Machine Learning For Information Retrieval

Machine Learning for Information Retrieval: Neural
Networks, Symbolic Learning, and Genetic Algorithms
Hsinchun Chen
University of Arizona, Management Information Systems Department, Karl Eller Graduate School of Management,
McClelland Hall 4302, Tucson, AZ 8572 1. E-mail: [email protected]
Information retrieval using probabilistic techniques has attracted significant attention on the part of researchers in
information and computer science over the past few decades. In the 198Os, knowledge-based
techniques also
made an impressive contribution to intelligent
information retrieval and indexing. More recently, information science researchers have turned to other newer artificial-intelligence-based
inductive learning techniques including
neural networks, symbolic learning, and genetic algorithms. These newer techniques, which are grounded on
diverse paradigms, have provided great opportunities for
researchers to enhance the information processing and retrieval capabilities of current information storage and retrieval systems. In this article, we first provide an overview
of these newer techniques and their use in information
science research. To familiarize readers with these techniques, we present three popular methods: the connectionist Hopfield network; the symbolic ID3/ID5R; and evolution-based genetic algorithms. We discuss their knowledge representations
and algorithms in the context of
information retrieval. Sample implementation
and testing
results from our own research are also provided for each
technique. We believe these techniques are promising in
their ability to analyze user queries, identify users information needs, and suggest alternatives for search. With
proper user-system
interactions,
these methods can
greatly complement
the prevailing
full-text,
keywordbased, probabilistic, and knowledge-based
techniques.
Introduction
In the past few decades,the availability of cheap and

effective storage devices and information systems has
prompted the rapid growth and proliferation of relational, graphical, and textual databases.Information collection and storageefforts have become easier,but effort
required to retrieve relevant information has become significantly greater, especially in large-scale databases.
Received September 29, 1993; revised March 25, 1994; accepted June
1, 1994.
0 1995 John Wiley &Sons. Inc.
JOURNAL
OF THE AMERICAN
SOCIETY FOR INFORMATION
This situation is particularly evident for textual databases,which are widely usedin traditional library science
environments, in business applications (e.g., manuals,
newsletters,and electronic data interchanges),and in scientific applications (e.g., electronic community systems
and scientific databases).Information stored in thesedatabasesoften has become voluminous. fragmented, and
unstructured after yearsof intensive use.Only userswith
extensive subject area knowledge, system knowledge,
and classification scheme knowledge (Chen & Dhar,
1990) are able to maneuver and explore in these textual
databases.
Most commercial information retrieval systems still
rely on conventional inverted index and Boolean querying techniques. Even full-text retrieval has produced less
than satisfactory results (Blair & Maron, 1985).Probabilistic retrieval techniques have been used to improve the
retrieval performance of information retrieval systems
(Bookstein & Swanson, 1975: Maron & Kuhns, 1960).
The approach is based on two main parameters, the
probability of relevance and the probability of irrelevance of a document. Despite various extensions, probabilistic methodology still requires the independenceassumption for terms and it suffers from difficulty of estimating term-occurrence parameters correctly (Gordon,
1988; Salton, 1989).
Since the late 1980s knowledge-based techniques
have been used extensively by information science researchers.These techniques have attempted to capture
searchersand information specialistsdomain knowledge and classification scheme knowledge, effective
searchstrategies,and query refinement heuristics in document retrieval systems design (Chen & Dhar, 1991).
Despite their usefulness,systemsof this type are consideredperformancesystems(Simon, 1991)-they only perform what they were programmed to do (i.e., they are
without learning ability). Significant efforts are often required to acquire knowledge from domain experts and
to maintain and update the knowledge base.
SCIENCE. 46(3):194-216,
1995
CCC 0002-8231/95/030194-23
A newer paradigm, generally considered to be the machine learning approach, has attracted attention of researchersin artificial intelligence, computer science,and
other functional disciplines such as engineering, medicine, and business (Carbonell, Michalski, & Mitchell,
1983; Michalski, 1983; Weiss & Kulikowski, 1991). In
contrast to performance systems, which acquire knowledgefrom human experts, machine learning systems acquire knowledge automatically from examples, that is,
from source data. The most frequently used techniques
include symbolic, inductive learning algorithms such as
ID3 (Quinlan, 1979) multiple-layered, feed-forward
neural networks such as backpropagation networks
(Rumelhart, Widrow, & Lehr, 1986), and evolutionbased genetic algorithms (Goldberg, 1989). Many information science researchershave started to experiment
with these techniques as well (Belew, 1989; Chen &
Lynch, 1992; Chen et al., 1993; Gordon, 1989; Kwok,
1989).
In this article, we aim to review the prevailing machine learning techniques and to present several sample
implementations in information retrieval to illustrate
the associated knowledge representations and algorithms. Our objectives are to bring these newer techniques to the attention of information science researchers by way of a comprehensive overview and discussion
of algorithms. We review the probabilistic and knowledge-basedtechniques and the emerging machine learning methods developed in artificial intelligence (AI). We
then summarize some recent work adopting AI techniques in information retrieval (IR). After the overview,
we present in detail a neural network implementation
(Hopfield network), a symbolic learning implementation
(ID3 and IDSR), and a genetic algorithms implementation. Detailed algorithms, selectedIR examples, and preliminary testing results are also provided. A summary
concludes the study.
Information Retrieval Using Probabilistic,
Knowledge-Based,
and Machine Learning
Techniques
In classical information retrieval models, relevance

