IMAP, layout is created by the DesignComposer (DC), which computes
layout for a given XML file, a layout profile and a style sheet (for details, see
http://www.dfki.de/˜kroener/DC.htm and [6]). This software is embedded in a
layout server, which provides the reader with a portal to the documents selected
by the CM. The server uses the DC to create layout in two steps: first, the
layout of the portal page is created. Then, when a user tries to access one of
the offered documents, a layout for the requested document is created on-the-fly
before delivery. If the creation fails, e.g., because of exceeding the generation
time bound, the non-optimized XML file (which is in IMAP an XHTML file) is
provided as fall-back solution.
The layout profile is created by a so-called Client Analyzer (CA), which is
implemented as an applet that is submitted by the layout server to the client
side. In the first place, the applet collects layout-related data, such as the browser
window’s dimension, the installed fonts, and the display’s DPI. Furthermore,
the CA enables the reader to communicate with the layout server and thus to
customize layout. Therefore it provides a GUI, which enables several kinds of
interaction.
530 S. Kraus, A. Kröner and L. Tsaban
Fig. 1. Two layout alternatives of the portal page; the layout profile is identical
First of all, the reader may submit his profile to the server. In that case,
the layout server decides using thresholds (e.g., about changes in the window
dimension) if the changes in the current profile require a re-generation of layout.
Furthermore, the reader may request a layout alternative. Such a request affects
that the server evokes the DC again with the current layout, which in turn starts
searching for a new layout for the last requested document (see Fig. 1). The CA
allows the reader also to edit parameters included in his profile, which have
been prepared by the author of the applied style sheet for customization. This
way, the reader may modify style aspects (e.g., colors and font sizes) and/or the
content selected for layout, just depending on which modifications the author
has granted.
References
1. L. Ardissono, L. Console, and I. Torre: On the application of personalization
techniques to news servers on the WWW. In Proc. AI*IA 99, 2000.
2. P. Baumgartner and A. Blohm: Automated deduction techniques for the management
of personalized documents. In Proc. of MKM 2001.
3. A. Borning, R. Lin, and K. Marriott: Constraint-Based Document Layout for the
Web. In Multimedia Systems 8.3 (2000), 177–189.
4. S. Feiner: A Grid-based Approach to Automating Display Layout. In Proc. of the
Graphics Interface 1988, 192–197, Morgan Kaufmann, Los Altos, CA.
5. T. Kamba, H. Sakagami, and Y. Koseki: ANATOGONOMY: a personalized newspa-
per on the World Wide Web. In International J. of HC Studies, 46, 1997, 789–803.
6. A. Kröner: Adaptive Layout of Dynamic Web Pages. In DISKI - Dissertationen zur
künstlichen Intelligenz 248, infix (2001), ISBN 3-89838-248-6
7. L. Tsaban: Intelligent Multimedia Authoring Tools For Electronic Publishing. Master
thesis, Bar-Ilan University, Ramat-Gan, 2001.
8. L. Weitzman and K. Wittenburg: Automatic Generation of Multimedia Documents
Using Relational Grammars. In Proc. of the ACM Multimedia 1994, 443–451, CA.
9. W.H. Graf: Constraint-Based Graphical Layout of Multimodal Presentations. In
M.T. Maybury and W. Wahlster, Readings in Intelligent User Interfaces, 1997.
A Hybrid Recommender System Combining
Collaborative Filtering with Neural Network
Meehee Lee1 , Pyungseok Choi2 , and Yongtae Woo2

1
Changshin College, Masan Kyungnam 630-764, Republic of Korea,
[email protected]
2
{choips,ytwoo}@sarim.changwon.ac.kr
Abstract. We propose a new recommender system which combines col-

laborative filtering(CF) with Self-Organizing Map(SOM) neural network.
First, all users are segmented by demographic characteristics and users
in each segment are clustered according to the preference of items using
the SOM neural network. To recommend items to a user, CF algorithm
is then applied on the cluster where the user belongs. As a result of ex-
perimentation for well-known movies, we show that the proposed system
satisfies the predictability of CF algorithm in GroupLens. Also, our sys-
tem improves the scalability and the performance of the traditional CF
technique.
1 Introduction
A recommender system in E-commerce means information filtering process to
supply personalized information by predicting user’s preferences to specific items.
The CF technique is widely used to suggest new items or to predict the utility
of a certain item for a particular user based on the user’s previous likings and
the opinions of other like minded users [1]. The CF technique, however, suffers
from some limitations such as early rater problem, sparsity, and gray sheep [2].
For the most part these limitations are related to the scalability and efficiency
of the k-NN approach. Essentially, k-NN requires that the neighborhood for-
mation phase be performed as an online process, and for very large data sets
this may lead to unacceptable latency for providing recommendations [3]. In
neighborhood-based CF systems, the time complexity can be very time consum-
ing with millions of users and items in the database and potentially make the
whole process unsuitable for realtime recommendation generation [1].
Recently, to improve the scalability of the CF algorithms and the quality
of the recommendations for the consumers, a hybrid model that combines CF
with other techniques has been studied. PTango has been combined content-
based prediction and CF for an online newspaper by basing a prediction on a
weighted average of the content-based prediction and the collaborative predic-
tion [4]. Good et. al have shown that CF can be used to combine personal IF
agents with the community opinion of the users to obtain better recommenda-
tions than those of agents or the users themselves [5]. O’Connor has proposed

532 M. Lee, P. Choi, and Y. Woo
a model combining clustering algorithm with CF to separate the set of items

based on users’ rating data to predict independently within each partition [6].
In this paper, we propose a hybrid recommender system, which combines CF
with the SOM neural network to improve the scalability and performance of the
traditional CF technique. We, first analyze user’s demographic characteristics to
segment users, and then cluster them according to preference of items using the
SOM neural network. In order to recommend items for a user, we determine a
cluster in advance where the features are similar to the user. A CF algorithm
is then applied to the users who belong to the cluster in order to predict user’s
preference for items.
2 A Collaborative Filtering Combining with SOM
2.1 Motivation
The preference of contents such as movies or music shows differences among users
depending on their taste. Also, the preference may be distinguished based on
the demographic characteristics such as sex, age, and academic career. However,
the CF technique utilizes the user’s preference for items, it is quite difficult to
consider the demographic characteristics. Moreover, as the tastes of other similar
users are considered in CF, it is not easy to reflect the features of the item itself.
The CF algorithm requires a large amount of computation that grows with
both the number of users and items because the pearson correlation coefficients
are calculated for the whole users. Therefore, a more sophisticated technique is
required to improve these problems.
2.2 The Proposed Hybrid Model
User Segmentation and Clustering. Demographic characteristics can be

used to identify the types of users that like a certain item. User segmentation
is a classification process in accordance with their demographic characteristics
for all users. Clustering is a learning process which groups users in each segment
based on the preference of items using the SOM neural network.
The SOM is an unsupervised learning model adequate to clustering, and
allows easier computations to the multidimensional data. Also, the SOM is little
affected from the size of data [7].
Applying Collaborative Filtering upon Cluster. The users in a cluster
have common demographic characteristics and similar preference for items. In
order to predict the preference for a user, CF algorithm is applied on the cluster
to which the user belongs. Therefore, our approach can reduce the computations
on the cluster which had been evidently decreased the size of the data. Figure 1
shows the conceptual diagram of the proposed hybrid recommender system.
Combining Collaborative Filtering with Neural Network 533
G jGmG
ijG
|G G
wG
UaG
yG
jG
kG pGwG ……UUG G … G ……UUG G … G
jG mG …… UUG G …G
|Gz G pGwG zGvGtG G

zGXG G
zGYG G
zGG G
Fig. 1. The Proposed Hybrid Model
3 Experimental Analysis
To verify the efficiency of the proposed model, we have tested for 54 well-known
movies and 174 users. We have gathered 5,331 rating values for the preference
of movie from 1 through 5 by the users explicitly. First, all users are segmented
by sex or age. Then, the users in each segment are clustered according to the
preference of genres using the SOM. To simplify the learning process, we have
used input patterns for the SOM as the preference values of genre, and generated
two clusters to each segment. The preference values of genre for a user are
derived from the average rating values of movies. The genres of movies had
been categorized into action, drama, comedy, fiction and thriller. Thus, each
cluster contains a set of users whose preferred genres are similar in the segment
determined by sex or age. Although the test set of movies were selected in a
restricted area, the preference of the movies are clearly distinguished by sex or
age.
To recommend movies to a user, the CF algorithm is then applied on the
cluster where the user belongs. As a result of experimentation, we show that the
proposed system satisfies the predictability of the CF algorithm in GroupLens.
MAE (Mean Absolute Error) value was compared to the proposed method and a
traditional CF algorithm. Figure 2 shows that our approach is a little better than
that of the traditional CF algorithm. Also, we could ensure that the computation
time of the proposed method was enhanced than that of the traditional CF
algorithm in GroupLens system.
4 Conclusion
In this paper, we have proposed a new hybrid recommender system that combines
the SOM and the collaborative filtering technique. This model has composed of
neighborhood generation and prediction generation. In neighborhood generation,
534 M. Lee, P. Choi, and Y. Woo
ʹΠΝΝΒΓΠΣΒΥΚΧΖ͑ΗΚΝΥΖΣΚΟΘ ΡΣΠΡΠΤΖΕ͑ΞΖΥΙΠΕ ʹΠΝΝΒΓΠΣΒΥΚΧΖ͑ΗΚΝΥΖΣΚΟΘ ΡΣΠΡΠΤΖΕ͑ΞΖΥΙΠΕ

;ͲͶ ;ͲͶ
ͪ͟͡͡ ͪ͟͡͡
ͩͦ͟͡ ͩͦ͟͡
ͩ͟͡͡ ͩ͟͡͡
ͨͦ͟͡ ͨͦ͟͡
ͨ͟͡͡ ͨ͟͡͡
ͧͦ͟͡ ͧͦ͟͡
ͧ͟͡͡ ͧ͟͡͡
Δ͢ Δͣ Δͤ Δͥ Δͦ Δͧ Δͨ Δͩ
ΔΝΦΤΥΖΣΤ ΔΝΦΤΥΖΣΤ
(a) CF vs Proposed Method(by sex) (b) CF vs Proposed Method(by age)
Fig. 2. The results of experiment
all users were segmented by their demographic characteristics such as sex or age.
Then, clusters were generated using the SOM which learns the preference of items
in each segment. In prediction generation, the CF algorithm was applied to a
certain cluster including the user, and the results were recommended to the user.
The system has tested on the movies data set to verify the efficiency of
the proposed technique. Experimental results show that the proposed system
yields a little better predictability than the traditional CF-based system. Also,
our system can reduces a large amount of the computation time to calculate
correlation coefficients. In conclusion, we can expect that our system improves
the scalability and the performance of the traditional CF technique.
References
1. Sarwar, B., Karypis, G., Konstan, J., Riedl, J.: Item-based Collaborative Filtering
Recommendation Algorithms. Proceedings of the 10th International World Wide
Web Conference(WWW10), Hong Kong, May (2000)
2. Sarwar, B., Karypis, G., Konstan, J., Riedl, J.: Analysis of recommendation algo-
rithms for e-commerce. Proceedings of ACM Conference on E-Commerce (2000)
158–167
3. Mobasher, B., Dai, H., Luo, T., Nakagawa, M.: Improving the Effectiveness of
Collaborative Filtering on Anonymous Web Usage Data. Proceedings of the IJ-
CAI2001 Workshop on Intelligent Techniques for Web Personalization(ITWP01),
August (2001)
4. Claypool, M., Gokhale, A., Miranda, T., Murnikov, P., Netes, D., Sartin, M.: Com-
bining Content-Based and Collaborative Filters in an Online Newspaper. ACM
SIGIR Workshop on Recommender Systems, Berkeley, CA. (1999)
5. Good, N., Schafer, B., Konstan, J., Borchers, A., Sarwar, B., Herloker, J., Riedl,
J.: Combining Collaborative Filtering With Personal Agents for Better Recom-
mendations. Proceedings of the AAAI conference. (1999) 439–446
6. O’Connor, M., Herlocker, J.: Clustering items for Collaborative Filtering. ACM
SIGIR Workshop on Recommender Systems, Berkeley, CA. (1999)
7. Kohonen, T.: The Self-Organizing Map. Proceedings of the IEEE, (1990) 1564-1479
An Adaptive Cooperative
Web Authoring Environment
Ana Marı́a Martı́nez-Enrı́quez1 , Dominique Decouchant2 ,

Alberto L. Morán2,4 , and Jesus Favela3
1
Depto de Ingenierı́a Eléctrica, CINVESTAV-IPN, D.F., México
[email protected]
2
Laboratoire “Logiciels, Systèmes, Réseaux”, Grenoble, France
3
Ciencias de la Computación, CICESE, Ensenada, B.C., México
4
Facultad de Ciencias UABC, Ensenada, B.C., México
[email protected]
Abstract. Using AllianceWeb, authors distributed around the world

can cooperate producing large documents in a consistent and concerted
way. In this paper, we highlight the main aspects of the group aware-
ness function that allows each author to diffuse his contribution to other
co-authors, and to control the way by which other contributions are inte-
grated into his environment. In order to support this function, essential
to every groupware application, we have designed a self-adaptive cooper-
ative interaction environment, parametrized by user preferences. Thus,
the characteristics of an adaptive group awareness agent are defined.
Keywords: Cooperative and distributed authoring, World Wide Web,

group awareness agent, GAA, adaptive cooperation interface.
1 Introduction
World Wide Web technology enables users distributed over the Internet to ac-
cess a growing amount of information. However, this information is stored and
managed on Web servers, and users only remain consumers of information. In
this way of operation, Web servers only reply to requests of client applications.
In spite of the existence of several interactive browser/editor tools, the in-
formation located on WWW servers is usually updated manually by the site
manager: documents are prepared in a private editing environment and then
pushed into the public space on a server. At the same time, many current trends
in Web publishing research focus on providing elaborate functions to access doc-
uments and publishing them on a remote site.
2 Cooperative Editing
The cooperative production of documents is the result of the coordination be-
tween different authors to carry on the activities needed for the joint production

536 A.M. Martı́nez-Enrı́quez et al.
of documents. Then, co-authors will use cooperative editing tools to enhance

their effectiveness, to enforce consistency, to control the validity of contribu-
tions, and finally to ensure the quality of the commonly produced work.
The PIÑAS (Platform for Interaction, Naming And Storage) middleware
[2], used to develop the AllianceWeb cooperative authoring system [1], aims to
determine the specific requirements and the dedicated techniques to support
Web cooperative authoring environments.
It supports cooperative authoring based on document fragmentation, and so
introduces several interesting distributed features: author identification, docu-
ment and resource naming, document replication, consistency of the concurrent
(distributed) production, and an ad-hoc distributed document storage system.
3 Adaptive and Deductive Group Awareness Function
Group awareness is a key feature of a cooperative application. Cooperative au-

thoring applications developed on top of PIÑAS allows each user 1) to diffuse
his contribution to other users, and 2) to control the integration of other contri-
butions into his working environment.
AllianceWeb uses a PIÑAS asynchronous notification mechanism that allows
the user to be notified of other coauthor editing actions. Currently, the group
awareness function allows each user to configure the notification level (automatic
updating or attentive/unattended reading) for each fragment he is consulting.
Considering the object space manipulated by the co-authors, and the set of
actions applied on the object space, the goal is to define a flexible, deductive and
adaptive cooperative group awareness environment. By these terms we mean:
Deductive - The system is able to analyze the cooperative working actions
of each user and to determine user objectives.
Adaptive - Taking into account each user’s goals, and with respect to their
individual preferences, it guides and dynamically adapts user environments.
In order to avoid the need for users to manage their cooperative environments
and to allow them to cooperatively work more efficiently, our system is defined
as an autonomous agent (see section 3.1) that manages the group awareness
environment (e.g. What are the suited tools to highlight in the user environment
to make him more efficient in the cooperation process? On what user activities
is it judicious to concentrate interaction or group awareness tools? etc).
3.1 A Group Awareness Agent
The goal of the Group Awareness Agent (GAA) is to determine the manipulated
entities and the executed cooperative actions, to modify the presentation of this
information on the user’s screen. It is important to notice that this presentation
is constrained by the user preferences applied using presentation filters.
Periodically the agent re-evaluates and updates the set of manipulated ob-
jects. This periodic re-evaluation is performed in parallel with the actions applied
by the user. This re-evaluation is completed both by statistic measures of the
An Adaptive Cooperative Web Authoring Environment 537
manipulated object space and by the analysis of the concurrent interactions

(concerned users, manipulated objects and action frequency).
The GAA is guided by user preferences (figure organization, formats, pre-
sentation, tool interfaces, etc) and user roles on the shared fragments.
For instance, let us consider the cooperative environment of user Jean, who
mainly acts as a reviewer on most of the shared document.
If author (document) = x
If role (x) = ‘‘reader"
If frequency ("reader") > 0.5
If annotationTool (x) = ‘‘installed"
Then
InstallInEnvironment (x) ← annotationTool
EndIf
Thus, the action components of the agent trigger specific cooperation actions
(annotation tool) in the environment of user Jean (”reader”).
The system announces to all users the degree of availability of everyone in
the global cooperation process. Then, each user may decide whether or not to
change his interest or his restrictions to avoid conflicts.
3.2 The Object and Action Spaces

In order to define a deductive and adaptive agent to support the group awareness
function, it is required to identify the manipulated objects, and the actions
performed on these objects.
The documents - Single entities as well as object compositions to support
information partitioning and sharing.
Some Special Objects - Cooperative tools may be present in the user
environment: an electronic white board, a multimedia teleconferencing tool, etc.
Objects for Group Awareness Support - These tools allow them to con-
sult the state variables (current work, availability, interests, available interaction
tools, etc) of other user environments and build or update his own interface.
Actions Applied on Documents - Among the observable actions, we can
identify: document creation, renaming or removing, addition, modification or
deletion of a document resource, document fragmentation, among others.
Document Editing Actions - The observable action list (GAA input in-
formation) is not restricted, and includes events explicitly or implicitly generated
by editing, resource managing, and cooperative activities management.
Each running instance of the AllianceWeb application is composed of two
separated active parts: the browser/editor and the assistant (see Fig. 1).
The browser/editor goal is to manage user interactions: document dis-
playing interpretation of user editing commands. All editing commands gener-
ate modification actions only for the local document copies. Eventually, they
generate some asynchronous remote requests that are delegated to the assistant.
The AllianceWeb assistant constitutes the interface with the servers. It
takes in charge all communication failures, and thus it periodically tries to re-
contact all temporarily unreachable servers. It works in autonomous mode.
538 A.M. Martı́nez-Enrı́quez et al.
user interface remote HTTP
modifications request server
hints for editor

GAA
AllianceWeb AllianceWeb
Actions Editing Module Assistant update local HTTP
Reasoning: server
Knowledge request
API ECF
Base
CGI script CGI script
Thot Editing
Perception
Kernel
other editing
events Document
events
Management
Draw Editor Local Copies of Documents
Annotation Tool AllianceWeb Application
Fig. 1. AllianceWeb and the Group Awareness Agent (GAA)
4 Conclusion
AllianceWeb allows several authors distributed over the Internet to produce
shared Web documents in a concerted and consistent way; to achieve this we
have created a seamless production/consultation space where co-authors coop-
eratively produce shared documentation. At the same time, all users doing Web
navigation can consult the documents whose versions are automatically updated.
The group awareness function allows each user to be aware of the production
of others and allows them not only to adjust their contribution, but also to
influence or to guide other user contributions.
We presented the first principles of a group awareness agent (GAA) that
allows the automatic adaptation of each user’s interaction environment, taking
into account their preferences and their actions in the shared object space.
Acknowledgements. This work is supported by ECOS and ANUIES orga-
nizations (project M98M01), by CONACyT projects (29729-A and 33067-A),
by CNRS-CONACyT cooperation projects (9018 / E130-505 and 10395 / E130-
706), and by SEP-SESIC and UABC (grant P/PROMEP: UABC-2000-08), with
the scholarship UABC-125 provided to the third author.
References
1. D. Decouchant, A. M. Martı́nez et E. Martı́nez, “Documents for Web Cooperative
Authoring”, In Proc. CRIWG’99, 5th CYTED-RITOS International Workshop on
Groupware, IEEE Computer Society, Cancun, Mexico, 15-18 September 1999.
2. D. Decouchant, J. Favela and A. M. Martı́nez-Enrı́quez, “PIÑAS: A Middleware
for Web Distributed Cooperative Authoring”, In Proc. of the 2001 Symposium on
Applications and the Internet (SAINT’2001), IEEE Computer Society and IPJ
Information Processing Society of Japan, San Diego, California (USA), pp. 187-
194, 8-12 January 2001.
Adapting Learner Evaluation by Plans
Alke Martens and Adelinde M. Uhrmacher
University of Rostock, Department of Computer Science,

18051 Rostock, Germany
{martens, lin}@informatik.uni-rostock.de
Abstract. State of the art intelligent tutoring systems (ITS) provide

ample opportunity for adaptability. Typically, they adapt single pages
and contents to the learner’s knowledge level and evaluate the learner
based on his past and current behavior. However, to decide whether
a learner behaves in a coherent manner, the learning system requires
some means for anticipating the learner’s goals and intentions. Thus,
not only the structure of the tutoring process, but also the evaluation
strategy should be adaptive. The ITS PLAIT reasons about and evalu-
ates the learner’s behavior according to a plan adapted for that purpose.
To facilitate the adaptation, the plan is structured into a three-layered
hierarchy. The topmost layer reflects the teaching objectives suggested
by pedagogy and the psychology of learning; the middle one embeds
domain-dependent refinement. The third layer represents the steps the
learner has taken and should optimally take.
1 Introduction
In a case-based intelligent tutoring system (ITS), the learner should not only
acquire knowledge about facts and how to apply them, but also knowledge about
learning behavior, and how to proceed in the context of a given training case. The
process of acting and interacting becomes part of the learning. To decide whether
a learner behaves in a coherent manner, the learning system has to anticipate the
learner’s goals and intentions. The tutoring system PLAIT (Planning Agents in
Intelligent Tutoring) adopts an approach to that effect in a case- and web-based
context. Based on techniques of hierarchical planning, plan recognition, and
agents, it tries to draw conclusions about the learner’s intentions. A similar has
been successfully employed in human-computer interaction, e.g. in the system
Collagen [1]. In PLAIT, the learner’s choice of sequence of steps is monitored
and continuously evaluated. Reflected in a user model, this information is used
to adapt a flexible tutoring process during runtime and to provide context-
sensitive help and explanation. The final evaluation offers hints and advice to the
learner, according to his progress in the training domain, and indicates possible
fundamental learning gaps.