feedback, document space modification, probabilistic
techniques, and Bayesian inference networks are among
the techniques most relevant to our research.In this section, we first summarize important findings in these areasand then present some results from knowledge-based
systems researchin information retrieval. However, our
main purpose will be to present research in machine
learning for information retrieval. Similarities and
differencesamong techniques will be discussed.
RelevanceFeedback and Probabilistic Models in IR
One of the most important and difficult operations in
information retrieval is to generatequeries that can suc-
JOURNAL
cinctly identify relevant documents and reject irrelevant

documents. Since it is often difficult to accomplish a successful searchat the initial try, it is customary to conduct
searches iteratively and reformulate query statements
based on evaluation of the previously retrieved documents. One method for automatically generating improved query formulations is the well-known relevancefeedback process (Ide, I97 1; Ide & Salton, 1971; Rocchio, 1971; Salton, 1989). A query can be improved iteratively by taking an available query vector (ofterms) and
adding terms from the relevant documents, while subtracting terms from the irrelevant documents. A single
iteration of relevance feedback usually produces improvements of from 40% to 60% in searchprecision (Salton, 1989). A similar approach can also be used to alter
the document representation. Document-vector modijcation changes and improves document indexes based
on the user relevance feedback of relevant and irrelevant
documents (Brauen, 1971). Using such a technique, the
vectors of documents previously retrieved in responseto
a given query are modified by moving relevant documents closer to the query and at the same time moving
irrelevant documents away from the query. While the
relevance feedback procedure is efficient and intuitively
appealing, it does not attempt to analyze characteristics
associatedwith the relevant and irrelevant documents to
infer what concepts (terms) are most appropriate for
representing a given query (or queries).
In probabilistic information retrieval, the goal is to estimate the probability of relevanceof a given document
to a user with respect to a given query. Probabilistic assumptions about the distribution of elements in the representations within relevant and irrelevant documents
are required. Using relevance feedback from a few documents, the model can be applied to estimate the probability of relevance for the remaining documents in a collection (Fuhr & Buckley, 1991; Fuhr & Pfeifer, 1994;
Gordon, 1988). To simplify computation, an assumption is usually made that terms are distributed independently (Maron & Kuhns, 1960). Fuhr and his coworkers
discussedprobabilistic models as an application of machine learning. They presented three different probabilistic learning strategiesfor information retrieval. First,
the classical binary independence retrieval model (Robertson & Sparck Jones, 1976; Yu & Salton, 1976) implemented a query-oriented strategy. In the relevance feedback phase, given a query, relevance information was
provided for a set of documents. In the application
phase,this model can be applied to all documents in the
collection, but only for the same initial query. The second document-oriented strategy collected relevancefeedback data for a specific document from a set of queries
(Maron & Kuhns, 1960). The parameters derived from
these data can be used only for the same document, but
for all queries submitted to the system. Neither of these
strategiescan be generalizedto all documents and for all
OF THE AMERICAN
SCIENCE-April
1995
195
. queries. Fuhr et al. proposed a third, feature-oriented