540 A. Martens and A.M. Uhrmacher
2 PLAIT - Reasoning and Intelligent Tutoring
PLAIT [6] is an extension of the case- and web-based ITS “Docs ’N Drugs - The
Virtual Policlinic” [2], which is used at the University of Ulm as a constitutive
part of the medical curriculum. The system is implemented as a classical ITS,
described in, e.g., [3]. The extension PLAIT proposes to enrich the system with
planning facilities, a plan library, and mechanisms for plan recognition. As de-
scribed in, e.g. [1], we assume that learner and intelligent tutor collaborate in
the learning environment. Thus, we have modeled them as two separated agents.
This allows mimicking the concurrent activities and the role of learner and tutor
within a training scenario. The idea to embed planning agents in the system
is based on the approach to perceive “plans as complex mental attitudes”, as
described in [5]. PLAIT works with an adaptive flexible tutoring process. Series
of interrelated steps lead from the initial page to the last page. Internally, each
page consists of a set of page elements and a set of possible next steps that
the learner could pursue. Page elements and steps are equipped with pre- and
post-conditions and a required experience level of the learner. The subset of page
elements and next steps, which is actually displayed is adapted to the learner’s
profile and history in current and past training cases. The interrelated steps and
pages form paths in the training cases; they are called the tutoring process.
Reasoning about the learner’s progress in the training case can happen in
different ways. Whereas reasoning about interaction elements and checkpoints
seems straightforward, reasoning about the learning process is more difficult. At
every state in the tutoring process, the learner’s actual choice of next step can
be evaluated differently. Either the chosen step could be the correct next step,
foreseen as optimal by the author. Or, the next step might be unnecessary, i.e.
not optimal but not totally wrong either, and would likely not immediately affect
the training case’s development. Or, the chosen step could be definitely wrong.
In that case, the immediate effects differ depending on the step’s importance and
content; only cumulative effects might prove critical. In addition, it has to be
taken into account that a learner might choose a series of steps that are wrong
with respect to the training case’s goals, but that are intrinsically coherent with
respect to his current mental strategy and knowledge. Thus, a chosen step can
be correct, given the facts the learner has acquired so far. Even if it is wrong in
the case’s overall context, it would be logically coherent with the former steps
the learner chose. Moreover, it could open an avenue to a series of interrelated
steps, all of which are wrong with respect to the original purpose of the case.
These steps have to be evaluated and supported differently compared to “simply”
wrong or unnecessary steps. Reasoning about the learner’s understanding of the
required behavior in the training case is achieved by comparing a pre-defined
optimal path in the training case to the path the learner actually has chosen.
The assumption is that the learner follows a certain strategy, which the tutoring
system perceives as the learner’s mental plan. The learner agent observes the
learner’s activities and reports them as a sequence of steps to the tutor agent.
The evaluation of the learner’s choices takes place in consideration of his former
Adapting Learner Evaluation by Plans 541
Abstract Tutor
Task 1 Task 2
Plan A Agent
Plan B Sub- Sub-
(Task Reduction Task 1 Task 2
Schema of Task 2)
Operational
Plan C Subsub Subsub
(Task Reduction Task 1 Task 2
Schema of Sub-Task 2)
Learner’s
Mental Plan Step 1 Step 4
Fig. 1. The hierarchy of plans in PLAIT [6]
decisions, i.e. the learner’s background knowledge and the sequence of steps he
has chosen becomes part of the reasoning [5].
2.1 A Three Layered Planner and Adapting Plans
The Hierarchical Task Network (HTN) in PLAIT is realized as a three layered

plan (Fig. 1). A goal in the HTN planner is a high level operator that shall be
achieved. Associated with every high level operator is a set of skeleton plans,
called operator reduction schemata, each of which can alternatively be used to
achieve the high level goal.
The abstract plan A at the topmost layer contains high level tasks that reflect
skills the learner should acquire, independent of the training domain. A skill is
part of a pedagogical strategy which the author of the training case defines, for
example according to [4]. Thus, a skill might be “knowledge of specific facts”,
“analysis of relationships”, etc. [4]. The high level tasks will be reduced into
subtasks by predefined task reduction schemata. Sub-tasks at plan level B are
constructed by a set of context dependent learning sub-goals that are domain
dependent. For example in the medical domain, sub-tasks are “order the cor-
rect examination before conducting a surgery”, “choose the correct diagnosis”,
etc. Tasks in plan B denote macro steps within the training domain. A further
task reduction leads to the operational plan C. Each of the sub-tasks at plan
level B will be reduced to operational subsub-tasks, which finally correspond to
concrete actions the author expected the user to apply. Those actions might be
e.g. choose “anamnesis” or order “X-ray of left arm”. The tutor agent compares
the operational plan C with the sequence of steps reported by the learner agent.
For a more detailed description see [6]. If the learner starts a training case, the
learner agent generates a tutoring process, which is adapted to the learner’s pro-
file. The learner’s history monitors the steps chosen in the current training case.
It allows conclusions about the learner’s actual state of knowledge. The learner
starts the training case with the first page, and a set of next step choices, both
adapted to his abilities. The learner makes his first choice of next step. Each
of the steps he chooses is reported to the tutor, which compares the steps with
the suggested subsub-tasks in plan C. If the learner chooses a step that is not
part of the optimal plan C, the tutor uses plan recognition. It queries the plan
542 A. Martens and A.M. Uhrmacher
library whether it contains a new sub-plan at the operational level C that is a

task reduction schema for the task in plan B and is consistent with the steps the
learner took so far. If such a new task reduction schema exists, the tutor will
skip the actual plan and embed the new sub-plan as plan C. The new sub-plan
will then be used as new reference plan. If no such plan is available in the plan
library, re-planning can be initiated. In some situations, the tutor agent will offer
the learner help, i.e. a plan repair strategy, to lead him back to the proper path.
The learner has the option to ignore this offer and continue following his own
ideas. The tutor agent will mark the subsub-task at layer C as not fulfilled. The
higher layers of the current plan indicate whether the learner lacks certain skills,
e.g. in applying knowledge in a certain situation. This information will be part of
the final evaluation. At the end of the training case, in addition to the objective
which the learner has finally achieved, the effort required to adapt the plan, and
the number and type of mistakes, are used as indicator for the learner’s success.
From the bottom to the top of the hierarchy, plans are used for evaluation.
3 Conclusion
The system PLAIT extends the functionality of the web- and case-based ITS
“Docs ’N Drugs”. It is aimed at supporting a “more intelligent” evaluation
of the learner’s progress. Moreover, the mechanisms contained in PLAIT will
help the training cases’ author to develop training cases in a structured man-
ner based on pedagogical theories. PLAIT employs two agents. The tutor agent
uses a three-layer HTN plan, techniques of plan recognition, and plan repair.
The three layers of the hierarchical plan reflect learning skills (based on learning
theory), learning sub-goals, and concrete operational sub tasks. The tutor agent
compares the learner’s actions as reported by the learner agent to the lowest
plan layer. It adapts the plans of the plan hierarchy according to the learner’s
activity. Thereby, it tries to trace, interpret, and anticipate the mental plan of
the learner. Based on the goals and the pedagogical strategies at the topmost
layer, conclusions about the learner’s skills are derived.
References
1. Lesh, N., Rich, C., Sidner, C.L.: Using Plan Recognition in Human-Computer-
Collaboration. Technical Report No. 98-14, IBM Watson Research Center (1998)
2. Docs ‘N Drugs - Die Virtuelle Poliklinik http://www.docs-n-drugs.de (2000)
3. Martens, A., Bernauer, J., Illmann, T., Seitz, A.: Docs ’n Drugs - The Virtual
Polyclinic. In:American Medical Informatics Conference, Washington USA (2001)
4. Bloom, B.S.: Taxonomy of Educational Objectives, Handbook I: Cognitive Domain.
David McKay Company, Inc, New York (1956), Reprint 1971
5. Pollack, M.E.: Plans as Complex Mental Attitudes. In: Cohen, P.R., Morgan, J.,
Pollack, M.E. (eds): Intentions in Communication, MIT Press (1990)
6. Martens, A., Uhrmacher, A.M.: Adaptive Tutor Processes and Mental Plans. Ac-
cepted for: Intelligent Tutoring Systems ITS, Biaritz, France (2002)
WETAS: A Web-Based Authoring System
for Constraint-Based ITS
Brent Martin and Antonija Mitrovic
ICTG, Department of Computer Science, University of Canterbury

Private Bag 4800, Christchurch, New Zealand
Abstract. Constraint-Based Modelling (CBM) is a student modelling technique

for Intelligent Tutoring Systems (ITS) that is rapidly maturing. We have
implemented several tutors using CBM, and demonstrated their suitability to, in
particular, open-ended domains. It is easier to build tutors in some domains
(e.g. open-ended) using CBM than other common approaches. We present
WETAS (Web-Enabled Tutor Authoring System), a tutoring engine that
facilitates the rapid implementation of ITS in new domains. We describe the
architecture of WETAS, and give examples of two domains we have
implemented.
1 Introduction
Constraint-Based Modelling (CBM) [5] is an effective approach that simplifies the
building of domain and student models. We have used CBM to develop SQL-Tutor
[3], an ITS for teaching the SQL database language. SQL-Tutor tailors instructional
sessions in three ways: by presenting feedback when students submit their answers,
by controlling problem difficulty, and by providing scaffolding information. Students
have shown significant gains in learning after as little as two hours of exposure to this
system [4].
While CBM reduces the effort of building domain models for ITS, the task of
building an ITS is nevertheless still large. To reduce the authoring effort, we have
developed WETAS, a web-based tutoring engine that performs all of the common
functions of text-based tutors. To demonstrate the flexibility of WETAS, we have re-
implemented SQL-Tutor, and developed a new Language Builder ITS (LBITS).
Although these domains share the property of being text-based, they have very
different problem/solution structures.
2 Constraint-Based Modelling
CBM [5] is a relatively new approach to domain and student modelling, based on the
theory of learning from performance errors [6]. It models the domain as a set of state
constraints, where each constraint represents a declarative concept that must be
learned and internalised before the student can achieve mastery of the domain.
Constraints represent restrictions on solution states, and take the form:
If <relevance condition> is true for the student’s
solution,
THEN <satisfaction condition> must also be true

544 B. Martin and A. Mitrovic
The relevance condition of each constraint is used to test whether the student’s
solution is in a pedagogically significant state. If so, the satisfaction condition is
checked. If it succeeds, no action is taken; if it fails, the student has made a mistake,
and appropriate feedback is given. Ohlsson does not impose any restrictions upon
how constraints are encoded and/or implemented. We have used a pattern-matching
representation designed for this purpose [2]. For example, the following constraint
from the SQL domain checks that names used in the ‘WHERE’ clause are valid
names from the database:
(147
‘You have used some names in the WHERE clause that are not
from this database.’
(match SS WHERE (?* (^name ?n) ?*)) ; relevance cond
(or (test SS (^valid-table (?n ?t)) ;

satisfaction
(test SS (âttribute-p (?n ?a ?t)))) ; condition
‘WHERE’)
3 Architecture
WETAS is a web-based tutoring engine that provides all of the domain-independent

functions for text-based ITS. It is implemented as a web server, written in Allegro
Common Lisp, and using the Allegroserve Web server (see www.franz.com).
WETAS supports students learning multiple domains at the same time; there is no
limit to the number of domains it may accommodate. Students interact through a
standard web browser such as Netscape or Internet Explorer. Figure 1 shows a screen
from SQL-Tutor implemented in WETAS. The interface has four main components:
the problem selection window (top), which presents the text of the problem, the
solution window (middle), which accepts the students input, the scaffolding window
(bottom), which provides general help about the domain, and the feedback window
(right), which presents system feedback in response to the student’s input.
WETAS performs as much of the implementation as possible, in a generic fashion.
In particular, it provides the following functions: problem selection, answer
evaluation, student modelling, feedback, and the user interface. The author need only
provide the domain-dependent components, namely the structure of the domain (e.g.
any curriculum subsets), the domain model (in the form of constraints), the
problem/solution set, the scaffolding information (if any), and, possibly, an input
parser, if any specific pre-processing of the input is necessary.
WETAS provides both the infrastructure (e.g. student interface) and the ‘intelligent’
parts of the ITS, namely the pedagogical module and the domain model interpreter.
The former makes decisions, based on the student model, on what problem to present
to the student next, and what feedback they should be given. The latter evaluates the
student’s answers, by comparing them to the domain model, and uses this information
to update the student model.
Problems are selected by determining the difficulty of each candidate exercise with
respect to the student. When the problems are authored, the system compares them to
the domain model to determine their structural complexity, which is based on the
WETAS: A Web-Based Authoring System for Constraint-Based ITS 545
Fig. 1. WETAS Interface
number and complexity of the constraints that are relevant to them. When a new
problem is to be selected, each candidate is compared to the student model, and the
difficulty is increased by a constant amount for each constraint the student currently
doesn’t know, and by a different constant for each constraint that has never been
relevant. Thus, the difficulty of each problem differs according to each student’s
strengths and weaknesses. The system then chooses the one with a difficulty that is
closest to the student’s current ability rating.
4 Example Domains
WETAS has been implemented in prototype form, and used to build two tutors, to
explore its capabilities and evaluate its effectiveness in reducing the ITS building
effort. These two implementations are now described.
SQL-Tutor. SQL-Tutor [3] teaches the SQL database query language to second and
third year students at the University of Canterbury, using Constraint-Based
Modelling. Students are given a textual representation of a database query that they
must perform, and a set of input fields (one per SQL clause) where they must write an
appropriate query. This system was implemented in 1998 as a standalone tutor, in
1999 as a Web-enabled tutor, and has been re-implemented in WETAS. The general
design and interface of WETAS borrows heavily from the original SQL-Tutor. We
had no problems implementing SQL-Tutor in WETAS.
Language Builder ITS (LBITS). Language Builder is an existing paper-based
teaching aid that is currently being converted to a computer system. It teaches basic
English language skills to elementary and secondary school students, by presenting
546 B. Martin and A. Mitrovic
them with a series of ‘puzzles’, such as crosswords, synonyms, rhyming words, and
plurals. We created an ITS from Language Builder (LBITS) by adding a domain
model so that feedback could be expanded from a simple right/wrong answer to more
detailed information about what is wrong, such as that the meaning of their answer
didn’t match the meaning of the clue, or they have got the letters ‘i’ and ‘e’ reversed.
These rules were inferred from a school spelling resource book [1]. We generated the
problem set automatically from a vocabulary list in [1]. The constraints were authored
in just five hours, and the problem set was created in around two days. LBITS
contains 315 constraints and 222 puzzles.
We have so far implemented a subset of the Language Builder puzzles, all of which
share the common format of a set of clues, each requiring one or more single-word
answers. The system has been evaluated at a New Zealand elementary school, where
the children found the system easy to use, and felt they had learned a lot. An analysis
of constraint failures supported this latter statement: a plot of the average failure rate
for each constraint versus the number of times it was relevant gave a ‘power curve’,
indicating the constraints were being learned.
5 Conclusions
Constraint-Based Modelling (CBM) is an effective approach that simplifies the
building of domain and student models for some domains, particularly open-ended
ones. We have developed a prototype authoring system called WETAS for CBM
tutors, which we intend to use to develop additional tutors for further research into
CBM, and for release into classrooms. We have reimplemented SQL-Tutor using
WETAS, and built a new tutor, LBITS. WETAS draws upon the strengths of CBM,
plus research carried out to date in practical implementations of CBM. It appears to be
a promising tool for the rapid development of new tutors, and a useful step towards
the large-scale deployment of Intelligent Tutoring Systems.
Acknowledgement. This research is supported by the University of Canterbury, grant

U6430.
References
1. Clutterbuck, P. M. (1990). The art of teaching spelling: a ready reference and classroom
active resource for Australian primary schools. Melbourne, Longman Australia Pty Ltd.
2. Martin, B. and Mitrovic, A. (2000). Tailoring Feedback by Correcting Student Answers. In
G. Gauthier, C. F., K VanLehn (Ed.), Proceedings of the Fifth International Conference on
Intelligent Tutoring Systems, Montreal, Springer, pp. 383-392.
3. Mitrovic, A. (1998). Experiences in Implementing Constraint-Based Modeling in SQL-
Tutor. In, Proceedings of the ITS'98, pp. 414-423.
4. Mitrovic, A. and Ohlsson, S. (1999). Evaluation of a Constraint-Based Tutor for a Database
Language. International Journal of Artificial Intelligence in Education 10, pp. 238-256.
5. Ohlsson, S. (1994). Constraint-Based Student Modeling. In Student Modeling: The Key to
Individualized Knowledge-Based Instruction, Springer-Verlag, pp. 167-189.
6. Ohlsson, S. (1996). Learning from Performance Errors. Psychological Veview 3(2), pp.
241-262.
Open-Ended Adaptive System
for Facilitating Knowledge Construction
in Web-Based Exploratory Learning
Hiroyuki Mitsuhara, Youji Ochi, and Yoneo Yano
Dept. of Information Science and Intelligent Systems, Faculty of Engineering, Tokushima

University, 2-1 Minamijosanjima, Tokushima, Japan
{mituhara, ochi, yano}@is.tokushima-u.ac.jp
Abstract. A typical learning method using the Internet is exploratory learning,

where learners construct knowledge through exploring the web autonomously.
A problem with web-based exploratory learning is insufficient hyperlinks. An
open-ended adaptive system has been developed that facilitates knowledge
construction in web-based exploratory learning by avoiding the impasse.
1 Introduction
Rapid spread of the Internet enables us to learn with informative web pages that exist
all over the world. A typical learning method using the Internet is exploratory
learning, where learners construct knowledge through exploring the web on the basis
of their own needs, goals, or interests. In web-based exploratory learning, the learners
frequently reach a learning impasse caused by insufficient hyperlinks. If a page is not
complete with sufficient hyperlinks to make up for learners' lack of knowledge, the
learners will not be able to understand the page. Thus insufficient hyperlinks lead the
learners to the impasse.
Search engines, which can supply suitable hyperlinks (pages), help escape from the
impasse but occasionally return meaningless hyperlinks. Adaptive web-based learning
systems, which alter the hypertext contents/structure to learners' characteristics, can
avoid the impasse. The existent adaptive web-based learning systems do not fulfill
learners' expectations that the open web (i.e. arbitrary web pages) will be adapted.
Almost all such systems limit the adaptation to the closed web (i.e. web pages inside
one server). Recently, this limitation has become a major focus in the domain of
adaptive hypermedia [1].
The goal of this study is to develop an open-ended adaptive system for facilitating
knowledge construction in web-based exploratory learning. Key factors for such a
system are the open web, adaptive hyperlink addition, and zero-input learner
modeling. The open web containing numerous pages is indispensable in fulfilling
learners' expectations. The adaptive hyperlink addition avoids the impasse by
enabling learners to refer to well-structured knowledge seamlessly from the open
web. A zero-input learner modeling is needed in order not to interrupt learners'
exploration activity. ITMS (Individualized Teaching Material System), a complete
system with all the key factors, has been developed.

548 H. Mitsuhara, Y. Ochi, and Y. Yano
Similar systems have appeared up to the present. AHA expands exploration scope
by adapting hyperlinks that connect with the open web [2]. KBS-Hypertext actualizes
extensible learning environment by integrating the open web with independent
courseware [3]. These systems can be classified as the system that utilizes the open
web from the closed web. However, ITMS is in the opposite classification.
2 ITMS
ITMS, which has a simple system composition (Fig. 1), can apply to diverse domains
of teaching. Basically, learners can use this system on a web browser without plug-in
software.
2.1 System Composition
The closed web: The closed web consists of knowledge components (KC for short)
created by teachers. KC is a fraction divided from educational web pages and can be
combined with KCs stored in distributed servers. Although heterogeneous KCs may
trigger off an incoherence that lowers learning effectiveness [5], these can conversely
enhance learners' ability to select correct knowledge and associate new knowledge
with already acquired knowledge.
Hyperlink addition engine: This engine first fetches a page requested by learners
from the open web. Secondly, it extracts the related knowledge from the page by
means of keyword matching. Thirdly, it combines corresponding KCs using
hierarchical stretch-text [4]. Additionally, this engine modifies original hyperlinks on
the page in order that the learners' requests are always sent to ITMS.
Learner model: ITMS builds a learner model that represents knowledge states,
‘acquired’, ‘unstable’, and ‘forgot’.
Adaptation engine: This engine adapts the combined KCs on the basis of knowledge
states. The adaptation techniques used are adaptive link annotation and adaptive link
sorting (from the taxonomy in [1]).
Open Web
Adaptation Learner
Adapted KCs Engine Model
Closed Web
Combined KCs
Web Page KC DB
Closed Web
(ID & Location)
Hyperlink
Addition Engine KCs
Closed Web KCs ITMS
Fig. 1. System Composition
Open-Ended Adaptive System for Knowledge Construction in Web-Based Learning 549
2.2 Learner Modeling
ITMS adopts zero-input learner modeling that estimates knowledge states from a
learner's actions to KC. It differs from existing modeling methods, which generally
consider that learners with understanding of the upper knowledge understand its lower
knowledge already. Our modeling method is based on the following hypotheses.
• The lower knowledge is difficult for learners to notice and easy to forget.
• They notice knowledge from the caption of the knowledge.
• When they notice forgotten knowledge, they refer to that knowledge and acquire
it.
• Even if they notice knowledge they have acquired, they do not refer to it.
• The state of knowledge that they did not notice deteriorates to ‘forgot’
In this paper, T represents the number of times that they do not notice knowledge
sequentially, Tf (Tf>=2 and natural numbers) represents the threshold for estimating a
knowledge state to be ‘forgot’. When they notice knowledge, T=0 is performed for
the knowledge. In the contrary case, T=T+1 is performed.
ITMS estimates the knowledge state by comparing T and Tf, and updates Tf by
observing whether a learner refers to the knowledge in the ‘forgot’ state. Referring to
such knowledge may indicate that the estimate was correct and the knowledge had
already been forgotten before he/she noticed it; therefore Tf=Tf -1 is performed.
Conversely, not referring to such knowledge may indicate that the estimate was
incorrect and the knowledge will keep acquired for a while; therefore Tf=Tf +1 is
performed. Low value Tf indicates that the knowledge is easy to forget. High value Tf
indicates that the knowledge is difficult to forget. Learners' continuous exploration
will bring Tf closer to the true tendency of forgetting.
2.3 Adaptation
The adaptation in ITMS facilitates equalizing knowledge construction by making

learners notice the knowledge in the ‘forgot’ state. This adaptation also facilitates our
zero-input learner modeling. Specific adaptation methods are to transfer such
knowledge from the initial layer to the top layer (i.e. adaptive link sorting) and to
highlight the caption of such knowledge with icons (i.e. adaptive link annotation).
Knowledge with low value Tf is frequently adapted. Conversely, knowledge with high
value Tf is infrequently adapted. Figure 2 shows an actual example of the adaptation.
3 Summary
Learners may increasingly be inclined to construct knowledge in web-based

exploratory learning. This paper has outlined an open-ended adaptive system, which
facilitates such knowledge construction by avoiding the learning impasse caused by
insufficient hyperlinks.
This research has many remaining discussing points. One future topic of interest
may be proving the reliability of zero-input leaner modeling through a large-scale
experiment.
550 H. Mitsuhara, Y. Ochi, and Y. Yano
Fig. 2. An actual example of the adaptation in ITMS: Here, a learner visits the page of ‘CAD
products’ (on the left). Words of ‘Bezier surface’ and ‘B-Spline surface’ appear on the page.
ITMS extracts the KCs concerning Bezier surface and B-Spline surface as related knowledge
and combines these KCs into a new page (on the top center). The KC of the ‘Inner Product’,
which is the lower knowledge of Bezier and B-Spline surface, is shown on another new page
(on the right). This occurs because the state of the KC is ‘forgot’. In this way, the adaptation
encourages the learner to notice the knowledge in the ‘forgot’ state and enables the learner to
directly refer to such knowledge. Each KC sketches its own linear knowledge structure, which
functions as a path indicator.
References
1. Brusilovsky, P.: Adaptive Hypermedia, User Models and User-Adapted Interaction 11,
pp87-110 (2001).
2. De Bra, P., and Calvi, L.: AHA! An open Adaptive Hypermedia Architecture, The New
Review of Hypermedia and Multimedia, vol. 4, pp. 115-139 (1998).
3. Henze, H. and Nejdl, W.: Extensible Adaptive Hypermedia Courseware: Integrating
Different Courses and Web Material, Proc. of AH2000, pp.109-120 (2000).
4. Hook, K., Karlgren, J., Waern, A., Dahlback, N., Jason, C.G., Karlgren, K., and Lemaire,
B.: A glass Box Approach to Adaptive Hypermedia, User Models and User-Adapted
Interaction 6(2-3), pp.157-184 (1996).
5. Seeberg, C., Steinacker, A., and Steinmetz, R.: Coherence in Modularly Composed
Adaptive Learning Documents, Proc. of AH2000, pp.375-379 (2000).
Adapting Web Interfaces by WHAT
Wilfred Ng and Jerry Yau

The Hong Kong University of Science and Technology, Hong Kong
{wilfred,csjerry}@cs.ust.hk
Abstract. We have previously presented a Java database application, called

WHAT, which is developed to analyse the users' navigation behaviour. Herein,
we discuss our on-going work, which makes use the output results from WHAT
to develop an enhanced interface for a commercial web site. The adapted inter-
face provides the additional frames of popular trails and pages, which help re-
duce the navigation problem and facilitate better web topology adaptation.
1 Introduction
Web usage mining is the process of extracting interesting patterns in web server logs
[1,2,6], which is related to the objectives of our WHAT system developed earlier on
[4,5]. WHAT (Web Hypertext Associated Trail) system is a Java-based application
developed for analysing web log data and mining web association rules as shown in
the architecture given in Fig. 1 (see [5] for details). WHAT is implemented on top of
an Oracle 8i server, which stores and manages the pre-processed web log data. We
also use Thin JDBC Driver and TCP/IP-based network. The essential idea is that
when users interact with a web site, data recording their behaviour is stored in the
corresponding web server logs. By analysing such log data, web designers/developers
can have a better understanding in the user navigation behaviour on their web sites.
Fig. 1. The Architecture of WHAT Mining System

552 W. Ng and J. Yau
2 The Methodology of WHAT System
The underlying principle of WHAT in analysing web log data is classified into the
following two areas: (1) To produce general statistical reports concerning web log
data, such as frequency counts of requested pages and their temporal information, and
(2) To mine the web association rules by using the modified Depth First Search
(DFS) algorithm and Incremental Step (IS) algorithm developed in [1], and then to
present the corresponding reports in appropriate visualisation tools. For the genera-
tion of statistical summary, users can impose queries to WHAT concerning analyses
on a web site such as ‘What are the total/average numbers of hits received per day?’,
‘What are the total/average numbers of pages viewed per day?’, and so on. The web
topology is modelled as a hypertext as proposed by [1]. A web server log file contains
information that characterises user accesses to the server, including the user identifi-
cations (or their IP addresses), the URL of the requested page and the timestamp of
the access. A user navigation session can be reconstructed as a sequence of page
requests such that no two consecutive requests are separated by more than 25 minutes
proposed in [2].
Fig. 2. A Weighted Directed Graph
We view a web site as a hypertext system. With the user navigation sessions recon-
structed, a weighted directed graph G = (N, E) can be built as follows: each node of N
corresponds to a visited page and each arc of E corresponds to a traversed link. Figure
2 presents a weight directed graph which models the navigation data embedded into a
web site. Since the model concerns only a collection of sessions, the concept of indi-
vidual session will be lost. Therefore, the resulting weighted graph is intended to
represent the navigation information of a group of users. Our aim is to mine trails of
these users in the manner of mass customisation. To capture the user's preferences
among the available links in a page, the concept of a web association rule can be used
to describe ‘when a user browses the hypertext system he or she will traverse the trail
with a certain probability’. A web association rule is expressed as (A1→A2) ∧ ! ∧ (An-
1→ An) where A1, … , An ∈ N and (Ai, Ai+1) ∈ E with i = 1, … , (n-1), meaning that
when a user traverses the hypertext system, there is a certain confidence that he or she
will follow the trail (A1, … , An). Confidence Cr in the rule is defined as the product of
the confidences of all the corresponding components as the equation given below:
n −1
Cr = ∏ C ( Ai → Ai +1 ) where C ( A, B ) = ( A, B)
i =1
∑ ( A, x)
{ x|( A, x )∈E )}
(1)
Adapting Web Interfaces by WHAT 553
In the graph shown in Fig. 2, C(B→C→D) = 5 (5 + 1) × 4 (2 + 4) = 0.55. Support Sr repre-

sents the average of times the links of the rule were traversed over the average of
times all the links of the graph were traversed. It is defined by the following equation:
(Rules with support greater than 1 consist of links traversed more than the average.)
 n −1   
S r =  ∑ ( Ai , Ai +1 ) (n − 1)   ∑ ( xi , xi +1 )
 {i|( x , x )∈E } E  (2)
 i =1   i i+1 
In the example of Fig. 3, S(B→C→D) = [(5 + 4) 2] [43 15] = 1.57. We adopt the modified
DFS algorithm and the IS algorithm proposed in [1] for mining all the trails in the
graph with support and confidence above specified thresholds.
3 Integrating WHAT with a Real Web Site
We demonstrate that WHAT adapts a commercial web site with respect to the server
web log and contributes to enhance the web interface. The web log files (74.8 MB) is
downloaded from [7]. The output results from WHAT, which provide the information
of a sequence of most popular trails and pages, are then fed into the web interface as
shown in Fig. 4. The raw web log data in the file is cleaned, transformed and stored in
Oracle database server by the cleaning module of WHAT. Note those entries gener-
ated by inline images are filtered out. Figure 3a presents some of the samples rules
mined with the DFS Algorithm under specified parameters of confidence, support and
rule lengths. Figure 3b presents the sample result of Requested Page Summary of the
ten most frequently requested pages (see [5] for details of the other services).
Fig. 3a. Sample Association Rules Fig. 3b. Screen Shots of Most Frequently
Identified Requested Page Summary
The interface of Hyperreal (in www.hyperreal.org/music/machines) is enhanced by

providing two additional frames given in Fig. 4: the first one is the frame of ‘hot
trails’, shown on the top of left-hand side, and the second one is the frame of the most
popular pages, shown on the bottom of the left-hand side. These frames are built upon
the output results from WHAT as already discussed in Sect. 2. The ‘Hot Trails’ frame
consists of the top 10 most popular trails defined by the association rules. We set the
confidence and support thresholds to be 0.5 and 1.0 in WHAT. The pages along a trail
can be collapsed into a ‘trail folder’ whenever necessary, and all the trail folders are
554 W. Ng and J. Yau
listed in the descending order of the support values. A visitor is able to click on the
hot trails, once opening the folder, to navigate the site following the trail suggestion,
or to go to any one of the most popular pages listed in the lower frame. In Fig. 4 we
can see the top 10 pages, which are the most frequently requested ones. We may pro-
vide in this frame other popular objects such as the most popular downloaded files, or
other most popular items depending on the related business objectives.
The Frame of The Frame of

Hot Trails Page Content
The Frame of
Popular Pages
Fig. 4. An Enhanced Interface of Hyperreal
4 Conclusions
This paper describes an integrated system WHAT, which is developed to mine web
association rules and to generate important statistics for web server log data on a real
web site. The underlying principle of this system is based on the formal web hypertext
model, which views the topology of a web site as a hypertext graph. Using the output
result of WHAT we develop an enhanced web interface, which provides a useful
guide for user navigating a complex web site and can be used to achieve the business
objectives such as placing advertisements along the most popular trails.
Acknowledgements
This work is supported by the Hong Kong Polytechnic University Grant

POLYU5095/00E and the Hong Kong University of Science and Technology Area-
Of-Excellence-IT Grant.
Adapting Web Interfaces by WHAT 555
References
1. Borges J. and Levene M. Mining Association Rules in Hypertext Databases.

Proc. of the 4th Int. Conf. on Knowledge Discovery and Data Mining, pp. 149 -
153, 1998.
2. Catledge L. D. and Pitkow J. E. Characterizing Browsing Strategies in the World
Wide Web. Computer Networks and ISDN Systems, 27(6), pp. 1065-1073, 1995.
3. Ng W. Evaluating the Client Side Approach and the Server Side Approach to the
WWW and DBMS Integration. Proc. of the 9th Int. Database Conf., pp. 72-82,
1999.
4. Ng W. Implementing WHAT to Support Web Usage Management. Proc. of the
International Conference on Internet Computing IC'2001, USA, pp. 35-40, 2001.
5. Ng W. and Chan C. WHAT: A Web Hypertext Associated Trail Mining System.
Proc. of the 9th IFIP 2.6 Working Conf. on Database Semantics, pp. 205-220,
2001.
6. Ng W. Capturing the Semantics of Web Log Data by Navigation Matrices. Proc.
of the 9th IFIP 2.6 Working Conf. on Database Semantics, pp. 169-183, 2001.
7. Raw Web Log Files from Hyperreal. http://www.cs.washington.edu/research/
adaptive/download.htm, 2001.
Considering Sensing-Intuitive Dimension
to Exposition-Exemplification in Adaptive Sequencing
Pedro Paredes and Pilar Rodriguez
Escuela Técnica Superior de Informática, U. A. M., Cantoblanco, 28049 Madrid, España

{Pedro.Paredes, Pilar.Rodriguez}@ii.uam.es
Abstract. This paper shows a way of using sensing-intuitive dimension of

learning styles of students in order to improve the efficiency of adaptive learn-
ing systems. Firstly, it introduces the procedure of extracting information about
sensing-intuitive students from the Felder-Soloman ILS questionnaire. Then, it
presents a mechanism of application of sensing-intuitive dimension to exposi-
tion-exemplification sequencing. The example used to explain the adaptation
effects is taken from a chess course developed with TANGOW, Task-based
Adaptive learNer Guidance On the Web.
1 Motivation
The theory of learning styles states that people have different approaches to learning
and studying [1]. We all have learning preferences that enable us to learn more effec-
tively. Adapting the course to the learning style of the student changes the point of
view of the learning process from teacher’s perspective to learner’s perspective. In
this sense, it is necessary to mention the experience of Arthur [2]. It is one of the first
systems which incorporates learning styles as a significant feature to their student
models.
There are a large number of learning styles models: The Myers-Briggs Type Indi-
cator (MBTI) derived from the theory of psychologist Carl Jung [3]; Kolb’s Learning
Style Model classifies students depending on how they perceive and process informa-
tion [4]; Herrmann Brain Dominance Instrument (HBDI) is based on four different
task-specialized quadrants of the brain [5]. Felder-Silverman Learning Style Model
categorizes an individual’s preferred learning style along a sliding scale of five di-
mensions [6,7]. One of them is called the sensing-intuitive dimension.
Sensing learners prefer learning first concrete and practical information oriented
toward facts and procedures while intuitive learners prefer conceptual and innovative
information oriented toward theories and meanings. In this paper we propose an adap-
tation procedure for moderate and strong sensing-intuitive learners, as detected by
means of the Felder-Soloman ILS questionnaire. In our approach, adaptation lies in
presenting examples before expositions to sensing learners and quite the opposite to
intuitive learners. In doing so, it makes use of this adaptation in TANGOW [8,9].

Sensing-Intuitive Dimension to Exposition-Exemplification in Adaptive Sequencing 557
2 Extracting Information about Sensing-Intuitive Students
Based on the Felder-Silverman classification, Felder and Soloman have built a ques-
tionnaire called ILS, Index of Learning Styles [10]. The aim of ILS questionnaire is to
determine the learning style preferred by each student. ILS questionnaire is still under
construction and nowadays it is formed by 44 questions with two possibly answers; a
or b. These questions are divided into four groups, with eleven questions each, corre-
sponding to four categories in Felder and Silverman’s classification.
To take measurements of student’s preferences in a given dimension, we have to
subtract the answers relating to one extreme to the other one of the same dimension.
Therefore, someone could be sensing-intuitive in a scale of odd values between 1 and
11. For example, if a student chose four of the answers classified as sensing and seven
of the answers classified as intuitive, then it is intuitive with a score of 7-4=3.
Felder and Soloman [10] interpret the results in sections. If you obtain a score of 1
or 3 you have a mild preference but you are well balanced. Otherwise, if your score is
5 or 7, you have a moderate preference and will learn more easily in teaching systems
which favor that dimension. Finally, if you score 9 or 11, you could have difficulty
learning in a system which does not support that preference.
We have chosen the Felder-Silverman model among the existing learning style
models because its ILS Questionnaire gives us the possibility of linking directly its
results to automatic adaptive environments. Thanks to the distribution of the ILS
Questionnaire in four different dimensions with two extremes we can build user mod-
els corresponding to each of these four dimensions.
3 Adapting Exposition-Exemplification Sequencing
Felder [7] claims that the goal of the teacher is to strengthen those learning styles that
students do not prefer. This procedure should improve the global learning capability
of the students. It is a long-term objective and the aim is that students perceive and
process information in all the ways. However in a Web based learning environment
we try to attain a short-term goal: an easiest and most effective learning process
through adaptation of learning courses to individual learning styles.
In this environment, we propose to make use of the ILS results to adapt the exposi-
tion-exemplification sequencing in a Web-based learning environment. The proce-
dure is as follows: firstly the student fills out the questionnaire; then the score ob-
tained points out the sensing-intuitive preference of the student: mild, moderate or
extreme; finally we use that preference to adapt the exposition-exemplification se-
quencing in case of moderate and extreme sensing-intuitive learners.
We have implemented the described adaptation procedure on TANGOW, Task-
based Adaptive learNer Guidance On the Web [8,9]. In TANGOW a course is de-
scribed in terms of Teaching Tasks and rules. Knowledge is represented by means of
TTs that need to be achieved. TTs may be exposition tasks (E), practical or examples
(e).
558 P. Paredes and P. Rodriguez
1
CONCEPTUAL TREE:
3
King Movements
1 E-e sequencing
Standard Special 2-E 2-e
Movements Movements
4-E 4-e 5-E 5-e
2 3
1
3
4 5
e-E sequencing
Short Long
Castling Castling 2-e 2-E
4-e 4-E 5-e 5-E
Fig. 1. Exposition-exemplification sequencing in sensing (lower) and intuitive (upper) stu-

dents in a piece of chess course about king movements.
Exposition TTs can have an example associated. Designers establish the default order,
that is, the order of the tasks and, consequently, the order of the task types (examples
first or exposition first). Our approach is to dynamically modify that last order ac-
cording to the learning style. If the student obtains a score of 1 or 3 his/her preference
to one or the other style is balanced and we apply the default order. On the contrary if
the student scores 5 or more than 5, the order of exposition-example tasks will be in
harmony with his/her learning preference.
In Fig.1 it is presented a conceptual tree with three E-e nodes. Nodes 2, 4, and 5
are E-e nodes; this means that student could see exposition before example (E-e) or
vice versa (e-E). This tree shows the King standard and special movements (short and
long castling). On the right side two different trees are presented showing the run-
time sequencing based on sensing-intuitive adaptation. The lower tree with e-E nodes
corresponds to moderate and extreme sensing students. The upper tree with E-e nodes
will be presented to moderate and extreme intuitive students. The default sequencing
is suitable for students with a mild preference and we assume that it is a common
practice to expose firstly and provide examples later (E-e sequencing).
This paper suggests the application of the results of the ILS questionnaire to adapt the
sensing-intuitive dimension to the exposition-exemplification sequencing. We assume
that a strong relationship between examples and their related theoretical expositions
Sensing-Intuitive Dimension to Exposition-Exemplification in Adaptive Sequencing 559
exists with independence of the conceptual representation. This relationship is more

powerful than any other sequencing consideration, and can eventually be combined
with other dimensions. In this sense, we are exploring the adaptation effects of com-
bining the sensing-intuitive with other dimensions [11].
In addition to sequential-global and sensing-intuitive dimensions, at present our re-
search examines the incorporation of other learning style dimensions, according to the
ILS, to the general model. It involves the study of possible conflicts between adaptive
actions based on different learning style dimensions and the resolution of these con-
flicts by establishing a priority of some dimensions against others.
Acknowledgments
The Spanish Interdepartmental Commission of Science and Technology (CICYT) has

sponsored this paper, project number TEL1999-0181 and project number TIC2001-
0685-C02-01.
References
1. Dunn, K., Dunn, R.: Teaching students through their individual learning styles. Reston,
VA.: National Council of Principles (1978)
2. Gilbert, J. E., Han, C. Y.: Adapting instruction in search of ‘a significant difference’.
Journal of Network and Computer Applications, 22 (1999).
3. Briggs Myers, I., McCaulley, M.: Manual: A Guide to the Development and Use of the
Myers-Briggs Type Indicator. Consulting Psychologist Press (1985)
4. Kolb, D.: Experiential learning: Experience as the source of learning and development.
Englewood Cliffs, NJ: Prentice-Hall (1984)
5. Herrmann, N.: The Creative Brain. Lake Lure, NC, Brain Books (1990)
6. Felder, R. M., Silverman, L. K.: Learning Styles and Teaching Styles in Engineering
Education. Engr. Education, 78 (7) (1988) 674-681
7. Felder, R. M.: Reaching the Second Tier: Learning and Teaching Styles in College Sci-
ence Education. J. Coll. Sci. Teaching, 23 (5) (1993) 286-290
8. Carro, R.M., Pulido, E., Rodríguez, P.: Task-based Adaptive learNer Guidance on the
Web. Proc of the 2nd Workshop on AS and UM on the WWW. CS Report 99-07, Eiden-
hoven University of Technology (1999) 49-57
9. Carro, R.M., Pulido, E., Rodríguez, P.: TANGOW: a Model for Internet Based Learning.
IJCEELLL, 11 (1-2) (2001).
10. Felder, R. M., Soloman, B. A.: Learning Styles and Strategies. December (2001) on-line
at: http://www2.ncsu.edu/unity/lockers/users/f/felder/public/ILSdir/styles.htm
11. Paredes, P., Rodríguez, P.; Tratamiento de los casos secuenciales-globales moderados y
extremos en un sistema de enseñanza adaptativa. Submitted to IPO2002, December (2001)
PlanEdit: An Adaptive Problem Solving Tool for Design
M.A. Redondo , C. Bravo , M. Ortega , and M.F. Verdejo2

1 1 1
1
Dpto. de Informática
Universidad de Castilla – La Mancha
Paseo de la Universidad, 4. 13071 – Ciudad Real. Spain
{mredondo,cbravo,mortega}@inf-cr.uclm.es
2
Dpto. de Lenguajes y Sistemas Informáticos
Universidad Nacional de Educación a Distancia
Ciudad Universitaria, s/n. 28040 – Madrid. Spain
[email protected]
Abstract. Experimental learning environments based on simulation usually

require monitoring and adaptation to the actions the user carries out. In this
work we present the monitoring and adaptive features of PlanEdit, a tool for
learning integral automation methods in buildings and housing by design and
simulation.
1 Introduction
Design and simulation environments for learning should provide students with
mechanisms to enhance effective learning. For that purpose, it is convenient to
interactively monitor the learner while he/she is solving a problem [1].
This work describes the monitoring and adaptability features of PlanEdit, an
evolution of the DomoSim environment [2], to learn the design of the automation of
buildings and housing. The environment included a knowledge-based editor to plan
the design of a model satisfying a set of requirements and a simulation program to test
the behavior of the model built.
The domain where a pressing need arises and where our investigation is applied is
the learning of the design of automated control facilities in buildings and housing,
also called Domotics. In this kind of training, the realization of practical experiments
is especially important. However, the material necessary to carry out these
assignments is usually expensive and in many cases is not adequately provided. This
problem gets worse with the difficulty to bring the student closer to real situations, to
replicate accidents and to simulate those chaotic situations, which may happen in the
real world; Domotics designs aim to cope with those troublesome situations too.
In order to soften this problem by means of the use of technology, we have
developed a distributed environment with support for distance learning of domotics
design: DomoSim-TPC [3]. It includes setting, realization, tracking, tasks analysis
and storage of collaborative learning activities. PlanEdit is one of the tools included in
this environment and it approaches learning from a individual perspective. This is a
supporting tool for the design of models that should satisfy a specification. The design
is approached in two steps, first a plan is built, and then a refinement of the plan is
carried out. A plan is specified as a partially ordered set of generic actions for the

PlanEdit: An Adaptive Problem Solving Tool for Design 561
construction of a model. We will now describe the characteristics considered in the

student model, the monitoring of the work carried out and how the system adapts to
the user.
2 Student Model
The user model consists of entities with information about the following categories:
- The student's profile that defines the student’s role in the system. This profile
stores restrictions and obligations on the type of actions that the student should
carry out.
- The user's interaction with the editor in order to plan the design.
- The sequence of actions that the student dynamically specifies in his/her search of
a solution to a problem. Additionally, the time dedicated to the elaboration and the
mistakes made in this process are considered. This way, each element in this
sequence is defined by a unique identifier, the moment the action is planned, the
mistake associated to it (if there is one) and the action preceding the current
action. This is, <item>::= <id> <time> <action> <prevAct> |
<id> <time> <action> <error> <prevAct>, where, <action>::=
<kind> <area> <section> <element>, and <error>::= ACTION
| SEQUENCE | DUPLICATE | NON_OBLIGATORY
3 Tutoring Module
The main objective of tracing the student’s resolution strategy is to determine if
he/she is close to a good solution. In many cases, there is not a unique solution. The
Optimal Plan of Design (OPD ) represents the set of possible solutions for a particular
problem. If he/she moves away from what the experts consider a good solution, the
system will display warnings offering help and reinforcement with the purpose of
communicating the student the mistakes he/she is making and driving him/her
towards a better solution. This circumstance can make the student reflect in search of
an explanation and the correction of the mistake. This way, hopefully, meta-learning
situations are promoted.
Figure 1 shows a schema which summarizes the mechanism used for monitoring
and guiding the learner's work. The beginning is the General Plan of Design (GPD)
for a generic kind of problem. In this plan there is a sequence of elements following
this schema <action>:<type>:<requirements>:<influences>.
The system can modify the GPD according to the specification of the parameters
that characterize the problem. As a result, the OPD is obtained for the proposed
problem. This constitutes a resolution schema with some flexibility in the tasks to
carry out. Additionally, the teacher sets up a help level that will display the kind of
messages that the system will provide in order to warn the student or to prevent the
mistakes that the solution proposed by the student can produce. Therefore, the
generated and displayed help messages will be a consequence of an inference process
starting from the OPD, the help level and the design proposed by the student.
562 M.A. Redondo et al.
Fig. 1. Components involved in the mechanism of tracing and monitoring the construction
procedure of the solution to a design problem.
In order to increase the controlled flexibility of the OPDs, a set of rules can be
defined. These rules define the form in which a strategy can be changed and the
possible side effects of the changes. We consider three sources of information that can
cause a variation in the structure of the resolution of design problems:
i) The characteristics of the problem

The design problems with similarities in their solutions can be classified into
categories. For each category a general strategy of resolution is defined: the GPD.
Nevertheless, a particular problem will be identified by a set of parameters. In
some cases these are fixed and in others, variable. These parameters may condition
the solution and allow adapting the construction mechanism to each student. To every
parameter we associate a rule of the type IF <condition> THEN
<modifyAction>. This rule can modify the design actions and their structure. This
is, the adaptation depends on the domain.
ii) Design actions

The design strategy is defined according to a series of actions with a high level of
abstraction. Among these certain dependences are established. The order the student
establishes when approaching each generic action defines his/her resolution strategy.
When reaching a state of the design at which different alternatives to continue are
offered, the student chooses one of them and discards the rest. He/she may choose a
non-obligatory option, but the fact of selecting it will imply the obligatory realization
of other actions. Therefore, the learner's decision will cause an adaptation of the OPD
depending on the alternative chosen.
PlanEdit: An Adaptive Problem Solving Tool for Design 563
iii) Help level

When the teacher proposes an activity in which a problem should be solved, he/she
specifies the help level. This level makes the use of scaffolding [4] possible, defining
the degree of difficulty of the problem and the level of help that the system offers to
the student. This will reinforce the intrinsic structure of the process of problem
solving in which the new knowledge is built through a successive process of
elaboration and integration.
The help level influences PlanEdit in two ways: the messages that will be
displayed in order to guide the learner during their work; and the set of possible
actions that will be shown to the student. The learner has to choose the most
appropriate action in each moment from them.
4 Conclusions
PlanEdit considers several aspects of the users, such as their procedural strategy to
solve a problem as a history trace of the path followed and their behavior as to
interaction with the system. With all this information, the tool adjusts the current
alternatives at each stage in the resolution process (generation and adaptation of
nodes). This adjustment shapes up the presence of reinforcing help and the
concealing/prohibition of alternatives to guide the learner around the resolution
process (adaptation of the path).
In our proposal, the path the learner outlines has to match one of the plans.
However, the outstanding characteristic of our tool in contrast to hypermedia systems
is that the learner does not follow a content-based path. Rather, he/she builds a path as
a plan or a solution strategy to a proposed problem. The system adapts to the
characteristics of the solution the students propose, trying to guide them so that they
can achieve their objective.
Our proposal is applicable to any domain where the resolution of design problems
in a structured way is approached. This resolution can be specified in terms of the
design actions to build a model. These actions can be inferred from the user
interaction.
References
1. Verdejo, M.F., (1992). A Framework for Instructional Planning and Discourse Modeling
in Intelligent Tutoring Systems. In E. Costa (ed.), New Directions for Intelligent Tutoring
Systems. Springer Verlag: Berlin, pp. 147-170.
2. Bravo, J., Ortega, M., & Verdejo, M.F., (1999). Planning in Distance Simulation
Environments. Full Paper in Communications and Networking in Education
COMNED’99. Aulanko, Hämeenlinna, Finlandia.
3. Bravo, C., Redondo, M.A., Bravo, J., & Ortega, M., (2000). DOMOSIM-COL: A
Simulation Collaborative Environment for the Learning of Domotic Design. Reviewed
Paper. Inroads - The SIGCSE Bulletin of ACM, vol. 32 (2), pp.65-67.
4. Rosson, M.B., & Carroll, J.M., (1996). Scaffolded Examples for Learning Object-Oriented
Design. Communications of ACM, vol. 39, num. 4.
Dynamic Generation of Interactive Dialogs
Based on Intelligent Agents
Martín González Rodríguez, Benjamin López Pérez, María del Puerto Paule Ruíz,
and Juan Ramón Pérez Pérez
Laboratory of Object Oriented Technologies (OOTLab – HCI Area) – Department of