strategy. In query-oriented and document-oriented strategies,the concept of abstraction was adopted implicitly
by regarding terms associatedwith the query or the document, instead of the query or document. In this featureoriented strategy, abstraction was accomplished by using
features of terms (e.g.,the number of query terms, length
of the document text, the with-document frequency of
a term, etc.) instead of terms themselves. The featureoriented strategy provides a more general form of probabilistic learning and produces bigger learning samples for
estimation; but the disadvantage is the heuristics required to define appropriate features for analysis. After
transforming terms into features, Fuhr et al. (1990)
adopted more sophisticated general-purpose statistical
and machine learning algorithms such as regression
methods and the decision-tree building ID3 algorithm
(Quinlan, 1986) for indexing and retrieval. In summary,
by using features of terms instead of terms, Fuhr et al.
were able to derive larger learning samples during relevance feedback. The general-purpose analytical techniques of regression methods and ID3 they adopted are
similar to the techniques to be discussedin this article.
The use of Bayesian classification and inference networks for information retrieval and indexing represents
an extension of the probabilistic models (Maron &
Kuhns, 1960; Turtle & Croft, 1990). The basic inference
network consists of a document network and a query
network (Turtle & Croft, 1990, 1991; Tzeras & Hartmann, 1993) that is intended to capture all of the significant probabilistic dependenciesamong the variables representedby nodes in the document and query networks.
Given the prior probabilities associated with the documents and the conditional probabilities associated with
the interior nodes, the posterior probability associated
with each node in the network can be computed using
Bayesian statistics. The feedback process in a Bayesian
inference network is similar to conventional relevance
feedback and the estimation problems are essentially
equivalent to those observed in probabilistic models.
Tzeras and Hartmann (1993) showed that the network
can be applied for automatic indexing in large subject
fields with encouraging results, although it does not perform better than the probabilistic indexing technique described in Fuhr et al. ( 1990). Turtle and Croft ( 1991)
showed that, given equivalent document representations
and query forms, the inference network model performed better than conventional probabilistic models.
Although relevance feedback and probabilistic
models exhibit interesting query or document refinement capabilities, their abstraction processesare based
on either simple addition/removal of terms or probabilistic assumptions and principles. Their learning behaviors are very different from those developed in symbolic
machine learning, neural networks, and genetic algorithms. In the following two subsections,we will first re-
196
JOURNAL
OF THE AMERICAN
view knowledge-based information retrieval, and then