Computing Science, University of Oviedo, Asturies, Spain
{martin, benja, paule, jrpp}@lsi.uniovi.es
Abstract. The design of the lexical and syntactical levels of the user interface
of any web based system is commonly based on general human-computer
interaction principles targeted to match the interaction requirements of the so-
called typical user, an abstract generalization of each user of an application.
However the identification of such typical user at the web design stage seems to
be an impossible task, considering the dynamic nature of the web, where the
user interface of a popular web systems can be used by thousands of different
users everyday. In order to avoid this problem, we have designed GADEA, an
intelligent user interface management system able to detect different kinds of
users by mean of distributed data gathering agents. The information obtained by
those agents is used to design interactive dialogues at execution time, which are
adapted to the specific cognitive, perceptive and motor characteristics of the
current user.
1 Introduction
Traditional design of user centred interfaces is based on the identification and

definition of the target audience of the application under development. Some design
guidelines include the identification and understanding of the target audience as the
most important steps to start designing a product [1, pp. 13; 2]. The idea is that once
the target audience has been understood an interface that effectively satisfies their
needs can be designed.
However, the quest for the typical user is opposite to the individuality and diversity
that makes up some much of our identity. If the design of the interaction mechanisms
of an application aims to make interfaces accessible and appealing to all users, it
shouldn’t rely on an abstract generalisation [2]. In the design of applications targeted
to a wide range of users it is almost impossible to determine the typical user without
falling in serious misconceptions. For example, we should ask ourselves whether its is
possible to describe the typical user of Microsoft Office or not. Another interesting
question without a clear answer is whether it is possible or not to identify the typical
user of generic web portals visited by thousands of user everyday such as for example
Lycos, Terra or MSN. For some authors such as Schneiderman [3], the answer to
those questions is definitely ‘no’ as ‘there is no average user’. For him, ‘either

Dynamic Generation of Interactive Dialogs Based on Intelligent Agents 565
compromises must be made or multiple versions of a system must be created’ in order

to satisfy the specific interaction needs of every single user.
Following this design principle we have developed the GADEA User Interface
Manager Expert System (UIMES) for the development of automatic generic adaptive
user interfaces. The tools provided by this system allow programmers to create
multiple versions of the user interface of an application dynamically. Each version of
the user interface is created and adapted dynamically in order to satisfy the interaction
requirements of the current user of the application. The design of each version will
depend on the specific state of the cognitive, perceptive and motor system of the
mentioned current user.
Current version of this UIMES has been entirely written in Java, so it can use as an
alternate User Interface Manager System (UIMS) for platforms such as Windows,
Linux or Macintosh. The internal design of this development framework also includes
the generation of dynamically adapted user interfaces in the web engineering and
adaptive hypermedia environments.
2 Overview of GADEA
User
Fu Gu
zzy Sets idelines
interactive DEVA
discourse (Discourse Expert Valuator for Adaptation)
CodeX
(Code eXplorer)
Us
er Mode
l
Se
Code mantic
ANTS
(Automatic Navigability Testing System)
Se
Code mantic
workspace
Fig. 1. GADEA’s main components

566 M. González Rodríguez et al.
The internal architecture of GADEA relies on three independent components

specially designed to cope with the problems derived of the three most important
features of the system. Those features are the separation of the client application’s
user interface from its functionality, the creation of a single tailored interactive dialog
for each user of an application and the constant update of each user model. The
components that address each problem are CodeX, DEVA and ANTS respectively
(see Fig. 1).
The CodeX (Code eXplorer) module represents the interface between GADEA and
the client applications of this UIMES. This module must convert the user interaction
requests into calls to specific methods inside the application’s domain workspace. To
reach this objective, this module automatically inspects the binary code of every client
application at execution time, looking for any user process defined by the
programmers during the design time. CodeX consider a ‘user process’ as those
methods designed to satisfy a specific user’s interaction request and will represent
them as options or commands in the respective application user interface. This
module is also in charge update every single piece information displayed by the user
interface, keeping track of possible changes in the value of public variables and data
structures associated to the interface by the application at both execution and design
time. All this information is obtained automatically inspecting the client application’s
code by mean of the structural reflection mechanism of the Java platform also
included as part of the servlet specification.
The interactive dialogs established to interact with the user inside the scope of ever
‘user process’ are created from the data that the application needs to perform a certain
task (provided by the user) or by the information that must be displayed as the result
of a user data request.
The current version of our prototype supports four kinds of data primitives used to
create an interactive dialog with the user. Those data primitives are strings, integers,
real and dates. Each of those primitives is encapsulated inside a JavaBean component
[4], which can be configured by mean of their own set of properties. Depending on
the kind of data required by the situation and kind of user process (input or output
process) the programmer must select and configure the appropriate primitive data
component, inserting it in an ‘interactive dialog object’. The primitive data
components can be grouped together creating semantic meaningful blocks called
‘chunks’. Semantically the chunks represent groups of related data elements that must
be perceived as a conceptual unit by the user. Internally, every chunk contains
primitive data components as well as other chunks, allowing the creation a hierarchy
of chunks.
The information collected by CodeX is sent periodically to DEVA (Dialog Expert
Valuator for Adaptation) which represent the agent software in charge of the
simulation of the behaviour of human expert in Human-Computer Interaction (HCI).
This module converts the application’s and user’s interaction requests into adaptive
interactive dialogs. Based on the general knowledge and guidelines provided by the
HCI discipline as well as by specific knowledge stored about the current user model
of the client application, DEVA uses its fuzzy inference engine to evaluate, to select
and to adapt the best interactive dialog available for the current user. The dynamic
creation of the user tailored interactive dialog will depend on the current user’s
cognitive, perceptive and motor skills, which might fluctuate over the time.
Finally, the ANTS module will be the agent responsible for keeping the
information about the users of any client application of GADEA updated. This
Dynamic Generation of Interactive Dialogs Based on Intelligent Agents 567
component makes use of different kinds of automatic remote agents designed to

observe the user behaviour in any of the interactive dialogs designed by DEVA.
Those agents obtain crucial information about the specific user’s skills, which is
stored in specific distributed user model.
3 Conclusions
The automatic code exploring system provided by GADEA allows an easy adaptation
of the low-level aspects of a user interface with little or almost no programming
effort. The approach adopted by CodeX in the design interactive dialogs in terms of
data primitive components instead in terms of static-based widgets allows designers to
focus their effort in the design of the high level interaction mechanisms (metaphors)
of their applications saving up time and resources.
The information captured by ANTS in every interactive dialog allows DEVA to
adapt the user interface to the current state of the user’s cognitive, perceptive and
motor system, which tends to fluctuate over time.
The use of an expert system (DEVA) to take every decision concerning with the
interaction style and techniques to be applied, guarantees a strong user interface
homogeneity in every application. As there isn’t any ambiguous natural language-
based HCI guideline to be interpreted by a human designer, this approach eradicates
troubles with consistency, which it is present in many computer platforms.
Currently, we are working on the design of a low-level adaptive Internet browser
based entirely on GADEA’s technology, able to be used with the same utility by
sighted and blind users, as well as users with different degrees of physical disability.
References
1. Apple Computer Inc; (1992) Macintosh Human Interface Guidelines. Adison-Wesley

Publishing Company. ISBN 0-201-62216-562216.
2. Reynolds, C.; (1997) A Critical Examination of Separable User Interface Management
Systems: Constructs for Individualisation. ACM SIGCHI. 29(3).
3. Schneiderman, B.; (1987) Designing the user Interface. Reading, MA: Addison-Wesley.
4. Sun Microsystems (1996). The JavaBeans 1.0 API specification.
http://www.sun.con/beans.
A Simple Architecture for Adaptive Hypertext
Oreste Signore
CNR-CNUCE, Area della Ricerca di Pisa - San Cataldo, via Moruzzi, 1 - 56124 Pisa
[email protected]
Abstract. We describe a general, flexible architecture based on web standards

(W3C Recommendations) where users access data through the mediation of in-
telligent agents. Architecture is based on simple but effective document and
user models, with weighted and semantic tagging of both documents and links.
User's profile and preferences affect the agents' behaviour, so that actions can
occur on the server as well at the browser side. Profile and preferences accom-
pany the URL reference, avoiding dependence from any central site.
1 Introduction
The Web is a Universal Information Space, and it is easily seen that there is need for
adaptive and intelligent systems. User needs in accessing information have been
widely addressed by researchers ([1,2]), and some possible solutions, like information
filtering, user profiling and two-level hypertext seem to give good results with a fairly
low overhead and complexity ([3,4,5,6]). An architecture where users access data
through the mediation of intelligent software agents can be a first step towards the
Semantic Web [7]. In this framework, a basic role is played by XML, as the basis for
the exchange of structured information. However, we should carefully consider the
effect of deep structuring and tagging. In the following, we will describe ABWR
(Adaptive Browsing of Web Resources), a general, flexible and web standards com-
pliant architecture where an intelligent software agent performs information filtering
based upon users' profiles and preferences.
2 Architecture of ABWR
The architecture has been driven by some basic principles, namely simplicity (evident
in both the underlying principles and the architecture), no centralization (as architec-
tures relying on some central site would be in contrast with the ideas that lead to the
development of the web, as described in [8]), light tagging (to emphasize the seman-
tic categories, more than the specific role played by the single, atomic information
item), weighted and semantically tagged links (to screen out useless links and suggest
appropriate interaction paradigms).

A Simple Architecture for Adaptive Hypertext 569
ABWR mainly relies on modules able to perform document filtering and personal-
ization. Some specific modules check the user behaviour during navigation. The ar-
chitecture aims to reduce the centrality of the agent, split in two equivalent compo-
nents: the ServerAgent and the ClientAgent. Generally speaking, ClientAgent plays an
active role when user behaviour is not stable or the server is unable to perform ac-
tions. The agent gets appropriate directives from the user Preferences. To avoid any
centralization and have a generally valid profile, the browser, when asking the server
for data, sends current profile and preferences.
Document Document Topics

semantics hierarchy
1..1
Data Server Document
structure
(from Use Case View)
1..n
n 1..n
Part Topic
1 0..n
Weight
Link Info
LinkItemWeight TopicWeight
1..n
LinkItem LinkType UserProfile
0..n 1 0..n
User
LinkTypeWeight/Threshold (from Use Case View)
Fig. 1. User, link and document model
The ABWR architecture (Fig. 1) is based on an abstract representation of docu-

ments and user, based on Document model, Link taxonomy and model, User model,
Preferences. They can easily be described according to the RDF syntax ([9]).
Semantic model of documents. Traditional documents are often organized in parts
(sections, paragraphs, etc.). XML documents, where every field is recognizable and
appropriately tagged, have more complex structures. However, we must clearly dis-
tinguish between structural and semantic information. Assigning semantics to the
document as a whole can just accommodate for very simple and flat documents. In
the majority of cases, we need a finer granularity level, with semantics attached to
specific parts of documents.
A relevant issue is to consider that users should be able to match their topics of in-
terest even if they are not experts in the field. Too specific topics would affect the
user/system communication. This approach leads to the definition of a limited and
rough general set of semantic categories (like Art, Economy, Politics, Music, Sport,
Literature, Science) that can be shared by a wide variety of users, and can be used to
570 O. Signore
define a user profile and to semantically characterize both various parts of the docu-
ments1 and links. Topics identifying concepts belonging to this predefined set of cog-
nitive categories in the semantic model can be tied to the document as a whole, or to a
document part. We have also to specify a weight, stating the relevance in the docu-
ment context. Every Topic is modelled as an element of a hierarchy of Topics, so
building a thesaurus, even if limited to a few terms and of reduced depth.
Link taxonomy and model. Links are an essential component of hypertext, allowing
the navigation on the web, and can implement the abstraction mechanism needed to
move from the data space to the concept space2. In the data space, documents are
connected by extensional links; in the concept space, associations among concepts
implement intensional links among documents ([3,4,5]). Semantic qualification of
links is equivalent to explicitly identify the meaning of the link in the document and
the role of involved resources. The reason why the link has been inserted in the
document, that is the nature of association (geographical, explicative, etc.), can be
addressed as link semanticType. To define a profile independently from the context, it
is useful to select a limited number of semantic types (Location, Time, Person, Lin-
guistic resource, etc.) It is worthwhile to note that different types of links, and there-
fore of the referred resource, can suggest different, and specialized, interaction para-
digm (time, map, classification, etc.). This can have a terrific effect on the potential
association mechanism: two documents can be linked through an intensional link
existing in the concept space, without having any extensional link specified in the
document ([10]). The semantic model of links should be appropriately mapped onto
the W3C Xlink Recommendation ([11]).
User Model. As a first approximation level, user model is defined in terms of a pro-
file. User profile is essentially dynamic. In our approach, the user profile specification
is tightly related to the semantic model of documents and links: user profile is defined
in terms of semantic categories, link types and link roles. For each of them, the user
profile will specify a degree of interest, assigning a weight. In addition, for each topic
or link type, a threshold value is defined.
Preferences. Several factors can affect filtering and personalization of documents.
Among them, device characteristics, abilities and behaviour (accounting for dynamic
aspects of the profile). Preferences act as directives to the agent, specifying, among
other properties, where manipulation should be performed (server or client side), if
the profile should be considered stable, the intended behaviour when user profile
topic is at a different detail level respect to document assigned topic.
A sample implementation. A prototype has been implemented3. There is an exten-
sive use of W3C Recommendations ([12]). Up to now, it has been tested on a very
limited document base.
1 In this paper, we will generically refer to ‘documents’ irrespective the sites will provide just
a set of (somehow indexed and searchable) ‘pages’, or dynamically access Information Re-
trieval Systems, or produce ‘documents’ from a traditional database.
2 The concept space is a simplified version of Semantic Web architecture’s ontology level.
3 Thanks are deserved to A. Cristofaro, who implemented a first version of the prototype.
A Simple Architecture for Adaptive Hypertext 571
3 Conclusion and Future Developments
Considering the difficulties that arise from the large amount of documents available
on the Web, and the growing interest in the semantic web research area, we proposed
a simple and hopefully effective approach, which resulted in a system architecture
and an operating prototype.
The relevant features are simplicity of architecture, flexibility of document, link
and user models, independence from centralized resources, intelligent software agent
operating on the server or the browser, depending on user's willingness or server
capabilities, openness to further enhancements, adherence to W3C Recommendations
and trends, like Semantic Web.
We are considering several future developments, like supporting a set of filtering
algorithms, more sophisticated user profile, automatic assignment of topics to docu-
ments, full support of new W3C Recommendations.
References
1. http://wwwis.win.tue.nl/ah/
2. http://ah2000.itc.it/
3. Signore O.: Issues on Hypertext Design, DEXA’95 - Database and Expert Systems Appli-
cation, Proceedings of the International Conference in London, United Kingdom 4-8 Sep-
tember 1995, Lecture Notes in Computer Science, N. 978, Springer Verlag, ISBN 3-540-
60303-4, pp. 283-292
4. Signore O.: Modelling Links in Hypertext/Hypermedia, in Multimedia Computing and
Museums, Selected papers from the Third International Conference on Hypermedia and
Interactivity in Museums (ICHIM’95 - MCN’95), October 9-13, San Diego, California
(USA), ISBN 1-88-5626-11-8, pp. 198-216
5. Signore O.: Exploiting Navigation Capabilities in Hypertext/Hypermedia, Proceedings of
HICSS '96, Maui Hawaii, 3-6 January 1996
6. Brusilovsky P.: Adaptive Hypermedia, User Modeling and User-Adapted Interaction 11:
87-110, 2001, on-line at:http://umuai.informatik.uni-essen.de/brusilovsky-umuai-
2001.pdf
7. http://www.semanticweb.org/
8. Berners-Lee T.: Weaving the Web: The Original Design and Ultimate Destiny of the
World Wide Web by Its Inventor, Harper San Francisco (1999), ISBN 0-06-251587-X
9. Lassila O., Swick R.: Resource Description Framework (RDF) Model and Syntax Specifi-
cation, W3C Recommendation 22 February 1999, http://www.w3.org/TR/REC-rdf-syntax
10. Signore O., Bartoli R., Fresta G., Loffredo M.: Implementing the Cognitive Layer of a
Hypermedia, Museum Interactive Multimedia 1997: Cultural Heritage Systems Design
and Interfaces – Selected papers from ICHIM 97 the Fourth International Conference on
Hypermedia and InterActivity in Museums, Paris, France, 3-5 September, 1997, (Edited
by David Bearman and Jennifer Trant) Archives & Museum Informatics (1997), p. 15-22,
ISBN 1-885626-14-2
11. XML Linking Language (XLink), W3C Working Draft 21 February 2000
http://www.w3.org/TR/xlink/
12. http://www.w3.org/TR/
Adaptive Learning Environment
for Teaching and Learning in WINDS
Marcus Specht, Milos Kravcik, Roland Klemke, Leonid Pesin,

and Rüdiger Hüttenhain
Fraunhofer Institute for Applied Information Technology FIT

53754 Sankt Augustin, Germany
{specht, kravcik, klemke, pesin, huettenhain}@fit.fraunhofer.de
Abstract. This paper introduces one approach to e-Learning describing the

Adaptive Learning Environment (ALE). The system provides a new
methodological approach to design education on the web. ALE will be used to
build a large knowledge base supporting Architecture and Civil Engineering
Design Courses and to experiment a comprehensive Virtual University of
Architecture and Engineering Design in the project WINDS. Here we outline
the system architecture and present its learning environment. The system
combines classical structuring of learning materials based on reusable learning
objects with an alternative structure - the course index.
1 Introduction
Web-based Intelligent Design and Tutoring System (WINDS) [1] is an ongoing

European project1 with the objective to implement a learning environment integrating
an intelligent tutoring system, a computer instruction management system and a set of
cooperative tools. We have developed the current version of the system and our
project partners prepare on-line courses that should be used from autumn 2002 at 21
universities.
ALE produces individualized courseware for the students depending on their
current state of knowledge, their preferences and learning styles. The author specifies
the metadata according to which various adaptation methods can be applied. In this
way, the system can adapt the sequence of learning objects according to the chosen
learning strategy. As a difference to earlier adaptive learning management systems
ALE also integrates the possibilities of using user modeling methods and data mining
for building expertise networks and supporting human resources management and
knowledge management for organizations. High quality expert and co-learner finding
algorithms can build on detailed user and learner models from ALE.
This paper describes the system architecture, presents the adaptive learning
environment, and outlines future perspectives.
1 Web based intelligent design system (WINDS) is EU funded project in the 5th framework of
the IST programme # IST-1999-10253.

Adaptive Learning Environment for Teaching and Learning in WINDS 573
2 Learning Objects and Roles
In ALE authors can create reusable and sharable learning objects and specify for them
metadata and relationships according to standard specifications like LOM [2] and
DCMI [3]. Learning objects are built of content blocks. Every content block has a
pedagogical role what gives additional possibilities for reusing and even recompiling
dynamic learning elements from content blocks of different pages. By taking into
account the pedagogical role of each content block also pedagogical patterns for
learning elements can be identified.
Additionally all learning objects in a course unit are linked dynamically by the
underlying index defined by the course authors. The index reflects a common
understanding of the domain and the subject matter. The index terms are described by
a term name, a description, synonyms, and relations to other terms. A learning object
can be viewed with highlighted index terms. This allows giving the learner different
approaches for working through the learning materials.
A user of the WINDS system can have one or more of the following roles
assigned: student, author (teacher), tutor, and administrator. Each role has access to
one special environment [4,5]. Additionally each user has a workspace to store and
exchange documents with colleagues. Synchronous and asynchronous collaboration
facilities are available too.
3 Adaptive Learning Environment
The WINDS learning environment is designed to support various learning strategies.

Paragraphs contain materials for expository (explanatory) education. Exploratory
(discovery) learning is encouraged not only by hyperlinks but also by index terms and
their interconnection with learning objects and external documents. Collaboration
facilities promotion of constructivist learning approaches.
The presented content can be adapted to the current knowledge level of the student
on the base of the User Model and the structure of learning objects. The user model
contains for each user all his or her educational events, called user episodes. These
episodes include events of different types:
• all the user’s actions with the learning objects
• all evaluations of completed tests made by the system
• teacher’s reviews of homework submitted by the student
• self assessment and self estimation of learners
Based on all these data the system can infer conclusions about the user’s
knowledge and goals. These are stored in the user model as well. Thus the user model
always reflects the current state of the user’s progress. The information is available
both for the teacher to control the student’s study process and for the system to adapt
the course presentation and navigation for the student. The learning environment in
WINDS consists of the following parts.
574 M. Specht et al.
3.1 Learning Object Display
The main part of the screen is occupied by the content of the current learning object.
Its complete path in the course tree is displayed together with navigation buttons to
support orientation and navigation in the course. A learning element can be displayed
with or without emphasized index terms hyper-linked with their explanations. Content
blocks in learning elements may be represented by texts, images, hyperlinks,
multimedia, or special formats (e.g. Word, PDF, PPT). Some of these content blocks
can be displayed as icons activating additional window possibly with a note or more
detailed information.
3.2 Annotated Course Structure
In another frame of the WINDS learning environment the course structure overview is
displayed. It supports student’s orientation in the course and helps to choose suitable
learning objects for the next study providing annotations for the learning objects.
The states of the learning objects in WINDS can be considered from several points
of view or in several dimensions. The system distinguishes three of them: interaction
history, tested knowledge, and user’s readiness. Additionally some other
characteristics can be defined depending on the followed objectives, e.g. current task
or context. Interaction history and tested knowledge can be expressed by quantitative
information defining the extent of the seen learning object or the tested knowledge,
user’s readiness is nominal information and depends on the prerequisites specified by
the author and on the interaction history of the user.
To represent various states or dimensions of an information space various visual
emphasis techniques can be used. Some studies (e. g. [6]) of visual perception
recommend the following visual attributes for different kind of information:
• Nominal information: texture, color (hue), position
• Quantitative information: length, position
Currently the system uses icon annotations. Other applicable annotation techniques
include hiding and presentation without hyperlinks. These can be also student’s
options.
3.3 Course Index
With each learning object corresponding index terms are displayed that provide a
means to interrelate heterogeneous course contents and to find individualised paths
through the learning materials. The WINDS course index component maintains the
index terms together with their respective descriptions, synonyms, relations between
terms, their occurrences in the course materials as well as in external documents. The
student can access all this information about a specific term by choosing it in the
Index frame. The index component can retrieve and highlight occurrences of index
terms within the course materials as well as within registered external documents
(such as web sites).
Adaptive Learning Environment for Teaching and Learning in WINDS 575
3.4 External Documents
External documents relevant to the course domain explain in more detail some
specific issues or provide up-to-date information like specialized portals do. Such
external documents serve also as resources for homework and projects. These
materials can go into more details than the course or give alternative views of the
domain. In the Documents frame the student can choose an external document related
to the current learning object and view this document either with or without
emphasized index terms.
3.5 Feedback and Recommendation Area
In the feedback and recommendation area the students can see their progress.
Additionally several recommendation strategies will be implemented which guide
learners in the course (prerequisite warnings, next step recommendations), learning
style support (learning material preferences, learning activity selections) and
cooperation initialisation (co-learner finding, expert finding and tutor support).
4 Conclusions and Further Research
This paper presented the current state of the ALE system, especially its learning
environment providing adaptive courseware. The learning objects can be delivered in
a variety of ways. The indexing system and the connection of the learning objects
allows for personalized coaching of students. We plan to extend the coaching
functionality by pedagogical agents. Different tutorial and navigation support
strategies will be implemented that allow the teacher to adjust the adaptive method to
the concrete target group and pedagogical framework. Open exploration of contents
will be supported on a level of index terms that where already explored by a learner.
In more strictly focused training settings the systems will keep the student close to the
default path of a curriculum based on the teacher’s specifications.
References
1. WINDS project: http://www.winds-university.org

2. LTSC IEEE Draft Standard for Learning Object Metadata, IEEE P1484.12/ D6.1 (2001)
3. Dublin Core Metadata Initiative (DCMI), http://dublincore.org/documents, DCMI
4. Specht, M., Kravcik, M., Pesin, L., Klemke, R.: Integrated Authoring Environment for
Web Based Courses in WINDS. Proceedings of ICL2001, Villach, Austria (2001)
5. Specht, M., Kravcik, M., Pesin, L., Klemke, R.: Authoring Adaptive Educational
Hypermedia in WINDS. Proceedings of ABIS2001, Dortmund, Germany (2001)
6. Bertin, J.: Semiologie Graphique. Paris: Editions de l’Ecole des Hautes Etudes en Sciences
(1967/1999)
Adaptation through Interaction Modelling
in Educational Hypermedia
Elio Toppano
Dipartimento di Matematica e Informatica

Università di Udine, Via delle Scienze 206, Loc. Rizzi, 33100 Udine, Italia
Abstract. The poster describes MMforTED, an adaptive educational

hypermedia system supporting the acquisition of several conceptualisations that
can be used for reasoning about technical artefacts from different perspectives.
Adaptation is performed by a next button that leads the learner to the best page
to be read next (direct guidance). The decision which page is best is taken by
analysing the properties of the course of interaction of the student up to a given
time and selecting the next page on the base of the affordances of potential
interaction it offers.
1 Introduction
Adaptive hypermedia are becoming increasingly popular as web-based educational

systems [1,2,3]. This poster describes a cognitive tool called MMforTED - an
acronym for MultiModelling for Technical Education - that has been developed to
foster the acquisition of several conceptualisations that can be used to represent and
inquiry about technical artefacts from different perspectives. The system enables
users to browse through a network of HTML pages illustrating descriptions of several
devices together with prototypical questions, ontologies and exercises supporting the
analysis and comprehension of a device structure and functioning. MMforTED has
been designed to be usable for diverse learning goals and thematic threads, so we do
not impose a single overarching perspective or global narrative structure to the
material. As shown by the experimental activity performed with students of various
ages and levels of proficiency, the lack of guidance produces disorientation and
cognitive overload especially in novice learners. To tackle this problem we have
introduced adaptivity through direct guidance [4]. Direct guidance implies that the
system informs the student which of the links on the current page will drive him or
her to the ‘best’ page in the hyper space. The best page is decided on the basis of
some characteristics of the user and on the base of the user's behaviour as it will be
explained in the following sections.
2 Adaptation through Interaction Modelling
The approach we followed to implement direct guidance is inspired to a recent

proposal made by Akhras and Self [5]: instead of modelling what the student knows

Adaptation through Interaction Modelling in Educational Hypermedia 577
we focus on the course of interaction and on the process of learning. More

specifically, we assume that a student using the MMforTED system passes through a
series of situations <(S1, t1); (S2, t2); ...,(Sn,tn) >, specifying a course of interaction,
via a sequence of events/actions (e.g. by activating hyperlinks). There are six different
types of situations S depending on the educational role of the pages that constitute the
hyperspace: situations showing a model of an artefact, representing assumptions lying
behind a model, providing the ontology used to build a model, presenting exercises,
questions, or relations between two models of the same artefacts. Each situation type
has a content represented by the specific instructional material it contains and affords
to a learner a specific pattern of interaction. The same situation type can be visited by
the student several times during a course of interaction. What must be stressed is that
the content of a situation is seen more as a resource that can be utilised, accessed, etc.
rather than a target knowledge that must be acquired.
When the student has reached situation (Sn,tn) there are several events (e1, e2, .. ,
ek) that could possibly follow. Each event choice extends the course of interaction in
a specific way. The aim of the adaptation mechanism is to select and suggest the
event that will optimise the learning opportunity for the student. There are basically
two ways for doing that. The first way is to select a situation (Sn+1, t+1) that affords
to the learner a pattern of interaction with desired properties. For example, the
mechanism can try to reduce the cognitive overload of a novice student by selecting
‘simple’ instructional material according to some criterion as discussed in the next
section. The second way is to select a situation that affords to the learner a course of
interaction <(S1, t1); .. ;(Sn+1,t+1)> that exhibits particular regularities which are
considered conducive to learning. In both cases, the adaptation mechanism tries to
provide profitable spaces of interaction to the learner based on some model of the
affordances of potential situations. This is a different focus to the traditional student
model that would be evaluating what a student has learned up to a given point.
2.1 Selecting Simple Models
A model simplification can be made by changing its conceptual content, the symbolic
representation or both these components. We use conceptual schemes to specify the
content of a model that is the types of entities, relations and properties (i.e. attributes
with associated domains of values) that have been used to build the model. Plex
structures are, instead, employed to externalise models [6]. A plex structure is a kind
of diagrammatic representation that is formed by interconnecting elements with N
attaching points called ‘napes’ by means of tie points. The conceptual schema
associated to a model gives the elements of the plex structure a semantic content.
In order to assess the ‘simplicity’ of a model we have introduced some metrics
that can be used to characterise internal features of a description. Internal features are
those that can be measured by examining a system description (the plex structure) on
its own separate from its environment. The metrics for internal features are [7]:
- Size: the number of occurrences of napes and tie points used to represent the
model by a plex structure;
- Density of connectivity: the ratio of the total occurrences of tie points to the total
number of napes. Density of connectivity increases as we add more connections
among napes;
578 E. Toppano
- Ontological variety: the total number of concept classes included in the conceptual
schema associated to a model;
- Concept density: for a concept type Ci, it is the ratio of the number of total
instances of Ci to the total number of instances of all concepts used in the model.
Lower counts of these metrics imply ‘simpler’. Metrics are used to adapt the
amount of information presented by a model to the learner in order to try to prevent
overloading and possibilities of distraction.
2.2 Enforcing Regularities of the Courses of Interaction
Three properties of the courses of interaction are currently monitored to perform

direct guidance namely cumulativeness, constructiveness and self-regulatedness. See
[5] for a formal definition of these properties. We are interested in applying these
properties to conceptualisations. Cumulativeness refers to the property that a course of
interaction exhibits when a conceptualisation (i.e. a conceptual schema) experienced
by the learner in a situation is in some way revisited in a later situation of the course
of interaction. Constructiveness refers to the property that a course of interaction
exhibits when a conceptualisation experienced by the learner in one situation is in
some way related to a new conceptualisation that the learner accesses in a later
situation. Self-regulatedness refers to the property that a course of interaction exhibits
when a learner's action performed in a situation is in some way evaluated by the
learner in another situation of the course of interaction and this evaluation is taken
into consideration to guide the next learner's actions or change the effects (e.g.
interpretation, understanding) of previous actions.
2.3 The Decision Making Approach
The above properties and metrics can be combined in various ways. For example, it is
possible to enforce cumulativeness with respect to a conceptualisation in order to
keep the course of interaction on the same conceptual basis of the models visited up
to a certain time and further minimise the amount of information to be presented to
the learner by decreasing the size or the density of connectivity of the proposed model
with respect to the currently focused one. Suppose, for example, that the learner is
visiting a page representing an artifact X from a structural perspective. The system
may suggest to visit a linked page describing another artifact Y that is represented
under the same perspective - and thus using the same conceptual schema - but using a
lower number of instances of the concepts types included in the schema. In this way,
the student can gain knowledge transferability by seeing multiple manifestations of
the same conceptualisation. Alternatively, it is possible to enforce constructiveness.
The system may suggest, for example, to visit a page representing the same artifact X
under a different perspective (e.g. a behavioral model) together with a page
illustrating codesignation relations between the two models. Codesignation relations
allow to integrate the two perspectives - the structural and the behavioral one - by
showing which elements of a model are connected to which elements of the other
model. In this way the student can gain cognitive flexibility by being exposed to
multiple interpretations (perspectives) of the same device. These decisions are taken
Adaptation through Interaction Modelling in Educational Hypermedia 579
according to a set of rules that consider the learner's profile (e.g. show simplified
material to novice learners and more complex details to advanced learners) and the
properties and characteristics of the past student's interaction with the MMforTED
system as they are inferred from the server access log.
Given a set of requirements (e.g. enforce cumulativeness, minimise the amount of
information) the decision mechanism is able to analyse - by accessing metaknowledge
about educational material - the characteristics of the models accessible from the
current position in the information space and to order the available alternatives on the
base of their appropriateness with respect to the desired properties. Informally,
requirements induce a preference ordering on possible choices and this, in turn,
determines the next page to be suggested to the learner. The decision making
mechanism, which is a modification of the algorithm used in the TAMS system [8],
embodies a hill climbing search process in the space of available alternative choices.
3 Conclusions
The instructional system we present in the poster performs direct guidance by

selecting material that affords to the learner a pattern of interaction or a course of
interaction with desired properties. This approach is in line with constructivistic
learning. Adaptation benefits from hierarchical organisation of domain conceptual
knowledge that enables to select material by performing smooth transitions from
coarse (abstract) material to fine grained detailed material and vice versa. The
adaptive version of MMforTED is still at its prototypical stage and no formal user
assessment have yet been carried out.
References
1. Murray, T.: MetaLinks: Authoring and Affordances for Conceptual and Narrative Flow in
Adaptive Hyperbooks. JAIED, Vol. 13, (2002) to appear
2. De Bra, P., Calvi, L.: AHA: a Generic Adaptive Hypermedia System. In: Proc. of the 2nd
Workshop on Adaptive Hypertext and Hypermedia, Hypertext '98, Pittsburgh (1998),
http://www.contrib.andrew.cmu.edu/~plb/HT98_workshop/DeBra.html
3. Brusilovsky, P.: Adaptive and Intelligent Technologies for Web-based Education. In:
Rollingen, C. and Peylo, C. (eds.) Kunstliche Intelligens, Special Issue on Intelligent
Systems and Teleteaching, Vol. 4, (1999) 19-25
4. De Bra, P.: Design Issues in Adaptive Web-Site Development. In: Proc. of the 2nd
Workshop on Adaptive Systems and User Modeling on the WWW (1999),
http://wwwis.win.tue.nl/asum99/debra/debra.html
5. Akhras, F.N., Self, J.A.: System Intelligence in Constructivistic Learning. JAIED, Vol. 11,
(2000), 344-376
6. Feder,J.: Plex Languages. Inform. Sci., Vol. 3, (1971), 225-241
7. Toppano, E.: Using Graph Transformations to Support Multilevel Reasoning in
Engineering Design. Machine Graphics & Vision, Vol. 8 (3), (1999), 395-425
8. Toppano, E.: Rational Model Selection in Large Engineering Knowledge Bases. Applied
Artificial Intelligence, Vol. 10 (3), (1996), 191-224
The Hybrid Model
for Adaptive Educational Hypermedia
1 2 1 3
Mohamed Ramzy Zakaria , Adam Moore , Helen Ashman , Craig Stewart ,
1
and Tim Brailsford
1
School of Information Technology and Computer Science
University of Nottingham, Nottingham, NG8 1BB, UK
{mrz,helen.ashman,tim.brailsford}@nottingham.ac.uk
2
School of Civil Engineering
[email protected]
3
School of Life and Environmental Sciences
[email protected]
Abstract. Web-based distance learning is becoming increasingly prevalent as

the Internet permeates every aspect of our culture, and many educational con-
tent management systems are now in use on the web. However, learners’ ex-
periences of these systems are almost invariably static, with information being
delivered regardless of their background or knowledge. Due to variation be-
tween learners’, it is suggested that these web-based distance-learning systems
would benefit from the capability of adapting their content to meet individual
needs. To effectively implement this adaptation of educational material, we re-
quire a user model that supplies the system with information about the learners
using the system, such as their backgrounds, knowledge, interests and learning
styles. This paper focuses on presenting a user model that combines the advan-
tages of two techniques (overlay and stereotyping) in a way that provides the
system with the ability to deliver information that is fully informed by the re-
quirements of individual users.
1 Introduction
The brave new era of the information age has ramifications for all disciplines, at the
most fundamental of levels. From education, to commerce and music, the Internet
impinges on every field where data and knowledge are currency. Arising out of this
worldwide network of communications comes the globalisation of information - in
which hypermedia tools are at the forefront enabling direct user access to information
[1]. As the amount of information on the web continues its exponential increase, the
number of users with different goals and interests also expands. It therefore becomes
increasingly important that the information available be adapted to suit each user’s
individual knowledge and aspirations.

The Hybrid Model for Adaptive Educational Hypermedia 581
For example, in traditional web-based educational hypermedia systems, the con-

tents are generally static, in so far as once written, their contents cannot be changed
without external intervention. This provides a uniform learning experience to all
learners, regardless of their needs and requirements. One example of such a system is
WebCT [2]. This is a hypermedia educational system developed in 1995 at the Uni-
versity of British Columbia. It is an environment for authoring and delivering educa-
tional materials over the web. WebCT presents a static and inflexible pedagogic
experience, without any kind of adaptation at the user level. Hence a web application,
such as those delivered via WebCT, which are designed with a particular class of
users in mind, may not suit those even marginally different from the original target
audience [3].
In response to this clear need, adaptive hypermedia systems have been created
such as AHA [4] and CHEOPS[5]. They build a model of the goals, interests, prefer-
ence and knowledge of their users; so that they may present them with the informa-
tion they need in a timely and appropriate manner [1].
This paper describes a hybrid user model, which is cooperative [6], (i.e. it collabo-
rates with users in gathering information, as they are required to supply the system
with some personal information, e.g. their occupation and preferences). This model
also involves users in the user modelling process, as the contents of topics or courses
are adapted according to their knowledge level about the topic they study. The hybrid
model described below has the benefit that it should suit any adaptive educational
hypermedia system, and we suggest that it is likely to provide a powerful addition to
any technology-based learning programme.
2 The Hybrid Model
2.1 Architecture
The hybrid model combines the use of two major techniques that are prevalent within
the user modelling community. The first of these is the overlay model, perhaps cur-
rently the most widely used technique of user modelling. This is used to measure the
knowledge level of users in any given topic or domain. A user’s knowledge according
to this model is considered to be an overlay of the total knowledge representing that
domain. This knowledge level is represented in the form of “Concept-Value” pairs
[7,8].
The second model is the stereotype; this technique assumes that knowledge is
customised for specific groups, with each user being assigned to one and only one
group at any given time. Thus, users who share the same background or knowledge
should be assigned to the same group. Users cannot change from one group (or class)
to another until they trigger the specific conditions of the new group [9,10].
Aspects of each of these models are utilised by the hybrid model as follows:
1. Overlay technique: the overlay measures the knowledge level of each learner
within certain subject domains. This knowledge level might represent the score
achieved in the system assessment at the end of each lesson, although any other pa-
582 M.R. Zakaria et al.
rameters the system authors may choose may also be used. For example, the score
achieved in self-assessment quizzes is a widely used and well-accepted metric of the
comprehension of information.
2. Level stereotype: level stereotypes mainly depend on the knowledge level of users.
For example, they may simply be defined as Beginner, Intermediate and Advanced,
but any classes may be used as appropriate to each system. According to the users
knowledge level they will be assigned to a single class of the level stereotype within
any given domain they of study. For example, a user studying biomechanics might
be assigned simultaneously to the novice class in biology and to the advanced class
in mathematics. Classes in the level stereotype are concerned with providing assis-
tance that is appropriate, and adapting the contents of the lesson to suit the learner.
Each class may define an article or set of articles, links to external documents, or to
lessons in other courses. For example, if a user belongs to one of the advanced
classes he may be provided with advanced articles or links to help the user to find
more about the topic or domain he studies. Level stereotypes not only adapt the
contents to suite a user’s level, but they also facilitate learning by identifying do-
main weaknesses in the topic under study.
3. Category stereotype: the hybrid model has been designed for systems that simul-
taneously run multiple courses for different levels of users. For example, the system
may be running courses for first year undergraduate as well as postgraduate users.
Overlay Model
M
at
ch
in hi
g s
us kn
er ow
Categories
’s
C1 C2 ……. Cn
ca led
Stereotype
te ge
go le
ry ve
an l
d
up
Levels
da
BE I A --
tin
g
Stereotype
Concept
End
User
Information pool B A
I
Fig. 1. The components of the hybrid model. An overlay model combines with a category and
level stereotype to retrieve appropriate content from the information pool to convey concepts
to the user. The user interacts with the system to inform and update the user model. BE – Be-
ginner, B – Basic, I – Intermediate, A – Advanced
For this reason the users need to be categorised, as the knowledge level of under-
graduate users in a certain stereotype level of a certain domain may not be the same
as that of postgraduate users in the same stereotype level of the same domain. For
example, consider two users, one of them a first year undergraduate and the second
one a studying for a higher degree. Both of these students are classified in the in-
termediate level stereotype for the biology domain. Both of them are in the same
level stereotype, but the intermediate level of postgraduates will be much more ad-
vanced than that of first year undergraduates. The categories stereotype helps the
system to distinguish between different users in the same level stereotype, and to
provide each of them the appropriate adaptation and help.
There is one other important aspect of this hybrid model - the information pool. This
is categorised by the domain model, and consists of a pool of articles, links, and other
items that encapsulate the resources of an adaptive system. The information pool is
likely to differ in both form and content from one system to another.
The overlay technique, level stereotype and category stereotype combine to pick
from the information pool the most convenient articles and links that suit each user’s
level, knowledge and background. Thus, according to a user’s knowledge level and
category, the most appropriate materials will be chosen from the information pool.
Figure 1 shows the hybrid model’s components, comprising of two stereotypes and
one overlay, to provide the maximum flexibility and to have the capability to serve a
wide range of users.
2.2 Mechanism
When a user logs on to the system for the first time, he/she will be given an initial
knowledge level value according to an estimate of prior knowledge about the subject
under study. That recommends the user to a certain level stereotype, and the category
is determined according to any parameters that the system authors may choose such
as user's occupation. Each time a user passes from one lesson to another the knowl-
edge level for that user is updated according to the score in the system assessment (as
well as other parameters the system’s authors may choose). According to the user’s
new knowledge level, the class assigned according to the level stereotype might be
changed or might be the same (i.e. if the user still has the same knowledge level). The
adaptation of the contents and the supporting articles are available according to the
class of the user in the level stereotype as well as the category. The steps involved in
adaptation are illustrated in Fig. 2.
3 Implementation of the Hybrid Model

in a Hypermedia Learning Environment
The hybrid model is currently being integrated into an adaptive educational hyperme-
dia system called WHURLE (Web-based Hierarchal Universal Reactive Learning
584 M.R. Zakaria et al.
Environment), [11,12,13] an XML-based integrated learning environment. In

WHURLE the content consists of atomic chunks, each of which consists of the small-
est conceptually self-contained unit of information that the author can envisage. The
domain of WHURLE content is contained within the melange, which consists of all
U s e r lo g in
U s e r c a t e g o r y s t e r e o t y p e a n d le v e l s t e r e o t y p e a r e
c h e c k e d th e n a d a p ta ti o n ta k e s p la c e
U s e r k n o w l e d g e le v e l i s u p d a t e d a t th e e n d o f e a c h le s s o n a c c o rd in g t o
h is s c o re i n s y s t e m a s s e s s m e n t in a d d itio n to a n y o th e r p a ra m e t e r s
U p d a tin g u s e r s t e r e o t y p e s
a c c o r d in g t o h is n e w k n o w le d g e
le v e l
Fig. 2. The hybrid model mechanism to adapt materials to users. Once the user logs in, the
system either adapts its material according to the existing user model, or creates a new one. At
the end of each lesson, the user model is updated.
the available chunks in any single installation. In WHURLE conditional transclusion

[14] is used to construct virtual documents, which contain one or more of these
chunks. The hybrid model is integrated into WHURLE as the filter for the lesson
plan. This filter generates an adapted virtual document dependant upon both the adap-
tation rules within the lesson plan and user profile. WHURLE lesson plans represent
the information pool of the hybrid model [11, 12, 13].
4 Conclusion
The hybrid model is a user model that gathers together the most commonly used
techniques of user-modelling for adaptive hypertext. This utilises the advantages of
each of these techniques in a way to provide a full understanding for the user’s needs
and requirements on several different levels. Using this technique we have imple-
mented adaptation within WHURLE to allow students to see pages of information
containing one or more chunks in a manner relevant to their skills, knowledge and
learning styles, following rules set by the author of the lesson that they are viewing.
Using this model we hope that the WHURLE system will provide a strong pedagogic
framework for a variety of web-based learning scenarios.
Acknowledgements
We wish to thank Peter Murray-Rust, Peter Davies, and Ban Seng Choo for many
useful discussions, and colleagues in the WTG, VSB and IBiS for their support and
encouragement. Craig Stewart is a research associate funded by the Hong Kong Uni-
versity Grants Committee.
References
1. Brusilovsky, P. (1996). “Methods and techniques of adaptive hypermedia”. User model-

ing and user-adapted interaction, 6(2-3), pp. 87-129.
2. Beshears, Fred. “WebCT overview”.
http://socrates.berkeley.edu:7521/articles/webct/WebCT-Presentation
3. Eklund,J; Brusilovsky,P; Schwarz,E.(1997). “Adaptive Textbooks on the World
WideWeb”, Proceedings of AUSWEB97, the third Australian conference on the world
wide web, Queensland, Australia, July 5-9, 1997, Southern Cross University press, pp.
186-192 .http://ausweb.scu.edu.au/proceedings/eklund/paper.html.
4. De Bra,P;Calvi,L.(1998).”AHA Adaptive Hypermedia Architecture”.NRHM journal, V.4,
pp 115 - 139
5. Ferrandino,S; Negro,A;Scarano,V.(1997).”CHEOPS :Adaptive Hypermedia on the World
Wide Web”. Proceedings of the European Workshop on Interactive Distributed Multime-
dia Systems and Telecommunicazion Services (IDMS '97).
6. Kay, J. 1995,“The UM toolkit for cooperative user models”. User Models and User
Adapted Interaction 4(3), 149-196.
7. Valley,K. (1997). “Learning Styles and Courseware Design”. Association of Learning
Technology Journal, 5(2), p42-51
8. Carr,B; Goldstein, I. (1977). “Overlays, a theory of modelling for computer aided instruc-
tion”. A technical report. AI memo 406, MIT, Cambridge, MA.
9. Rich, E. (1983). “Users are individuals: individualizing user models”, Journal of man-
machine studies vol.18, 199-214.
10. Benaki,E; Karkalestsis,V; Spyropoulos,C. (1997). “User modelling in WWW: the UMIE
prototype”. Proceedings of sixth international conference on user modelling, Chia La-
guna, Sardinia, 2-5 June1997
http://www.contrib.andrew.cmu.edu/~plb/UM97_workshop/Benaki/Benaki.html
11. Brailsford, TJ ; Moore, A ; Stewart, CD ; Zakaria, MR ; Choo, BS ; Davies, PMC.(2001).
“Towards a framework for effective web-based distributed learning”. WWW10 proceed-
ings, HongKong.
12. Moore,A; Brailsford,T.J, Stewart, C.D. (2001). “Personally tailored teaching in
WHURLE using conditional transclusion”. The twelfth ACM conference on hypertext and
hypermedia. August 14-13, 2001, Denmark.
13. Brailsford,T; Stewart,C; Zakaria,M; Moore,A.(2002).”Autonavigation, Links and Narra-
tive in an Adaptive Web-Based Integrated Learning Environment”. Proceedings of
www2002, Hawaii, USA.
14. Nelson, T.H. (1995). “The Heart of Connection: hypermedia unified by transclusion”.
Communications of the ACM.
Navigation Modelling in Adaptive Hypermedia
Peter Dolog and Mária Bieliková
Department of Computer Science and Engineering

Slovak University of Technology
Ilkovičova 3, 812 19 Bratislava, Slovakia,
{dolog, bielikova}@dcs.elf.stuba.sk,
http://www.dcs.elf.stuba.sk/{dologp,bielik}
Abstract. In this paper we reflect the need for modelling in a system-

atic production of adaptive hypermedia applications. Proposed approach
is based on the Unified Modelling Language (UML). State diagrams are
used to model possible paths through hypertext. The user model ex-
pressed by a class diagram determines structural and behavioural fea-
tures, which are used for specification of adaptations in states and tran-
sitions contained in state diagrams.
1 Introduction
One of the main goals of any adaptive hypermedia application is to increase
user efficiency measured either in the time spent searching for information or
the amount of information absorbed by the user. Another important issue is to
aid developers of such systems which are going to be more and more complex.
The increased complexity of hypermedia applications raises the need to employ
modelling in hypermedia development process.
The modelling of a hypermedia is extensively studied only in past two decades.
Models help us understand developed system by simplifying some of details.
Adaptation of navigation, together with a user model, should be addressed
in hypermedia application modelling. The goal of this paper is to present an
approach to modelling adaptive navigation. The Unified Modelling Language,
namely state diagrams together with sequence and class diagrams are employed
for these purposes.
2 Process of Adaptive Navigation Modelling

A navigation model represents possible paths through information chunks and
their contextual grouping. To develop such navigation model, an analytical model
of information structure and user roles should exist.
Basic modelling technique for navigation modelling in our approach is a state
diagram, which enables to model dynamic character of navigation. We proposed

This work was partially supported by Slovak Science Grant Agency, grant No.
G1/7611/20.