provide an extensive discussion of the recent machine
learning paradigms for information retrieval.
Knowledge-BasedSystems in IR
Creating computer systemswith knowledge or intelligence has long been the goal of researchersin artificial
intelligence. Many interesting knowledge-basedsystems
have been developed in the past few decadesfor such applications as medical diagnosis, engineering troubleshooting, and business decisionmaking (Hayes-Roth &
Jacobstein, 1994). Most of these systems have been developed basedon the manual knowledge acquisition process, a significant bottleneck for knowledge-based systems development. A recent approach to knowledge elicitation is referred to as knowledge mining or
knowledge discovery (Frawley, Pietetsky-Shapiro, &
Matheus, 1991; Pietetsky-Shapiro, 1989). Grounded on
various AI-based machine learning techniques, the approach is automatic and it acquires knowledge or identifies patterns directly from examples or databases.We review some important work in knowledge-basedsystems
in IR and learning systems in IR, respectively, in the next
two subsections.
There have been many attempts to capture information specialistsdomain knowledge, search strategies,
and query refinement heuristics in document retrieval
systems design. Some of such systems are computerdelegated,in that decisionmaking has been delegatedto
the system and some are computer-assisted, wherein
users and the computer form a partnership (Buckland &
Florian, 199 1). Becausecomputer-assisted systems have
been shown to be more adaptable and useful for search
tasks than computer-delegated systems, many knowledge-basedsystems of this type have been developed for
IR over the past decade.
CoalSORT (Monarch & Carbonell, 1987) a knowledge-basedsystem, facilitates the use of bibliographic databases in coal technology. A semantic network, representing an experts domain knowledge, embodies the systems intelligence. PLEXUS, developed by Vickery and
Brooks (1987), is an expert system that helps users find
information about gardening. Natural language queries
are accepted.The system has a knowledge base of search
strategiesand term classifications similar to a thesaurus.
EP-X (Smith et al., 1989) is a prototype knowledgebasedsystem that assistsin searchingenvironmental pollution literature. This system makes extensive use of domain knowledge, represented as hierarchically defined
semantic primitives and frames. The system interacts
with users to suggest broadening or narrowing operations. GRANT, developed by Cohen and Kjeldsen
( 1987) is an expert system for finding sourcesof funding
for given research proposals. Its search method-constrained spreading activation in a semantic network-
SCIENCE-April
1995
makes inferences about the goals of the user and thus

finds information that the user has not explicitly requested but that is likely to be useful. Foxs CODER system (Fox, 1987) consists of a thesaurus that was generated from the Handbook of Artificial Intelligence and
Collins Dictionary. In CANSEARCH (Pollitt, 1987) a
thesaurus is presented as a menu. Users browse and select terms for their queries from the menu. It was designed to enable doctors to search the MEDLINE medical databasefor cancer literature. The Intelligent Intermediary for Information Retrieval (13R),developed by
Croft and Thompson ( 1987), consists of a group of experts that communicate via a common data structure,
called the blackboard. The system consists of a user
model builder, a query model builder, a thesaurus expert, a search expert (for suggesting statistics-based
search strategies), a browser expert, and an explainer.
The IOTA system, developed by Chiaramella and Defude ( 1987),includes natural languageprocessingof queries, deductive capabilities (related to user modeling,
search strategies definition, use of expert and domain
knowledge), management of full-text documents, and
relevanceevaluation of answers.Chen and Dhars ( 1991)
METACAT incorporates several human search strategies and a portion of the Library of Congress Subject
Headings (LCSH) for bibliographic search. The system
also includes a branch-and-bound algorithm for an automatic thesaurus (LCSH) consultation process.
The National Library of Medicines thesaurus projects are probably the largest-scale effort that uses the
knowledge in existing thesauri. In one of the projects,
Rada and Martin (Martin & Rada, 1987; Rada et al.,
1989)conducted experiments for the automatic addition
of concepts to MeSH (Medical Subject Headings) by including the CMIT (Current Medical Information and
Terminology) and SNOMED (Systematized Nomenclature of Medicine) thesauri. Accessto various sets of documents can be facilitated by using thesauri and the connections that are made among thesauri. The Unified
Medical Language System (UMLS) project is a longterm effort to build an intelligent automated system that
understands biomedical terms and their interrelationships and uses this understanding to help users retrieve
and organize information from machine-readable
sources (Humphreys & Lindbergh, 1989; Lindbergh &
Humphreys, 1990; McCray & Hole, 1990). The UMLS
includes a Metathesaurus, a Semantic Network, and an
Information Sources Map. The Metathesaurus contains
information about biomedical concepts and their representation in more than ten different vocabularies and
thesauri. The Semantic Network contains information
about the types of terms (e.g., disease, virus, etc.)
in the Metathesaurus and the permissible relationships
among these types. The Information Sources Map contains information about the scope, location, vocabulary,
JOURNAL
and access conditions of biomedical databases of all

kinds.
Another important component of information retrieval is user modeling capability, which is a unique
characteristic of reference librarians. During the user-librarian consultation process, the librarian develops an
understanding of the type of user being dealt with on the
basis of verbal and nonverbal clues. Usually, the educational level of the user, the type of question, the way the
question is phrased, the purpose of the search, and the
expected searchresults all play major roles in helping the
librarian determine the needs of the user. The librarian,
in essence,createsmodels of the user profile and the task
requirements during the consultation process.
User modeling has played a crucial role in applications such as question-answering systems,intelligent tutoring systems, and consultation systems (Appelt, 1985;
Chen & Dhar, 1990; Sleeman, 1985; Swarthout, 1985;
Zissos & Witten, 1985). An intelligent interface for document retrieval systems must also exhibit the user-modeling capability of experienced human intermediaries.
Daniels proposed a frame-basedrepresentation for a user
model and rules for interacting with the users. She has
shown that user modeling is a necessaryfunction in the
presearch information interaction (Daniels, 1986).
Richs Grundy system builds models of its users,with the
aid of stereotypes,and then usesthose models to guide it
in its task, suggestingnovels that people may find interesting (Rich, 1979a, 1979b, 1983). IR-NLI II (Brajnik,
Guida, & Tasso, 1988) incorporates user modeling into
a domain-independent bibliographic retrieval expert system. A user model is built based on the users amount of
domain knowledge and searchexperience.
Despite successesin numerous domains, the development process for knowledge-basedsystemsis often slow
and painstaking. Knowledge engineersor system designers need to be able to identify subject and classification
knowledge from some sources(usually some domain experts) and to represent the knowledge in computer systems. The inference engines of such systems, which
mainly emulate human problem-solving strategies and
cognitive processes(Chen & Dhar, 1991), may not be
applicable acrossdifferent applications.
After examining the potential contribution of knowledge-basedtechniques (natural languageprocessingand
expert systems,in particular) to the information retrieval
and management tasks, Sparck Jones ( 1991) warned that
it is important not to overestimate the potential of such
techniques for IR. She argued that for really hard tasks
we will not be able to replace humans by machines in the
foreseeablefuture and many information operations are
rather shallow, linguistic tasks, which do not involve
elaborate reasoning or complex knowledge. However,
she believed AI can contribute to specializedsystemsand
in situations where users and systemscomplement each
other (i.e., computer-assistedsystems).
OF THE AMERICAN
SCIENCE-April
1995
197
Learning Systems: Neural Networks, Symbolic

Learning, and Genetic Algorithms
Unlike the manual knowledge acquisition process
and the linguistics-based natural language processing
technique used in knowledge-based systems design,
learning systems rely on algorithms to extract knowledge
or identify patterns in examples or data. Various statistics-based algorithms have been developed by management scientists and have been used extensively over the
past few decades for quantitative data analysis. These algorithms examine quantitative data for the purposes of
(Parsaye et al., 1989): (1) clustering descriptors with
common characteristics, for example, nearest neighbor
methods, factor analysis, and principal components
analysis; (2) hypothesis testing for differences among
different populations, for example, t-test and analysis of
variance (ANOVA); (3) trend analysis, for example, time
series analysis; and (4) correlation between variables, for
example, correlation coefficient, discriminant analysis,
and linear/multiple
regression analysis (Freund, 197 1;
Montgomery,
1976). These analysis techniques often
rely on complex mathematical models, stringent assumptions, or special underlying distributions.