Navigation Modelling in Adaptive Hypermedia 587
five basic steps of navigation modelling process: identifying basic interaction

scheme, identifying states, identifying transitions, identifying events, and map-
ping the user model elements to the state diagram. These steps can be performed
in parallel and in iterations. Moreover, proposed approach can be used at several
levels of abstraction (of a hypermedia system).
Adaptive navigation strongly depends on a user modelling. The user model
incorporates various characteristics of users. Hypermedia application usage data
are also represented in the user model. In our approach a user is modelled by a
class diagram similarly to [12]. The user model is derived from user roles. Struc-
ture of the user model follows the well known Adaptive Hypermedia Application
Model (AHAM) [6]. The user model should at least contain a class, which repre-
sents the level of user knowledge. Other classes (user preferences, goals, interests,
knowledge, background, hyperspace experience, etc.) can be incorporated when
it is needed [9]. The user model contains operations for reading the current state
of user characteristics and for updating user characteristics. Environment or con-
text data [2] are carried during mapping to navigation model likewise data in
the user model.
I. Identifying basic interaction scheme. The first step in navigation modelling

is basic interaction scheme modelling. This is intended to identify a sequence of
interactions between main system roles. The UML sequence diagram is used for
these purposes.
II. Identifying states. States in a navigation model fulfil the role of information
chunks [8]. They can be grouped into superstates. The states are created from
an information model. There are two possibilities of mapping: (1) a superstate
mapped to a class with substates mapped to class attributes, and (2) a superstate
mapped to a class instance with substates mapped to class instance attributes.
Parallel substates are mapped to attributes of a class or its instances, which
are presented simultaneously. Attributes, which do not need to be presented
simultaneously are grouped into ordinary substates. The classes, which are ag-
gregates of another class are mapped to parallel or ordinary substates of that
class’ state. In addition, these substates are determined by the cardinality of the
aggregation relationship. Specialised classes are mapped to ordinary substates.
Special information chunks derived from several attributes and/or classes or
special states needed for purposes of navigation can also be considered. States
can be extended with a history. The history indicates that a user can start his
browsing where he finished when exited system last time.
III. Identifying transitions. A transition represents an active interconnection be-

tween information chunks. Association relationships from information model are
transformed to transitions. When it is needed, additional transitions can be in-
corporated into the model. The fork and join pseudostates, and SyncState are
intended to model a synchronisation of parallel states. The first two are intended
for splitting or joining transitions. The latter is for synchronizing substates of
parallel regions.
588 P. Dolog and M. Bieliková
A condition can be assigned to the transition. A transition can also be con-

ditionally forked; i.e. the transition can end in several states. Transition can also
have associated time event for modelling sequential hypermedia timing. A tran-
sition can also have associated side effect actions, which together with transition
conditions are very important for adaptation modelling.
IV. Identifying events. Events raise transitions in a state machine. Events can
be directly mapped to presentation elements, which have associated actions.
They are mediators between navigation model and presentation model of actions.
Events can be joined to a generalisation/specialisation tree. An event can be
mapped to more than one transition.
V. Mapping user model elements to state diagram. The adaptive behaviour is

modelled by an introduction of features of user model classes into state diagrams.
Accessible attributes of user model classes are mapped to guards conditions.
They are tested for specific values, which have to be satisfied when transition is
raised. Operations are mapped to actions of transitions. They are used for up-
grading the user model state or for specific operations with the user model and/or
information chunks. Operations for retrieving current user model state can also
be used in guard conditions. Guard conditions of transition specify local rules
of adaptation. Global rules of adaptation, can be specified as guards of internal
transitions, parts of entry, exit or do actions, and conditions of superstates.
3 Modelling of Techniques for Adaptive Hypermedia
Several efficient techniques for adaptive hypermedia were proposed [3]. We se-
lected some of these techniques for presentation of capabilities of the proposed
approach. The examples cover both link-level and content-level adaptation.
Figure 1 depicts part of an adaptive navigation model, which was created ac-
cording to the approach proposed in this paper. The example figures the model
of a lecture on functional programming (FP). This lecture consists of four topics:
Functional Programming, Programming Schemes, Examples of Linear Lists
Processing and Examples of Non-Linear Lists Processing. The aim of
such lecture is to exercise programming of basic list processing functions in the
Common Lisp language. First, some introduction is needed. This is carried out by
Introduction to FP, Computation in FP, and Introduction to LISP frag-
ments. Next, the introduction to Programming schemes (Programming Schemes
state) is performed. It is represented by Introduction substate and simple cat-
egories of Linear Lists Processing and Non-Linear Lists Processing.
Adaptation rules are involved in transition labels or as internal transitions
of states. Events handle user interaction or internal system events. Conditions
and actions are taken from the user model.
Conditional text is modelled by Entry internal transition of a state. It is
followed by a condition, which determines whether the fragment is displayed
or not. In Fig. 1, conditional text is represented by the Introduction to LISP
Functional Programming Programming Schemes

Next
Introduction to FP Introduction
Computation in FP Linear Lists Processing

Entry/SortOutgoingLinks()
Click
Introduction to LISP Simple Commented
Entry[User.UserKnowledge. Click
CurrentLOK() > Low ]
[User.UserKnowledge. [User.UserKnowledge.
CurrentLOK() > Low ] CurrentLOK() <=
Low ]
Examples of Linear Lists
Processing
Non-Linear Lists Processing
Searching
Searching Reduction
Reduction Next
Reduction List Generation
Reduction By By [User.UserKnowledge.
Function CurrentLOK() > Middle ] [User.UserKnowledge.
Function
CurrentLOK() >
Mapping Middle ]
Mapping
[User.UserKnowledge.
List CurrentLOK() < Examples of Non-Linear
Generation Middle ] Lists Processing
Fig. 1. A part of an adaptive navigation model.
state. There is the condition, which examines the level of current user knowledge.
The fragment is displayed only if the level of knowledge is greater than Low.
Stretchtext can be modelled by two approaches. The first is to represent
stretched and unstretched text as two different alternative states with transition
between them. The transition has associated Click event. The second approach
is to model stretchtext with two parallel states, where one state is condition-
ally constrained. It means that the state, which is conditionally constrained
is presented only if the condition is satisfied. Condition is mostly based on a
history (usage data). In Fig. 1, stretchtext is represented in the Linear List
Processing state containing Simple and Commented version of introduction.
The text is displayed unstretched if current user level of knowledge is greater
than Low. Otherwise, Commented version is displayed. These two alternatives can
be switched by clicking (Click event).
Commented text is modelled similarly. The difference between commented
text and stretchtext is that stretchtext can be clicked and thus unstretched.
Alternative pages or fragments are modelled by conditional transition split.
The decision symbol (junction) represents modelling element for alternation. An-
other possibility is to employ the diamond symbol. In Fig. 1, there are two alter-
natives when the next event occurs. If current level of user knowledge is less than
Middle only the Examples of Linear Lists Processing state is displayed.
590 P. Dolog and M. Bieliková
Otherwise, Examples of Non-Linear Lists Processing as more challenging

task is provided by adaptive hypermedia application.
Direct guidance can be modelled by assigning the next or back event to
particular transition. Another approach is to use directly state diagram as a
guide, which is interpreted and displayed as a map and the current position and
allowed links are sufficiently indicated (for example like marking in petri nets).
Hiding of a link is modelled by guard condition of a transition. When the
condition is not satisfied, link is not enabled (not displayed). The transition in
Fig. 1 labelled by List Generation event is an example of such a link. The
link is displayed (transition is allowed) only if current user level of knowledge is
greater than Middle.
Link sorting is an operation. It is obviously performed when a user reached
particular fragment, where the links have to be sorted. Such operation can be
modelled as the Entry action of particular state. Figure 1 depicts Linear List
Processing state, which represents such fragment. The SortOutgoingLinks()
action is invoked when the user reached this state.
The state/transition diagram can be used also as a model of adaptive map.
Special manipulation function can be provided, which interprets meta-model and
rules of states, and according to them displays particular part of the model to
the user.
4 Related Work and Conclusions
In this paper we described an approach to adaptive hypermedia modelling. Ap-

plication of a state/transition diagram modelling technique is the first advantage
of the proposed approach. A hypermedia application reacts to user actions. Thus,
it seems that state-transition diagrams are more natural for modelling navigation
than structural techniques such as in OOHDM [14], UHDM [12] or WebML [4]
(for a review of others, see for example [7]).
Provided guidelines for user modelling and the integration of a user model
and navigation model is another advantage. User modelling is supported by
UHDM [12], WebML [4] and W2000 [1] but independently from navigation.
According to [10] the adaptation specification can be involved as slice or slice
relationship condition. But slices and their relationships do not satisfactorily deal
with interactions. The authors do not explicitly discuss relationships between
user model and presentation or application model. In [13] the graph formalism
is employed for modelling paths. Adaptation is specified as text composition
templates with linguistic rules.
HMBS/M [5] and χTrellis [11] are based on behavioural techniques. However,
the former only allows to map states to class instances. Both approaches do not
emphasize on adaptation and user modelling.
Our approach can be used for modelling known techniques of adaptive navi-
gation and presentation. Our further research is oriented to extension of proposed
approach to support implementation modelling.
References
1. Luciano Baresi, Franca Garzotto, and Paolo Paolini. Extending UML for modeling
web applications. In Proc. of 34th Anual Hawaii International Conference on
System Sciences (HICSS’34), Maui, Hawai, January 2001. IEEE Press.
2. Mária Bieliková. Adaptive presentation of evolving information using XML. In
T. Okamoto, R. Hartley, Kinshuk, and J.P. Klus, editors, Proc. of IEEE Interna-
tional Conference of Advanced Learning Technologies (ICALT’2001), pages 193–
196, Madison, USA, August 2001. IEEE Press.
3. Peter Brusilovsky. Methods and techniques of adaptive hypermedia. User Modeling
and User-Adapted Interaction, 6(2-3):87–129, 1996.
4. Stefano Ceri, Piero Fraternali, and Aldo Bongio. Web Modeling Language
(WebML): a modeling language for designing web sites. Computer Networks and
ISDN Systems, 33(1–6):137–157, June 2000.
5. Marcia Regina de Carvalho, Maria Cristina Ferreira de Oliveira, and Paulo Cesar
Masiero. HMBS/M - an object oriented method for hypermedia design. In Proc.
of Brazilian Symposium on Multimedia and Hypermedia Systems (SBMIDIA’99),
pages 43–62, Goiânia, June 1999.
6. Paul De Bra, Geert-Jan Houben, and Hongjing Wu. AHAM: A dexter-based ref-
erence model for adaptive hypermedia. In K. Tochtermann, J. Westbomke, U.K.
Wiil, and J. Leggett, editors, Proc. of ACM Conference on Hypertext and Hyper-
media, pages 147–156, Darmstadt, Germany, February 1999.
7. Peter Dolog. Modelling in hypermedia development, August 2001. Technical Re-
port (A Written Part of PhD Examination). Department of Computer Science and
Engineering, Slovak University of Technology.
8. Peter Dolog and Mária Bieliková. Modelling browsing semantics in hypertexts
using UML. In J. Zendulka, editor, Proc. of ISM’2001 - Information Systems
Modelling, pages 181–188, Hradec nad Moravicı́, Czech Republic, May 2001.
9. Peter Dolog and Mária Bieliková. Hypermedia modelling using UML. In Proc.
of ISM’2002 - Information Systems Modelling, Rožnov pod Radhoštěm, Czech
Republic, April 2002.
10. Flavius Frasincar, Geert Jan Houben, and Richard Vdovjak. A RMM-based
methodology for hypermedia presentation design. In A. Caplinskas and J. Eder,
editors, Proc. of ADBIS 2001 - Advances in Databases and Information Systems,
pages 323–337, Vilnius, Lithuania, September 2001. Springer, LNCS 2151.
11. Richard Furuta and P. David Stotts. A formally-defined hypertextual basis for
integrating task and information, 1994. Tech. Report TAMU-HRL 94-007.
12. Nora Koch. Software engineering for adaptive hypermedia systems? In Paul De
Bra, editor, Proc. of Third Workshop on Adaptive Hypertext and Hypermedia, 8th
International Conference on User Modeling, July 2001.
13. Daniela Petrelli, Daniele Baggio, and Giovanni Pezzulo. Adaptive hypertext design
environments: Putting principles into practise. In Proc. of International Conference
on Adaptive Hypermedia and Adaptive Web-Based Systems (AH’2000), pages 202–
213, Trento, Italy, August 2000. Springer, LNCS 1892.
14. Daniel Schwabe and Gustavo Rossi. An object-oriented approach to web-based
application design. Theory and Practise of Object Systems (TAPOS), Special Issue
on the Internet, 4(4):207–225, October 1998.
An Adaptive Agent Model
for e-Commerce Architecture
Ana B. Gil1, Francisco García1, and Zahia Guessoum2
1 University of Salamanca, Dept. Informática y Automática

Facultad de Ciencias, Plaza de la Merced s/n, 37008, Spain
{abg, fgarcia}@usal.es
2 LIP6, UPMC 4 Place Jussieu, Case 169, 75252 Paris Cedex 05
[email protected]
Abstract. For building easily e-commerce applications, adaptive agent models

are useful. These systems therefore should be equipped with the ability to
improve their performances themselves, working over the basis of building a
model with the goals, preferences and knowledge of each individual user in the
environment. This paper presents a new architecture to support sales in Web
stores centered on business based on product catalogs. The main goal of this
architecture is focused on endowing capability of adaptivity in several ways:
contents, navigation and presentation, while analyzing the different ways to
develop the adaptivity in e-commerce.
Keywords: Adaptive Agents, E-commerce, Negotiation, User Model.
1 Introduction
The Internet, a network of computer networks that have no central control or

organization, is changing the way people think about and do business. The basis of
competition and wealth creation in the e-commerce is not only a good use of
information, quality, re-engineering process, speed, effective decision-making,
empowerment or other countless management techniques popular today. It is also the
flexibility and adaptability when working on business. Lots of customers with
different expertise and interest access to the online marketplaces, therefore they may
need support to explore the catalogs and to select the products for purchase making
the best choice.
In recent years, researchers from different fields have pushed toward greater
flexibility and intelligent adaptation in their systems. The development of intelligent
adaptive agents has been rapidly evolving in many fields of science. Such agents
should have the capability of dynamically adapting their parameters; improve their
knowledge base or operation methods in order to accomplish a set of tasks. These
agents should therefore be equipped with the ability to self-improve their
performance, applying the principles and concepts of this kind of learning in real-
world contexts.

An Adaptive Agent Model for e-Commerce Architecture 593
The purpose of this paper is to present a new architecture to support sales in Web
stores centered on business based on e-catalogs. The paper is organized as follows:
Section 2 explains the proposed agent-based e-commerce architecture and examines
the components of this architecture that are actually implemented and some work
plans for be added. Finally, Section 3 closes the paper, presenting our conclusions and
some ideas on further work needed.
2 An Adaptive Agent Model for e-Commerce
Online marketplaces are an opportunity and a threat to retail merchants. They offer
traditional merchants an additional channel to advertise and sell products to
consumers thus potentially increasing sales. They provide the hope and the promise of
the global perfect market. A commerce model typically includes activities such
as advertising, searching, negotiating, ordering, delivering, paying, using, and
servicing [2].
EC is one of the most important applications for Agent Technologies [1]. Agents
are not limited to simple information gathering tasks, but they are increasingly
becoming involved in the more complex process of actual trading: making purchases
and deals. Therefore Multi-Agent System (MAS) have to be endowed with adaptive
behavior because the environment in this kind of scope is complex and dynamic.
2.1 Adaptive Agent Model
An adaptive agent can be defined as an agent with the ability to strengthen rules,
which “work” best, and the ability to discover new plausible and possibly “better”
rules applying several techniques. To build a generic adaptive agent model and
following the tradition of explicating and separating representation of control in meta-
level architecture, Guessoum [4] proposes a meta-behavior in the agent architecture.
This meta-behavior gives each agent the ability to make appropriate decisions about
control or to adapt its behaviors over time to new circumstances. It provides the agent
with a self-control mechanism to dynamically schedule its behaviors in accordance
with its internal state and its world state.
A standard adaptive agent model has two main components: (i) a behavior and (ii)
a meta-behavior. Behavior and meta-behavior are both based on two kinds of
elements: actions and conditions, and they have both a decisional system. Actions,
conditions and decisional system are the minimal structure for each kind of behavior.
The meta-behavior allows to dynamically updating the set of rules describing the
agent behaviors relies on data about the agent itself, its environment, and the decision
system used by the behavior and the way to modify it too. So the meta-behavior
provides two kinds of adaptation, (i) structural adaptation, adapting the structure of
the agent to the evolution of his environment and (ii) behavioral adaptation, adapting
the decision-making process of the agent to the evolution of its environment.
The main feature provided is that these two kinds of adaptation can be combined to
have at same time agents that adapt their structure and their behavior.
594 A.B. Gil, F. García, and Z. Guessoum
2.2 An Adaptive Approach for e-CoUSAL
Our group is interested in the definition of e-commerce models that allow the entry of
the Small and Medium Enterprises (SMEs) into the virtual commerce bandwagon.
The model here presented it is based on an architecture for product catalog-based e-
commerce architecture, which we call e-CoUSAL [3]. This is an architecture based on
two main components: a visual catalog-designer tool and an e-commerce web server.
shopping
Supplier organiz ation Final client

commerce
management
e-catalog payment system
design
E-c ommerce
server
publication
Visual catal og-designer tool

Bank
Fig. 1. General structure of the proposed architecture
The first one, the authoring tool is used by the enterprise to generate and manage
the e-catalog that presents its business in the e-commerce site. The e-catalog will be
published in the e-commerce web server automatically. Then, the e-commerce site
needs to adapt itself to show the correct interfaces and the actual information to its
end-users through its e-commerce services.
The e-commerce site supports a catalog-based policy; this means that this kind of
system diverges quite far from classic on-line information systems. While a
hyperspace of information items still constitutes a major part of these systems,
browsing this hyperspace is not a major activity, but only a by-product of the major
activity (such as shopping for goods). In fact, the better these systems work, the less
browsing should be required. Adaptive characteristics are particularly interesting here.
In this section we introduce the different components that comprise the proposed
architecture and propose the future work to make it adaptive.
2.2.1 The e-CoUSAL e-Commerce Architecture

e-CoUSAL is an e-commerce architecture especially directed to SMEs, where the
amount of investment in technology solutions cannot be very large. Therefore two
main objectives are considered to develop this model: flexibility in the business
environment and that the business property should be able to maintain its electronic
business in self-sufficient way.
As we can see in Fig. 1, the central element is the e-commerce server, which
interconnects the different parts involved in the commercial environment. The SME is
responsible for the inclusion, across a specialized software tool (as described
explained in next section), of the e-catalog that introduces the offers into dynamically
generated web pages in the server where a client could find and buy the commodity.
2.2.2 The Visual Catalog-Designer Tool

The practical use of this tool is subordinated to being part and link vehicle of an e-
commerce oriented architecture while introducing a general platform that makes
possible an e-catalog based business strategy over the Internet. In this way,
enterprises can create their own e-catalogs using this tool, and then they can offer
their goods for sale through the e-commerce architecture.
A symbiotic relationship arises from this tool and the e-commerce site where the e-
catalogs are published. The server imposes a minimum set of criteria for product and
catalog. On the other hand, the tool offers the server all the necessary information for
the publication of the organization contents. This information exchanged is the logical
definition of the products that will be published, their conceptual organization in an e-
catalog, and the precise visualization format. For this information interchange the
markup language XML is used due to its flexibility for modeling, its document
semantic validation capability, and the separation between the data and their
visualization format.
This tool has customizing capabilities in its interface. The most important one is its
language independence through locate messages defined in external XML files so it
can be extended easily to any language.
Fig.2. Catalog layout Fig.3. Catalog sheet definition
The working process with this tool is structured around the work-view concept,
which reduces the amount of information that is shown to the user of the tool, thus
avoiding an unnecessary information overload. The main work-views defined in the
presented visual design tool are as follows: the template definition view, the product
view, and the catalog manager view.
The template definition view consists of the data template and the product
template. The data template could be defined as the mechanism that allows defining
the data format of one field for product description then helping the user to define the
596 A.B. Gil, F. García, and Z. Guessoum
contents in a more effective way. On the other hand, a product template serves as a
model for the later definition of precise products into e-catalogs.
All the functionality for precise products is gathered in the product work-view.
While templates make our work easier, the products are the conceptual definition of
each element that appears in the e-catalog.
The catalog manager view presents the grouping of the products in e-catalogs for
their later publication and maintenance in the server. When an e-catalog is defined
there is a separation between its conceptual definition and its visualization in the
server. This characteristic directs the e-catalog composition process as can be seen in
Figures 2 and 3.
Another non-functional view is introduced to guarantee the consistency among the
other work-views. This is the repository view.
2.2.3 Application of Adaptive Agents to e-CoSUSAL

The adaptive agent model explained in previous section is our ground basis to be
implemented in the actual e-CoUSAL architecture prototype, where specialized
adaptive agents have to be designed to carry on the activities of the front-end of the
Web store. For this purpose, we aim to implement with the multiagent platform
DIMA tool, founded on open and modular agent architecture.
Meta-Knowledge
e-commerce
Explicit site
Explicit
input output
Final Client
Fig. 4. e-commerce adaptive agent model
The requirements for the adaptation system architecture can occur at three levels:
presentation, information and navigation. These three faces are combined and
connected to conform the proposal architecture.
The user in order to the level of expertise and interest needs an agent that assists
him with the navigating the web site (Navigation). The main goal is to endow
business-object to the architecture, therefore negotiation processes reinforced with
adaptive user interfaces (Presentation) for making offers in way of discount
stockbrokers (Information) to the final client will be another task.
Therefore, it’s necessary monitoring the interactions on the client-side taking into
account the system customer is working on (kind of medium, downloading time,
content attributes) also while agents analyzing web logs, determining similarity
between documents and searching for making recommendations and learning about
users. At this point there are twofold ways of learning, (i) about a concrete user, and
(ii) to learn from patterns of use. For modeling this information and combined all the
users aspects, dynamic user modeling techniques are necessary.
The adaptive model we propose for e-commerce (see Fig. 4), interacts with the
final client due to negotiation techniques helped with the e-commerce site client
interface, adapting in according to the inputs. So this structure allows agents to

register the users interests and personalized pages to offer relevant information to the
client. Negotiation will locate the mutually preferred deals. User’s preferences are
expressed to the agent that performs negotiation. On this way a flexible virtual
business adapts its strategy and the site where is hosted.
This paper has shown a brief look into Adaptive Systems and how have a real
importance in e-commerce because of adaptive models are more suitable for building
solutions for real-life applications, therefore this framework is a great promise for the
further advancement in this ground.
We have presented an adaptive multiagent model for e-commerce that could serve
as starting point for creating an e-commerce site. We plan to implement the presented
model with the profiled adaptive architecture with DIMA that offers to build the
adaptive structure and experiment with several adaptation mechanisms. The
adaptation is mainly across the negotiation agent that will allow the SME personalize
their business policies (discount, payments and so on) and also the customization one
to adapt the interaction to the both principal actors, the end-user and also the SME-
user.
References
1. Gil, A. B., García, F. J., Guessoum, Z. “Adaptive Agents for E-commerce Applications”.
In New Methods and Tools Supporting E-Commerce – Proceedings of the ZOCO
Meeting. Catedral, Salamanca. 2002.
2. J. Eriksson, N. Finne and S. Jason. SICS MarketSpace, an Agent-Based Market
Infrastructure. Agent Mediated Electronic Commerce. Noriega, P., Sierra, C.(Ed.). Lecture
Notes in Computer Science. VOL. 1571. Springer-Verlag, pp 41-53. 1999.
3. J. García, Mª N. Moreno and J. A. Hernández. E-CoUSAL: An e-commerce architecture
for small and medium enterprises. In Advances in Business Solutions. Ed. Catedral. (In
press). 2002.
4. Z. Guessoum, M. Quenault and R. Durand. An Adaptive Agent Model. Proc. AIB’S,
York, 2001.
Personalized Adaptation to Device Characteristics
Eelco Herder and Betsy van Dijk