The
findings are then presented in mathematical formulas
and parameters.
Learning Systems: An Overview. The symbolic machine learning technique, the resurgent neural networks
approach, and evolution-based genetic algorithms provide drastically different methods of data analysis and
knowledge discovery (Chen et al., in press; Fisher &
McKusik, 1989; Kitano, 1990; Mooney et al., 1989:
Weiss & Kapouleas, 1989: Weiss & Kulikowski, 199 1).
These techniques, which are diverse in their origins and
behaviors, have shown unique capabilities for analyzing

both qualitative, symbolic data and quantitative, numeric data. We provide below a brief overview of these
three classesof techniques, along with a representative
technique for each class.
l
198
Symboliclearningand 103: Symbolic machine learning techniques, which can be classified based on such
underlying learning strategiesas rote learning, learning
by being told, learning by analogy, learning from examples, and learning from discovery (Carbonell, Michalski, & Mitchell, 1983), have been studied extensively by AI researchers over the past two decades.
Among these techniques, learning from examples, a
special case of inductive learning, appears to be the
most promising symbolic machine learning technique
for knowledge discovery or data analysis. It induces a
general concept description that best describes the positive and negative examples. Examples of algorithms
which require both positive and negative examples are
Quinlans (1983) ID3 and Mitchells (1982) Version
Space. Some algorithms are batch-oriented, such as
JOURNAL
OF THE AMERICAN
Stepp and Michalskis CLUSTER/RD algorithm

(Stepp & Michalski, 1987) and ID3; but some are incremental, such as Utgoffs IDSR (Utgoff, 1989). Many
algorithms create a hierarchical arrangement of concepts for describing classesof objects, including Lebowitz UNIMEM (Lebowitz, 1987), Fishers COBWEB
(Fisher & McKusick, 1989) and Briemans CART
(Brieman et al., 1984). Most of the symbolic learning
algorithms produce production rules or concept hierarchies as outputs. These representations are easy to
understand and their implementation is typically
efficient (especially when compared with neural networks and genetic algorithms).
Among the numerous symbolic learning algorithms
which have been developed over the past 15 years,
Quinlans ID3 decision-tree building algorithm and its
descendants (Quinlan, 1983, 1986) are popular and
powerful algorithms for inductive learning. ID3 takes
objects of a known class. described in terms of a fixed
collection of properties or attributes, and produces a
decision tree incorporating these attributes that correctly classifiesall the given objects. It usesan information-economics approach aimed at minimizing the expected number of tests to classify an object. Its output
can be summarized in terms of IF-THEN rules.
9 Reuralnetworksand backpropagation:The foundation of the neural networks paradigm was laid in the
1950sand this approach has attracted significant attention in the past decade due to the development of more
powerful hardware and neural algorithms (Rumelhart,
Widrow, & Lehr, 1994). Nearly all connectionist algorithms have a strong learning component. In symbolic
machine learning, knowledge is represented in the
form of symbolic descriptions of the learned concepts,
for example, production rules or concept hierarchies.
In connectionist learning, on the other hand, knowledge is learned and remembered by a network of interconnected neurons, weighted synapses,and threshold
logic units (Lippmann, 1987; Rumelhart, Hinton, &
McClelland, 1986). Learning algorithms can be applied to adjust connection weights so that the network
can predict or classify unknown examples correctly.
Neural networks have been adopted in various engineering, business, military, and biomedical domains
(Chen et al., 1994; Lippmann, 1987; Simpson, 1990;
Widrow, Rumelhart, & Lehr, 1994). For example,
Hopfield networks have been used extensively in the
area ofglobal optimization and search (Hopfield, 1982;
Tank & Hopfield, 1987); Kohonen networks have been
adopted in unsupervised learning and pattern recognition (Kohonen, 1989). For a good overview of various
artificial neural systems, readers are referred to Lippmann (1987).
Among the numerous artificial neural networks
that have been proposed recently, backpropagation
networks have been extremely popular for their unique
learning capability (Widrow et al., 1994). Backpropagation networks (Rumelhart, 1986) are fully connected, layered, feed-forward models. Activations flow
from the input layer through the hidden layer, then to
the output layer. A backpropagation network typically
SCIENCE-April
1995
starts out with a random set of weights. The network