University of Twente
{e.herder, e.m.a.g.vandijk}@cs.utwente.nl
Abstract. Device characteristics, such as screen size and means of interaction,

and the context in which a device is used, seriously affect the user’s mental rep-
resentation of an information environment and its intended use. We hypothe-
size that user characteristics are valuable resources for determining which in-
formation is of interest in specific situations. Our project goal is to design
mechanisms for adapting navigation support to device characteristics and its
context of use, thereby considering that user goals and the resulting expected
navigation behavior might be subject to change.
1 Introduction
Mobile access to information is becoming more and more important. A handheld

device equipped with a browser and a wireless connection provides an opportunity to
connect to the Internet at any time from anywhere [2]. As we know, handheld devices
are smaller than desktop devices, in order to ‘fit into your pocket’ more easily. This
has its consequences for screen size and interaction mechanisms. Although screen
resolution and input devices can be improved, interaction mechanisms for handheld
devices will remain limited because of their size. Advanced adaptive presentation and
adaptive navigation support techniques are said to be indispensable for making a
small-screen interface more usable [1].
Hypertext systems have the ability to produce complex, richly interconnected and
cross-referenced bodies of information. Because of their non-linearity, users fre-
quently experience cognitive overload and disorientation [11]. Various navigation
aids have been developed in order to provide the context needed to prevent this prob-
lem. Adaptive hypermedia systems build a model of the goals, preferences and
knowledge of each individual user, and use this model throughout the interaction in
order to adapt to the user’s needs [1]. It has been demonstrated that navigation sug-
gestions based on user goals as well as on navigation behavior of previous users have
a positive effect on users’ spatial and temporal awareness [6]. Sorting, hiding and
annotating links as well as map adaptation are some example methods that may help
users to find their paths in hyperspace.
It is expected that mobile Internet users primarily want to extract particular bits of
information relevant to a current task. Navigation therefore requires strong support

Personalized Adaptation to Device Characteristics 599
for this goal-directed activity [2]. Several systems have been developed that summa-
rize ‘regular’ web pages in order to display them conveniently on a small-screen
handheld device. Some of them (e.g. [2]) also provide specialized navigation facilities
such as alternative link sorting schemes. Unfortunately, these systems merely make
use of content information, although user characteristics are a valuable resource for
determining which information may be of interest.
In contrast to handheld interfaces, virtual (three-dimensional) worlds are typically
used for exploration and entertainment [3]. Although the environment still needs to
offer goal-directed navigation aids, it is likely that users also will want to look around
for related information they might be interested in. Ideally, such an interface should
posses challenging navigation aids, should invite exploration and should be fun to
use. Enhanced perceived fun will most likely lead to increased time spent in a virtual
world [5], which is interesting from an e-commerce point of view [3].
When providing information through such different interfaces as handheld devices,
virtual worlds and regular web pages – which all serve different user goals – adapta-
tion to the environment is an important issue. This environment does not only consist
of the device being used, but also of the context of use (e.g. waiting at a bus stop) [1].
Our research is aimed at determining what navigation support is needed in different
situations and how user characteristics can be used for selecting the areas of interest.
For this purpose, we will consider three different interfaces which all present the
same information environment:
− a handheld device interface (which is typically used in a goal-directed way)
− a virtual world (which is typically used in an explorative way)
a ‘regular’ web page (which is a mix of both ends of the spectrum)
For each of these interfaces, mechanisms will be developed for adapting navigation
support to the richness of interaction means provided, its – possibly varying – context
of use and the resulting user needs, which might be subject to change as well.
2 Adapting Navigation Aids to Device Characteristics
While navigating a hypermedia document, people use environmental cues and artifi-
cial aids such as site maps so that they can achieve their goals without getting lost.
Their navigation behavior can be categorized as either goal-directed or explorative.
Though mutually exclusive, these activities may become connected: exploration may
result in goal-directed search activities, but the opposite direction is likely as well.
These two navigation types call for different types of navigation assistance [3].
Handheld interfaces most likely invite a goal-directed kind of use – not only be-
cause of the limited interaction mechanisms of such devices, but also because of the
situations in which they are used. Interfaces that are rich in interaction means, such as
virtual worlds, invite mainly explorative behavior.
Online information can be accessed using such an increasing variety of equipment
from almost any place. It is desirable to make use of the same information source for
representation on different devices; duplicate effort could seriously tax human and
machine resources [2] and, even worse, may lead to inconsistency. As stated before,
600 E. Herder and B. van Dijk
navigation support should be adapted to the device characteristics and the context in
which it is used.
Equally important to selecting the right navigation suggestions, is the number of
suggestions that should be given to the user. An overwhelming amount of information
will cause the user to have to spend considerable time scanning through, whereas too
little information might not provide sufficient cues. It is obvious that the ideal number
of navigation suggestions is highly dependent on the device characteristics and the
context of use; users with small screens are reported to follow links less frequently
than those provided with larger screens [2].
As suggested in the introduction, there is not one approach for selecting navigation
aids that fits all users. When user characteristics are taken into account in the process
of adapting navigation support to handheld devices, this might improve the effective-
ness of such small screen interfaces [2].
Alternatively, the same user characteristics can be used for creating a virtual world
that invites exploration by means of visually attractive navigation aids and embodied
agents that support the users’ explorative behavior, leading them to those places they
are interested in [3]. Such a virtual world might be modeled after the real world (e.g.
cities, buildings); it might as well be symbolic, with a topology that is based on non-
geographical relations between information nodes.
3 A Pragmatic, User-Centered Approach
Past research has generally been overly technology-oriented, forgetting that the origi-
nal goal was to deal with usability problems [13]. In our project PALS Anywhere1 we
try to overcome this problem by separating research into a cognitive track and a tech-
nological track, carried out in parallel by researchers from both disciplines.
Cognitive theory is needed for distinguishing the environmental factors that affect
human cognitive task performance. There is no such comprehensive theory, so we
will have to develop limited or practical theories on the influence of accepted features
of cognition on navigation behavior [10], in order to obtain a foundation for the
mechanisms to be incorporated in our models. The applicability of the cognitive theo-
ries and the functionality of the resulting system will be validated by usability tests.
Since there is not one single approach that is most suitable for modeling both user
and environment [7], it is important to know what knowledge such models should
express and how this knowledge can be obtained, before a proper choice can be made.
Observations of user behavior, for example, are commonly represented as statistical
data. However, it might be desirable to propagate this primary knowledge to higher-
level assumptions that apply to more than one specific device or interface. Such con-
ceptual data is typically represented by explicit, logic-based formalisms, which have
strong inferential capacities [12]. Statistical models, on the other hand, are more apt
1 PALS stands for Personal Assistant for onLine Services; the project is supported by the Dutch
Innovative Research Program IOP-MMI. Our research partners are TNO Human Factors and
the University of Utrecht, faculty of Mathematics and Computer Science.
Personalized Adaptation to Device Characteristics 601
to deal with uncertainty and the incremental acquisition of observational data [14]. At
present, the only motivation for preferring a particular modeling technique is its ob-
served success [14]. Since we aspire to find generic formalisms based on practical
empirical theories, we intend to build our models of both user and environment in-
crementally and eclectically.
The models will be implemented in an agent-based framework that is able to rea-
son about the interface, the user and environmental characteristics. This framework
will provide the core functionality of a collaborating personal assistant [8] that learns
about its users and helps them to reach their goals, either solicited or on its own initia-
tive.
Hypermedia documents can be seen as a collection of nodes (separate chunks of
information) that are richly interconnected. This structure facilitates capturing its
contents and its structure in simple but powerful formalisms, such as adjacency matri-
ces and directed graphs. These representations enable assessment of navigational
patterns [9]. With the joint features of these formalisms and user modeling techniques
an overlay can be derived that contains only those connections that are relevant to the
user. Naturally, in the information model the various relation categories should be
categorized. An obvious solution is the use of typed links, such as analogy, abstrac-
tion and simultaneity [4].
This overlay is not primarily meant to be shown to the user, but merely as a start-
ing point for the navigation assistance design. Each connection category can be trans-
lated into an appropriate structuring element, for instance annotated hyperlinks and
contextual menus on web pages or topological layout and landmarks in virtual
worlds.
4 Research Goals
Device characteristics – such as screen size, interface design and means of interaction
– and the context in which these devices are used, have their impact on the user’s
mental representation of an information environment. In our project we want to de-
termine how spatial and temporal abilities are affected and what navigation aids can
help users to find the information they need. An important consideration is the ex-
pected navigation behavior an interface should support, which is either goal-directed,
explorative or a mix of both.
We hypothesize that adaptive hypermedia techniques are an important means for
adapting an interface to the device and the situation in which it is used. Our research
goal is to model user, device and context of use – the ‘Trinity of Context’ – in order
to adapt navigation support to the continuously varying user needs.
A clever selection of navigation aids will offer the user more freedom in naviga-
tion. On tiny screens, omitting all items that are not of interest leaves more space for
relevant navigation support. Analogously, virtual worlds will become more survey-
able; tailored landmarks and personalized assistance in wayfinding prevent the user
from getting lost.
602 E. Herder and B. van Dijk
We will evaluate our theories on three existing devices, significantly different in

intended use, representation and richness of interaction means. Notwithstanding
technological improvements, devices will remain different on these factors, because
of the contexts of use for which they are designed. Therefore we expect the outcomes
of our research will not only be of immediate use, but also be useful in designing
navigation aids of novel interfaces.
References
1. Brusilovsky, P.: Adaptive Hypermedia. User Modeling and User-Adapted Interaction 11.
Kluwer Academic Publishers, The Netherlands, 2001, pp. 87-110
2. Buyukkokten, O., Garcia-Molina, H., Paepcke, A., Winograd, T.: Power Browser: Effi-
cient Web Browsing for PDAs. CHI 2000 Conference Proceedings. ACM, New York,
2000, pp. 430-437
3. Van Dijk, B., Op den Akker, R., Nijholt, A., Zwiers, J.: Navigation Assistance in Virtual
Worlds. Proceedings 2001 Informing Science Conference. Krakow, 2001
4. Greer, J.E., Philip, T.: Guided Navigation Through Hyperspace. Proceedings of the work-
th
shop “Intelligent Educational Systems on the World Wide Web”. 8 World Conference of
the AIED Society, Kobe, 1997, pp. 18-22
5. Hassenzahl, M., Platz, A., Burmester, M., Lehner, K.: Hedonic and Ergonomic Quality
Aspects Determine a Software’s Appeal. CHI 2000 Conference Proceedings. ACM, New
York, 2000, pp. 430-437
6. Kaplan, C., Fenwick, J., Chen, J.: Adaptive Hypertext Navigation Based On User Goals
and Context. User Modeling and User-Adapted Interaction 3. Kluwer Academic Publish-
ers, The Netherlands, 1993, pp. 193-220
7. Kobsa, A.: Generic User Modeling Systems. User Modeling and User-Adapted Interac-
tion 11. Kluwer Academic Publishers, The Netherlands, 2001, pp. 49-63
8. Maes, P.: Agents that Reduce Work and Information Overload. Communications of the
ACM 37, no. 7. ACM, New York, 1994, pp. 31-40
9. McEneaney, J.E.: Graphic and numerical methods to assess navigation in hypertext.
International Journal of Human-Computer Studies 55. Academic Press, London, 2001,
pp. 761-786
10. Neerincx, M., Lindenberg, J., Rypkema, J., Van Besouw, S.: A practical cognitive theory
of Web-navigation: Explaining age-related performance differences. Position Paper CHI
2000 Workshop Basic Research Symposium. ACM, The Hague, 2000
11. Park, J., Kim, J.: Effects of Contextual Navigation Aids on Browsing Diverse Web Sys-
tems. CHI 2000 Conference Proceedings. ACM, New York, 2000, pp. 257-271
12. Pohl, W., Nick, A.: Machine Learning and Knowledge Representation in the LaboUr
Approach to User Modeling. Proceedings of the 7th International Conference on User
Modeling. Banff, Canada, 1999, pp. 197-188
13. Schneider-Hufschmidt, M., Kuehme, T.: Adaptive User Interfaces: principles and prac-
tice. Elsevier Science, Amsterdam, 1993
14. Zukerman, I., Albrecht, D.W.: Predictive Statistical Models for User Modeling. User
Modeling and User-Adapted Interaction 11. Kluwer Academic Publishers, The Nether-
lands, 2001, pp. 5-18
Users Modeling for Adaptive Call Centers
Ilaria Torre
Department of Computer Sciences – University of Torino

Corso Svizzera 185 - 10149 Torino (Italy)
[email protected]
Abstract. The project described in this paper applies the principles of

adaptivity to a “traditional” Call Center in order to support the Operator in the
interaction with the Customer. The system uses the models of both the
Customer and the Operator and builds up the stepwise answer through an
adaptive workflow.
1 Introduction
The term Call Center can be used with several meanings. In this context, a Call Center
is a system which manages the phone calls some human agents (operators) receive
from the users (typically, the customers of a company) or perform toward them. For
more precision, the first type of calls are defined as inbound calls and the second ones
as outbound calls. A classic example of the former are the calls for solving a problem
or getting information, while an example of the latter are the promotional calls.
The operators can play different roles and have different levels of autonomy, but
the important thing is that they are human, with a voice and a heart. Many studies are
demonstrating that lots of people prefer a cool contact on the other side of the line,
instead of the cold contextual helps, mailing systems, automatic answerers, SMS, WAP
and so on. As a matter of fact, the companies have been increasing the
implementations of CTI (Computer Telephony Integration) solutions which are able
to integrate all these channels and manage all the messages in a single workflow,
where the human agent has a central role.
But besides the advantages of supporting customers with hearty agents and higher
flexibility, such a solution represents a problem for the companies as it implies high
costs of work, costs of training, frequent turnover, no standard answers, no control of
the interactions, etc. The project described in this paper aims at finding out an
intelligent solution for such a matter, in order to exploit the advantages of both the
human and the software agents and minimize the disadvantages. As it will be shown,
the idea is that of using adaptivity to manage the workflow of the answering process.
To achieve that, the Call Center System can be seen as a Three Tier Adaptive
System where the function of the web server is performed by the operator: (s)he is an
alive interface1 between the customers and the system (the company). But to represent
an interface, the work of the operator has to be mechanized, becoming similar to a
1 An interesting example of human to human helping system, integrated in an automatic help-

desk, is in [1].

604 I. Torre
software agent’s work and, therefore, yielding the same benefits: interchangeability of
the operators, low costs of training, standard and correct answers and sharing of
knowledge about the users. But it is not enough, adding adaptivity to the workflow,
other benefits can be gained: efficient routing of the calls to the most appropriate
operators, adaptation of the system to the level of expertise of the operator and above
all ability of the operator to adapt to the users even at their first interaction.
Given the importance of the matter, many CRM2 vendors have been developing
solutions to manage the workflow of the answering process, but no one customizes
the screen views that the support systems show to the operators in the answering
workflow, combining together the profiles of the customer and of the operator.
2 How the System Works: The Adaptive Work Flow
Starting from the moment the customer dials up and provides her/his identification
code, or does not provide any code, as a prospect, to the moment the phone call ends,
the whole process is structured as an adaptive workflow (see the figure below which
presents a simplified Use Case of an inbound call), managed by two agents:
?
Calls Operator
Routing Support Problems KB
System System Marketing DB
Customer Operator
1 Voice Call 1. Inbound call: the Customer dials up.

2. Adaptive filtering: the Routing System selects
2 Selects operator the Operator that is right for answering. The rules
used for the selection take into account the model of
the Customer and those of all the Operators.
3. Adaptive supporting: the Support System opens
3Loads the first customized screen a screen, customizing the content and the interface
in accordance with some specific business rules, that
use the models of both the Customer and the
selected Operator.
Handshake
4 4. Handshake between the Customer and the
Asks the question Operator. The Customer asks the question.
5 Starts the answer retrieval process 5. The Operator sets up the parameters to begin the
step by step query, toward the Support System.
6. Adaptive supporting: the Support System builds
the answer, or the first step of the answer, and uses
6 Builds the answer decision trees to define the workflow.
7. Adaptive supporting: the Support System loads
the next screen, taking into account the Customer
model, the Operator model and the answer from the
7 Loads the next customized screen Knowledge Base The screen can contain: i) the
definitive answer, ii) a request of more details, iii)
8 several kinds of suggestions for the operator.
…………..…………………………….. 8. If the screen asks the Operator for modifying or
…………..……………………………..
…………..…………………………….. adding some input parameters, 5,6,7 can be a cycle.
9. If the screen suggests to redirect the call, and/or
9 Sends the call back the Operator has difficulties to go on with the
…………..…………………………….. process, he sends the call back to the Routing
…………..……………………………..
…………..…………………………….. System, which will Select another Operator (cycle
starting at step 2: ! Adaptive Filtering).
Asks for more details to accomplish the answer / or / Provides the final answer 10. If the screen contains a request for more details,
the Operator asks the Customer and a new cycle
10 starts at step 5; if it contains the final answer, the
Answers / or / Expresses his/her level of satisfaction for the answer received Operator provides it and receives her/his comments.
11. If the Customer is satisfied with the answer, the
11 Starts the marketing process Operator can start an action in order to promoting
…………..…………………………….. or getting information about the Customer. He sets
…………..…………………………….. up the parameters to receive Support by the system,
…………..…………………………….. which will begin a new cycle starting at step 7:
! Adaptive Supporting
Users Modeling for Adaptive Call Centers 605
• the “Call Routing System” RS, is activated as soon as the customer dials up and is
responsible for routing the calls to the most suitable operator: adaptive filtering,
• the “Operator Support System” SS, is involved in a more complex activity, that is
the adaptive support of the operator in building the answer or in making a
marketing proposition. This task is complex as it sums up together three jobs:
i)support for problem solving ii) adaptation of the support to the characteristics of
the user iii) adaptation of the support to the characteristics of the operator.
Finally, the component, called “Problems Knowledge Base and Marketing Data
Base”, represents a whole of different kinds of knowledge bases that the SS uses to
support the operator in the process, in particular they contain:
• the rules that define the workflow of the interaction with the customer, build the
stepwise answers and generate the screens for the operator,
• the object of the rules, namely, the knowledge regarding the solutions of technical
problems, information on services and commercial propositions.
3 Users Modeling and Adaptation
Among the components described above, the users modeling knowledge has not been
mentioned. It would have reduced the readabilities of the schema, requiring at least
two new flows for almost all the actual flows of the schema. Indeed, the actions
performed by each agent are preceded by: i) reasoning on the customer’s model and
on the models of all the operators, for the routing agent and ii) reasoning on the
models of the customer and of the selected operator, for the supporting agent.
The definition of the features (of the customer and of the operator), necessary for
users modeling, covers an important part of the project. It has been performed
following the approach in [3], which provides a methodology for identifying the
important user modeling dimensions (UMD) for a given system. Such a method starts
from the hypothesis that the choice of the UMD depends, on the one hand, on the
adaptive goals and tasks of the system, and, on the other, on the domain features.
Goals and features, once decomposed in a sort of tree functional analysis, become the
axes of a matrix which allows us to determine, at each point of intersection, the UMD
that are essential for performing the adaptation. For example, given the goal’s cascade
of “improving the image of the company”!“satisfying customers”!“solving
problems” !“choosing the right operator for the calling customer”, etc., and the
domain features “inbound call”!“technical call”, etc., the features of the customer
that have to be considered are: “customer status (open calls, churn probability, new
account, etc.), “customer commercial value”, “service level required (time of answer,
time of solution, etc.)”, etc. and the features of the operator that have to be taken into
account are: expertise, domain knowledge, speed, previous interaction with that
customer, status (free of calls and messages, with calls in queue, etc), etc.
Following such a methodology all the necessary features of the users have been
identified, and organized in the respective models.
606 I. Torre
• The Customer User Model is distributed and divided in four parts: i) the first one
contains the descriptive features. The data come from the operational systems3
(trading and banking systems, call center, etc.) and are loaded in the
DataWareHouse every night ii) a second part contains some synthetic data
calculated on the basis of the previous data (e.g. risk level, commercial value,
expertise, satisfaction, etc.), iii) a third part contains a ranking of the operators
sorted in accordance with the degree of match with the customer’s model. This
sorting allows the RS to switch the calls very rapidly to the most suitable
operator, iv) the fourth part consists of a daily table with the relevant actions of
the day (pendent calls, request of services, etc.).
• The Operator User Model is divided in four parts too: i) descriptive data ii)
synthetic predictive data concerning expertise, skill, domain knowledge, speed of
answer, commercial and communicative ability, etc. iii) extended predictive data,
regarding the domain knowledge, iv) status (free, with calls in queue, etc.)
The adaptation is performed by the RS and the SS, using different techniques. The
RS selects the operator firing some rules like: “if the customer has no pendent calls
and there are no successful closed calls in the day, then select the first operator in the
rank of the customer’s model and check her/his status”. The rank is calculated as a
weighed scoring. For example a high probability of churn of the customer requires
that the call is routed toward operators with high expertise, domain knowledge and
commercial ability: this is obtained as a weighted combination of such features in the
count for the rank.
As regards the SS, it uses decision trees (see [4]) to identify, step by step, the best
node toward the solution and compose the customized screen for the operator. For any
answer, that the operator receives from the customer, the decision tree produces the
next screen to support the operator. A typical screen contains a central frame with the
answer from the knowledge base or with a new question the operator has to ask the
customer in order to refine her/his problem, a set of dynamic list boxes the operator
has to select communicating the system the questions/answers provided by the
customer, a bar with the progress of the answer process, a customized tool bar with
helps and facilities (more ore less descriptive, according with the operator’s model),
information about the customer and advices on how to manage her/him, according
with the customer’s model (commercial value, risk level, domain knowledge, etc.).
As a conclusion, the project puts together technologies from different fields, trying
to unify them under the common feature of adaptivity. To reduce the complexity
deriving from this, in the first phase, the system will rely on a given Call Center
System - Genesys -, integrating the functions of adaptive routing and adaptive
workflow.
2 The CRM (Customer Relationship Management) area is very wide as it covers all the fields
and applications related to the management of the customer (marketing automation systems,
personalization agents, ecommerce, olap, erp and contact systems. Some of the vendors that
manage, inside their enterprise product or with specific modules, the answering workflow
process are: Siebel, e.Piphany, Pivotal, Remedy, Genesys, etc. (see [2] for an overview).
3 The project aims at being general, but it actually concerns the Call Center of a Bank.
Users Modeling for Adaptive Call Centers 607
References
1. Greer, J., McCalla, G., Cooke, J., Collins, J., Kumar, V., Bishop, A., Vassileva: J.The
Intelligent HelpDesk: Supporting Peer Help in a University Course. In Proc. ITS'98.
LNCS 1452, Springer-Verlag, Berlin (1998) 494-503
2. International Journal of Call Centre Management, http://www.winthrop-publications.co.uk
3. I. Torre: A modular approach for user modelling. In Adaptive Hypermedia and Adaptive
Web-Based Systems 2000. LNCS 1892, Springer-Verlag (2000) 414-420
4. M. Stolze, M. Ströbel: Utility-Based Decision Tree Optimization: A Framework for
Adaptive Interviewing. In Proc. User Modelling 01. LNCS, Springer-Verlag (2001) 105-
116
The Exploratory of Personal Assistants
Lim Tek Yong and Tang Enya Kong
School of Computer Science

University Science Malaysia
11700 Minden, Penang, Malaysia
604-6577888 ext 3002
{tylim, enyakong}@cs.usm.my
Abstract. This paper presents a perspective of using personal assistants in

personalized systems, particularly in both understanding the user model and
generating appropriate adaptation and response. The paper discusses the use of
a personal assistant as a data input source and the expected stereotypes based on
mind-sets. This is the framework of exploring a user-centered design approach
to building up the intelligent personalized system.
1 Introduction
“Users using the web are only one click away from a competitor.” This is the
common phrase used to encourage customer relationship management on the Internet.
In electronic commerce, the development of relationships with remote users is vital to
maintaining customer loyalty. Computers are now acting as an agent to attract and
retain users. Therefore, information technologies have transformed into relationship
technology. Thanks to the advances in technology, dynamic pages can be created on
the fly. Web sites have taken the steps by delivering tailored products and services on
the web in order to establish, maintain and extend a customer base. They will create
stable long-term relationships with repeat customers in the long run. A lot of
personalization systems are being proposed by researchers and companies to cater
web pages to the needs of the current users (Brusilovsky, 2001). However, some data
collection is not appropriate to be used to model the user profile precisely (Kobsa,
Koenemann and Pohl, 2001). Besides that, quite a number of reported researches are
not supported by empirical experimentation (Chin, 2001). So, one might not able to
justify whether the proposed personalized system is usable and efficient. Therefore,
this paper will look into how a personalized system uses a personal assistant as a data
input in the next section.
2 Personal Assistant
The acquisition task is the first step in identifying the data input for personalized
systems (Kobsa, Koenemann and Pohl, 2001). There are three types of data input for
personalized systems, namely data about users’ characteristics, data about users’
computer usage and data about users’ hardware, software and physical environment.