adjusts its weights each time it seesan input-output
pair. Each pair is processedat two stages,a forward
pass and a backward pass. The forward pass involves
presenting a sample input to the network and letting
activations flow until they reach the output layer. During the backward pass, the networks actual output is
compared with the target output and error estimates
are computed for the output units. The weights connected to the output units are adjusted to reduce the
errors (a gradient descentmethod). The error estimates
of the output units are then used to derive error estimates for the units in the hidden layer. Finally, errors
are propagatedback to the connections stemming from
the input units. The backpropagation network updates
its weights incrementally until the network stabilizes.
Simulated evolution and genetic algorithms., During
the past decade there has been a growing interest in
algorithms which rely on analogiesto natural processes
and Darwinian survival ofthe fittest. The emergenceof
massively parallel computers made these algorithms of
practical interest. There are currently three main avenues of research in simulated evolution: genetic algorithms; evolution strategies; and evolutionary programming (Fogel, 1994). Each method emphasizes a
different facet of natural evolution. Genetic algorithms
stresschromosomal operations such as crossover and
mutation (Booker, Goldberg, & Holland, 1990;
Holland, 1975). Evolution strategies emphasize individual behavioral changes.Evolutionary programming
stressesbehavioral changes at the level of the species
(Fogel, 1962, 1964). Fogel ( 1994) also provides an excellent review of the history and recent efforts in this
area. Among these methods, genetic algorithms have
been used successfully for various optimization problems in engineering and biomedical domains.
Genetic algorithms were developed based on the
principle of genetics(Goldberg, 1989; Koza, 1992; Michalewicz, 1992). In such algorithms a population of
individuals (potential solutions) undergoesa sequence
of unary (mutation) and higher order (crossover)transformations. These individuals strive for survival: a selection (reproduction) scheme,biased toward selecting
fitter individuals, produces the individuals for the next
generation. After some number ofgenerations the program converges-the best individual representsthe optimum solution.
analysis methods, backpropagation net, and decisiontree-based inductive learning methods (ID3-like) were
found to achieve comparable performance for several
data sets. Fisher and McKusick (1989) found that using
batch learning, backpropagation performed as well as
ID3, but it was more noise-resistant. They also compared
the effect of incremental learning versus batch learning.
Kitano ( 1990) performed systematic, empirical studies
on the speed of convergence of backpropagation networks and genetic algorithms. The results indicated that
genetic search is, at best, equally efficient as faster variants of a backpropagation algorithm in very small scale
networks, but far less efficient in larger networks. Earlier
research by Montana and Davis (1989), however,
showed that using some domain-specific genetic operators to train the backpropagation network, instead of using the conventional backpropagation delta learning
rule, improved performance. Harp, Samad, and Guha
( 1989) also achieved good results by using GAS for neural
network design.
Systems developed by Kitano ( 1990) and Harp et al.
( 1989) are also considered hybrid systems (genetic algorithms and neural networks), as are systems like COGIN
(Green & Smith, 1991) which performed symbolic induction using genetic algorithms and SC-net (Hall & Romaniuk, 1990), which is a fuzzy connectionist expert system. Other hybrid systems developed in recent years employ symbolic and neural net characteristics. For
example, Touretzky and Hinton (1988) and Gallant
( 1988) proposed connectionist production systems, and
Derthick (1988) and Shastri (199 1) developed different
connectionist semantic networks.
Learning Systems in IR. The adaptive learning
techniques cited have also drawn attention from researchers in information science in recent years. In particular, Doszkocs, Reggia, & Lin ( 1990) provided an excellent review of connectionist models for information
retrieval and Lewis ( 199 1) has briefly surveyed previous
research on machine learning in information retrieval
and discussed promising areas for future research at the
intersection of these two fields.
l
Over the past years there have been several studies

which compared the performance of these techniques for
different applications as well as some systems which used
hybrid representations and learning techniques. We
summarize some of these studies below.
Mooney et al. (1985) found that ID3 was faster than a
backpropagation net, but the backpropagation net was
more adaptive to noisy data sets. The performances of
these two techniques were comparable, however. Weiss
and Kapouleas (1989, 1991) suggested using a resampling technique, such as leave-one-out for evaluation, instead of using a hold-out testing data set. Discriminant
Neural networks and IR: Neural networks computing,