The Exploratory of Personal Assistants 609
However, the paradox of an active user and the task orientation of users should be
taken into account as the users visit a site. The user should not be exposed to a
lengthy registration procedures or initial interviews (e.g. in a short survey format).
Therefore, some researchers suggested quizzes as entertainment and offering
incentives such as free merchandise. The other possible alternative is using a dialog
system integrated with current personalized systems. The area of user modeling for
dialogues has recently seen a welcome and increasing use of empirical methods
(Zukerman and Litman, 2001). We proposed using a personal assistant to extend the
dialogue system on an electronic shopping web site.
Why are personal assistants important in the user modeling? There are two
important reasons. The first reason is that the “human figure” communicating with
users captures their attention and makes them alert to messages delivered to them
(Reategui and Campbell, 2001). A flower shop reported an increase in sales of up to
three -times after a personal assistant, who helps the user to buy flowers and
bouquets, was implemented on the Internet. Perhaps personalized information is more
acceptable and salient if a personal assistant instead of a system-like interface
presents it. A combination of a personal assistant and an information intensive site
can build customer trust and acceptance (Urban, Sultan and Qualls, 1999). Besides
that, several experiments using the “wizard of oz” method have proven that users
benefit from the personal assistant during their online shopping session. However,
most of them were using a pre-recorded session of interaction (McBreen, Shade, Jack
and Wyard, 2000), using pre-defined questions to ask the personal assistant (Urban,
Sultan and Qualls, 1999) and a real human as the assistant on the web (Aberg and
Shahmehri, 2001). The users were having limited options of exploring the personal
assistant’s capabilities.
The second reason is because the personal assistant would allow users to develop a
social relationship and allow a new way of data acquisitions for the user model. De
Angeli, Johnson and Coventry (2001) have found that users would treat the personal
assistant as a new relationship. Their research has shown some understanding of
social rules driving the interaction between a user and a chatterbot. One of the most
important processes in the formation and maintenance of a relationship is that of self-
disclosure (the act of revealing private and personal information to others). It is a
clear sign of trust and commitment to the relationship. In human-relationship and in
Internet Chat Rooms, self-disclosure is intended to be a mutual process (refer to
Example 1). Both parties are required to exchange intimate and personal details as
well as feelings with each other.
Example 1. A typical self-disclosure in Internet chat room.
A: Hello
B: Hi there!
A: a/s/l please.
B: 25 f Malaysia. how about u?
A: me 26 m Singapore.
B: What are u looking 4?
A: looking 4 new friendship.
…
However, most chatterbots only simulate conversation without utilizing any
knowledge about the users and their actual behavior during the online session.
Secondly, they are not able to serve as a medium for customer advice if they are not
able to interpret the individual dialog situation and allow goal-directed strategies to be
610 L.T. Yong and T.E. Kong
pursued. Therefore, some exploration needed to be carried out to determine the real-
time of user chatting with a personal assistant while searching the online shopping
website. Besides that, we need to know how far the self-disclosure can be used as data
input for a personalized system.
In order to explore the effect of a personal assistant, we are using the fixed
stereotypes as the first step for representation references rather than having none.
Ralph’s (2000) preliminary study has shown that chatterbots with a sophisticated
repertoire of conversational skills will fail to be more than entertaining, if the personal
assistant does not treat a user as an individual having specific needs, preferences, etc.
Therefore, an understanding of different user types will be beneficial to help
developing an acceptable and believable personal assistant. Moe (2001) has identified
four types of visitors on the Internet. They are
1. Directed-purchase visits. This user is ready to purchase right away.
2. Search and deliberation visits. This user is researching the information and
eventually intends to buy.
3. Hedonic-browsing visits. This user is doing electronic window-shopping, that is
shopping for pleasure or recreation.
4. Knowledge-building visits. This user is engaged in exploration to know more about
the marketplace – a pursuit that may affect long-term shopping.
In the old-times, experienced sales people learned to distinguish between shoppers
based on their in-store behavior. Sales people are more helpful when shoppers appear
to be very focused in looking for a specific product. However, they are more likely to
ignore shoppers who are merely just “window-shopping”. In an electronic world,
some reports argued that the behavior of users would be changed since most
electronic marketplaces provide a more convenient way for goal-directed users to
search for a product (Wind and Mahajan, 2001; De Kare-Silver, 2000). Now, the
question is “Would the personal assistant be more helpful for users who are goal-
directed, and be more sociable for users with exploratory behavior or vice-versa?” A
user-centered design can be used to refine basic assumptions and the whole
personalized system can be easily redesigned if needed (Petrelli, Angeli and
Convertino, 1999). The reason is that their study pointed out some unpredicted
situations and confirmed some working hypotheses.
3 Mind-Set
The concept of mind-set was originally suggested by Wurtzberg motivational

psychologists (Kulpe, 1904, cited in Gollwitzer, 1996). A mind-set refers to a
“specific cognitive orientation”, imbued with distinct and unique features and can be
of different types. Each mind-set is associated with different thought contents and
different modes of information processing (Heckhausen & Gollwitzer, 1987). A
mind-set can divided in two types: goal-directed mind-sets and experiential mind-sets.
Goal-oriented mind-sets consist of both deliberative and implemental mind-sets. A
deliberative mind-set refers to a cognitive orientation where the consumer is intent on
collecting and processing information, and is common in the problem-identification,
information search, and decision-making stages of consumer behavior. On the other
hand, an implemental mind-set refers to an action-focused, cognitive orientation that
The Exploratory of Personal Assistants 611
occurs after the decision has been made, and serves to facilitate smooth action
execution for goal attainment.
An exploratory mind-set refers to a consumer’s cognitive orientation to encounter
new experiences and to satisfy his or her curiosity. On the other hand, experiential
mind-set is the hedonic mind-set where the individual de-emphasizes cognitions, and
focuses instead on the sensory elements of experience.
Security and privacy infrastructures should be used that both protect user data
against possible attacks and allow users to remain highly anonymous with respect to
the personalized system while still enjoying personalization. One of the suggested
approaches is to provide comprehensive and intelligible advance notice to users about
all the data that is to be collected, processed and transferred. As a result, this will
increase users’ trust in the application and is mandated by virtually all privacy laws.
Furthermore, users can personalize (“opt-in”) the processing of their data in voluntary
consent.
4 Ongoing Work
This paper presents a perspective of using a personal assistant in personalized

systems, particularly in both understanding the user model and generating appropriate
adaptation and response. The research question is
1. What types of data input can be collected from first-time visitors, returning
visitors, infrequent and frequent user needs when using a personal assistant?
2. What types of adaptation are needed for personal assistants when encountering
first-time visitors, returning visitors, infrequent and frequent user needs?
At present, Moe (2001) proposed a typology of shopping strategies based on the
user’s click-stream in a website. Her result showed that each user can be classified
into the four mind-set categories based on a series of click-streams. Our research is to
investigate the user interaction, based on mind-set, with the personal assistant when
visiting/shopping on a website. The first step is to identify the data input that can be
used as a proper indication of the type of users. We are planning to build a prototype
to simulate the personal assistant on a shopping website. Then we will perform online
focus group sessions to collect the experiment data. The second step is to identify
types of adaptation that are needed to support all these users. In this stage, we will
prepare adaptation materials that are suitable for each type of users. Then we will
perform a quantitative study online to collect the experiment data.
Therefore, our research is focused to comprehend users’ acceptance of and
reactions to the introduction of human characteristics in user interfaces. Defining the
users’ characteristics that may affect their acceptance of personal assistants and
helping predict their behavior during the interaction is a major challenge. Before such
adaptation can be designed and models can be developed, researchers must first
understand the underlying interaction patterns of the users, and how to identify
different types of behavior. This paper is a first step in that direction.
To conclude, creativity can support the different styles of presenting information to
users. For example, a portal can create a simple style if the users are goal-directed
searchers. They can create a more elaborate and artistic style if the users are
experiential surfers. Therefore, a personalized system with personal assistant has to be
intelligent enough to support all these users.
612 L.T. Yong and T.E. Kong
References
1. Aberg, J. and Shahmehri, N. (2001) An Empirical Study of Human Web

Assistants: Implications for User Support in Web Information Systems.
Proceedings of SIGCHI ’01 Seatle, WA, USA.
2. Brusilovsky, P. (2001) Adaptive Hypermedia. Journal of User Modeling and
User-Adapted Interaction, 11, 87-110.
3. Chin, D.N. (2001) Empirical Evaluation of User Models and User-Adapted
Systems. Journal of User Modeling and User-Adapted Interaction, 11, 181-194.
4. De Angeli, A, Johnson, G.I. and Coventry, L. (2001) The unfriendly user:
exploring social reactions to chatterbots. In M.G. Helander, H.M. Kalid, T. Ming
Po (Eds). Proceedings of the International Conference oh Affective Human
Factor Design, Asean Academic Press, pp. 467-474.
5. De Angeli, A., Lynch, P. and Johnson G. (2001) Personifying the e-market: A
framework for social agents. Proceedings of Interact 2001 Conference on
Human-Computer Interaction, Tokyo, Japan, July 9-13, 2001.
6. De Kare-Silver, M. (2000) E-Shock 2000 The electronic shopping revolution:
strategies for retailers and manufacturers. Macmillan Press Ltd, London.
7. Gollwitzer, P.M. (1996) “the Volitional Benefits of Planning.” In Gollwitzer,
P.M. and Bargh, J. A. (eds), The Psychology of Action: Linking Cognition and
Motivation to Behavior (pp. 287-312) New York: Guilford Press.
8. Heckhausen, H. and Gollwitzer, P.M. (1987). “Thought Contents and Cognitive
Functioning in Motivational versus Volitional States of Mind.” Motivation and
Emotion, 11, 101-120.
9. Kobsa, A., Koenemann, J. and Pohl, W. (2001) Personalized Hypermedia
Presentation Techniques for Improving Online Customer Relationship. The
Knowledge Engineering Review 16(2), 111-155.
10. McBreen, H., Shade, P., Jack, M. and Wyard, P. (2000) Experimental
Assessment of the Effectiveness of Synthetic Personae for Multi-Modal E-Retail
Applications. Proceedings of Agents 2000 Barcelona Spain.
11. Moe, W.W. (2001) Buying, Searching or Browsing: Differentiating between
Online Shoppers Using In-Store Navigational Clickstream. Journal of Consumer
Psychology (forthcoming)
12. Petrelli, D., De Angeli, A. and Convertino (1999) A User-Centered Approach to
User Modeling. Proceedings of the Seventh International Conference,UM'99,
255-264
13. Reategui, E. and Campbell, J.A. (2001) The role of personified Agents in
Recommendation Delivery. Proceedings of 2001 ACM SIGIR Workshop on
Recommender Systems, New Orleans, LA, USA.
14. Urban, G. L., Sultan, F. and Qualls, W. (1999) Design and Evaluation of A Trust
Based Advisor on the Internet. MIT
http://ebusiness.mit.edu/research/papers/Urban.pdf
15. Wahlster, W. (2001) Virtual Sales Agents for Electronic Commerce. ACAI 2001
Prague.
16. Wind, J. and Mahajan, V. (2001) Digital Marketing: Global Strategies From the
World’s Leading Experts, John Wiley & Sons, Inc., New York.
17. Zukerman, I and Litman, D. (2001) Natural Language Processing and User
Modeling: Synergies and Limitations. Journal of User Modeling and User-
Adapted Interaction, 11, 129-158.
Author Index
Aaronson, J. 1 Chen, G.-L. 164

Abrahão, S.A. 358 Chen, J. 164
Abraham, D. 456 Choi, P. 531
Aedo, I. 317, 419 Cini, A. 497
Aerts, A. 388 Cisternino, A. 392
Ardissono, L. 14 Coelho, T.A.S. 274
Arjona, J.L. 24 Coffey, J.W. 380
Armstrong, R. 428 Conejo, R. 376, 480
Aroyo, L. 122 Conlan, O. 100
Arruabarrena, R. 363 Console, L. 112
Ashman, H. 580 Corchuelo, R. 24
Avesani, P. 462 Correa, J.S. 338
Cotter, P. 328
Bailey, C. 36 Cristea, A. 79, 122
Bajuelos, a.A.L. 90 Czarkowski, M. 384
Barbosa, S.D.J. 274
Barra, M. 47 David, A.A. 480
Baudisch, P. 58 Decouchant, D. 535
Beck, R.E. 488 Dempster, E. 492
Behrendt, W. 368 Dı́az, P. 317, 419, 476
Bental, D. 492 Dijk, B. van 598
Berka, T. 368 Dolog, P. 586
Beschoren, K. 338
Esuli, A. 392
Bieliková, M. 586
Bontcheva, K. 69 Favela, J. 535
Boticario, J.G. 143 Fernández de Arriba, M. 501
Brailsford, T. 580 Fogli, D. 484
Bra, P. De 388 Fons, J. 358
Bravo, C. 560 Frasincar, F. 133
Breda, A.M. 90 Fresta, G. 484
Brueckner, L. 58 Funk, P. 436
Bruen, C. 100
Brunstein, A. 372 Gams, E. 368
Bry, F. 472 Garcı́a-Cabrera, L. 284
Buendı́a, F. 476 Garcı́a, E. 317
Bueno, D. 376, 480 Garcı́a, F. 592
Garcı́a, F.J. 505
Calvi, L. 79 Gargan, M. 100
Carmona, C. 376, 480 Gaudioso, E. 143
Carrara, P. 484 Gensel, J. 444
Carro, R.M. 90 Gentile, E. 432
Casanova, M.A. 274 Gil, A.B. 505, 592
Cassel, L.N. 488 Gioria, S. 112
Castillo, G. 90 Goel, M. 510
Cawsey, A. 492 González, J. 264
614 Author Index
González, M. 358 Lucena, C.J.P. de 274

González Rodrı́guez, M. 564
Gouli, E. 153 Ma, W.-Y. 164
Goy, A. 14 Macı́as, M. 264
Grigoriadou, M. 153 MacKinnon, L. 492
Grimsley, M. 397 Madeira, M. 338
Guessoum, Z. 592 Maes, S.H. 401
Gutiérrez, J. 363 Maglio, P. 47
Gu, X.-D. 164 Maier, P. 428
Guzman, E. 376 Malek, M. 406
Martens, A. 539
Hall, W. 36, 428 Martin, B. 296, 543
Hasegawa, S. 518 Martı́nez-Enrı́quez, A.M. 535
Healey, J. 401 Martin, H. 444
Henze, N. 174 Martinovska, C. 411
Herder, E. 598 Martins, L.C. 274
Hosn, R. 401 Marwick, D. 492
Houben, G.-J. 133 Massa, P. 462
Hübscher, R. 184 Masthoff, J. 415
Hüttenhain, R. 572 McGowan, J.P. 514
Medina-Medina, N. 284
Jameson, A. 193 Meehan, A. 397
Miettinen, M. 242
Kanawati, R. 406 Millard, D.E. 36
Kashihara, A. 518 Mitrovic, A. 296, 543
Kay, J. 203, 384 Mitsuhara, H. 547
Kim, Y. 522 Mizzaro, S. 306
Klemke, R. 572 Montero, S. 419
Koch, N. 213 Moore, A. 580
Kong, T.E. 608 Morán, A.L. 535
Koychev, I. 223 Mun, H. 424
Kraus, M. 472 Mussio, P. 484
Kraus, S. 527
Nakai, Y. 452
Kravcik, M. 572
Naumann, A. 372
Krems, J.F. 372
Negro, A. 47
Kröner, A. 527
Nejdl, W. 174
Kukulenz, D. 233
Ng, M.H. 428
Kummerfeld, B. 203
Ng, W. 551
Kurhila, J. 242
Nokelainen, P. 242
Kushmerick, N. 514
Nori, M. 462
Langweg, H. 440 Ochi, Y. 547
Lauder, P. 203 Ok, S. 424, 522
Lee, M. 531 Oliveira, J.P.M. de 338
Lieberman, H. 2, 253 Ortega, M. 560
Liu, H. 2, 253
López Brugos, J.A. 501 Pacey, D. 492
López-Cuadrado, J. 363 Pacini, G. 392
López Pérez, B. 564 Papanikolaou, K. 153
Lombardi, I. 112 Paredes, P. 556
Author Index 615
Parets-Llorca, J. 284 Susi, A. 462

Pastor, O. 358
Paternò, F. 505 Tasso, C. 306
Paule Ruı́z, M.P 564 Tirri, H. 242
Pérez Pérez, J.R. 564 Toppano, E. 576
Pérez, T.A. 363 Torasso, P. 14
Pesin, L. 572 Toro, M. 24
Petrone, G. 14 Torre, I. 112, 603
Plantamura, P. 432 Toyoda, J. 518
Puntambekar, S. 184 Tsaban, L. 527
Redondo, M.A. 560 Uhrmacher, A.M. 539

Reich, S. 368
Vadillo, J.A. 363
Rodriguez, P. 556
Valdeni de Lima, J. 497
Rodrı́guez-Fortiz, M.J. 284
Verdejo, M.F. 560
Roselli, T. 432
Verdin, R. 338
Rossano, V. 432
Vicari, R.M. 338, 466
Ruiz, A. 24
Villanova-Oliver, M. 444
Sánchez, F. 264 Wade, V. 100
Saponaro, V. 432 Wahlster, W. 12
Sarkar, S. 510 Wainer, R. 338
Scarano, V. 47 Waniek, J. 372
Schwarzkopf, E. 193 Warpechowski, M. 338
Segnan, M. 14 Weal, M.J. 36
Selker, T. 253 Weber, G. 448
Sicilia, M.-Á. 317 Weibelzahl, S. 448
Signore, O. 568 Williams, H. 492
Silveira, R.A. 466 Wirsing, M. 213
Simi, M. 392 Wolz, U. 488
Smits, D. 388 Woo, Y. 424, 522, 531
Smyth, B. 328, 514
Sollenborn, M. 436 Yacef, K. 456
Souto, M.A.M. 338 Yamada, S. 452
Spalka, A. 440 Yano, Y. 547
Specht, M. 572 Yau, J. 551
Stash, N. 388 Yong, L.T. 608
Stewart, C. 580
Sunderam, A.V. 348 Zakaria, M.R. 580
Surano, V. 112 Zanella, R. 338

3 540 47952 X

Uploaded by

Copyright:

Available Formats

3 540 47952 X

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

3 540 47952 X

Uploaded by

Copyright:

Available Formats

Lecture Notes in Computer Science 2347

Edited by G. Goos, J. Hartmanis, and J. van Leeuwen

Cataloging-in-Publication Data applied for

Die Deutsche Bibliothek - CIP-Einheitsaufnahme

CR Subject Classification (1998): H.5.4, H.4, H.5, H.3

© Springer-Verlag Berlin Heidelberg 2002

Hypermedia as a means for accessing information was ﬁrst proposed by Vannevar

May 2002 Paul De Bra

AH 2002 was organized by the University of Málaga, Spain.

Local Organization Chair

Doctoral Consortium Chairs

Nicola Henze, Institut für Technische Informatik, Hannover

Local Organization Committee

Resource-Adaptive Interfaces to Hybrid Navigation Systems . . . . . . . . . . . . . 12

Automatic Extraction of Semantically-Meaningful Information

Towards Open Adaptive Hypermedia . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

GAS: Group Adaptive System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

TV Scout: Lowering the Entry Barrier

Adaptivity, Adaptability, and Reading Behaviour:

Towards Generic Adaptive Systems: Analysis of a Case Study . . . . . . . . . . . 79

A Methodology for Developing Adaptive Educational-Game Environments 90

Multi-model, Metadata Driven Approach to Adaptive Hypermedia Services

Adaptation and Personalization on Board Cars:

Adaptive Authoring of Adaptive Educational Hypermedia . . . . . . . . . . . . . . . 122

Hypermedia Presentation Adaptation on the Semantic Web . . . . . . . . . . . . . 133

User Data Management and Usage Model Acquisition

Personalizing Assessment in Adaptive Educational Hypermedia Systems . . 153

Visual Based Content Understanding towards Web Adaptation . . . . . . . . . . 164

Knowledge Modeling for Open Adaptive Hypermedia . . . . . . . . . . . . . . . . . . . 174

Adaptive Navigation for Learners in Hypermedia Is Scaﬀolded Navigation . 184

Pros and Cons of Controllability: An Empirical Study . . . . . . . . . . . . . . . . . . 193

Personis: A Server for User Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203

The Munich Reference Model for Adaptive Hypermedia Applications . . . . . 213

Tracking Changing User Interests through Prior-Learning of Context . . . . . 223

Prediction of Navigation Proﬁles in a Distributed Internet Environment

Educo – A Collaborative Learning Environment

GOOSE: A Goal-Oriented Search Engine with Commonsense . . . . . . . . . . . . 253

On Adaptability of Web Sites for Visually Handicapped People . . . . . . . . . . 264

A Framework for Filtering and Packaging Hypermedia Documents . . . . . . . 274

Adaptation in an Evolutionary Hypermedia System:

Evaluating the Eﬀects of Open Student Models on Learning . . . . . . . . . . . . . 296

Ephemeral and Persistent Personalization in Adaptive Information Access

Fuzzy Linguistic Summaries in Rule-Based Adaptive Hypermedia Systems 317

The Plight of the Navigator:

Towards an Adaptive Web Training Environment

Automated Personalization of Internet News . . . . . . . . . . . . . . . . . . . . . . . . . . 348

On Evaluating Adaptive Systems for Education . . . . . . . . . . . . . . . . . . . . . . . . 363

Recommending Internet-Domains Using Trails and Neural Networks . . . . . . 368

Learning Grammar with Adaptive Hypertexts: Reading or Searching? . . . . 372

SIGUE: Making Web Courses Adaptive . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 376

An Ontology-Guided Approach to Content Adaptation in LEO:

A Scrutable Adaptive Hypertext . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 384

AHA! Meets AHAM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 388

A Web-Based Selling Agent That Maintains Customer Loyalty

Behavioral Sequences: A New Log-Coding Scheme

A Fuzzy-Based Approach to User Model Reﬁnement

Towards an Authoring Coach for Adaptive Web-Based Instruction . . . . . . . 415

Generation of Personalized Web Courses Using RBAC . . . . . . . . . . . . . . . . . . 419

An Automatic Rating Technique Based on XML Document . . . . . . . . . . . . . 424