in particular, seems to fit well with conventional retrieval models such as the vector spacemodel (Salton,
1989) and the probabilistic model (Maron & Kuhns,
1960). Doszkocs et al. (1990) provided an excellent
overview of the use of connectionist models in information retrieval. These models include severalrelated
information processing approaches, such as artificial
neural networks, spreadingactivation models, associative networks, and parallel distributed processing. In
contrast to more conventional information processing
models, connectionist models are self-processingin
that no external program operateson the network: the
network literally processesitself, with intelligent be-
JOURNAL OF THE AMERICAN SOCIETY FOR INFORMATION
SCIENCE-April
1995
199
havior emerging from the local interactions that occur concurrently between the numerous network
nodes through their synaptic connections. By taking a
broader definition of connectionist models, these authors were able to discussthe well-known vector space
model, cosine measures of similarity, and automatic
clustering and thesaurus in the context of netH!orkrepresentation. Based on the network representation,
spreading activation methods such as constrained
spreading activation adopted in GRANT (Cohen &
Kjeldsen, 1987) and the branch-and-bound algorithm
adopted in METACAT (Chen & Dhar, 1991) can be
considered as variants of connectionist activation.
However, only a few systems are considered classical
connectionist systems that typically consist of
weighted, unlabeled links and exhibit some adaptive
learning capabilities.
The work of Belew is probably the earliest connectionist model adopted in IR. In AIR (Belew, 1989). he
developed a three-layer neural network of authors. index terms, and documents. The system used relevance
feedback from its users to change its representation of
authors, index terms, and documents over time. The
result was a representation of the consensual meaning
of keywords and documents shared by some group of
users. One of his major contributions was the use of a
modified correlational learning rule. The learning process created many new connections between documents and index terms. Rose and Belew (199 1) extended AIR to a hybrid connectionist and symbolic
system called SCALIR which used analogical reasoning to find relevant documents for legal research.Kwok
( 1989) also developed a similar three-layer network of
queries, index terms, and documents. A modified Hebbian learning rule was usedto reformulate probabilistic
information retrieval. Wilkinson and Hingston ( 1991,
1992) incorporated the vector space model in a neural
network for document retrieval. Their network also
consisted of three layers: queries, terms, and documents. They have shown that spreading activation
through related terms can help improve retrieval performance.
While the above systems represent information retrieval applications in terms of their main components
ofdocuments. queries, index terms, authors, etc., other
researchers used different neural networks for more
specific tasks. Lin, Soergel, & Marchionini (199 1)
adopted a Kohonen network for information retrieval.
Kohonens feature map, which produced a two-dimensional grid representation for N-dimensional features,
was applied to construct a self-organizing (unsupervised learning), visual representation of the semantic
relationships between input documents. In MacLeod
and Robertson (199 I), a neural algorithm was used for
document clustering. The algorithm compared favorably with conventional hierarchical clustering algorithms. Chen et al. (1992, 1993, in press) reported a
seriesof experiments and system developments which
generated an automatically created weighted network
of keywords from large textual databases and integrated it with several existing man-made thesauri (e.g.,
200
JOURNAL
OF THE AMERICAN
LCSH). Instead of using a three-layer design, Chens

systems developed a single-layer, interconnected,
weighted/labeled network of keywords (concepts) for
concept-based information retrieval. A blackboardbaseddesign which supported browsing and automatic
concept exploration using the Hopfield neural networks parallel relaxation method was adopted to facilitate the usage of severalthesauri (Chen et al., 1993).In
Chen, Basu, and Ng (in press-a),the performance of a
branch-and-bound serial search algorithm was compared with that of the parallel Hopfield network activation in a hybrid neural-semantic network (one neural network and two semantic networks). Both methods achieved similar performance, but the Hopfield
activation method appearedto activate concepts from
different networks more evenly.
Symbolic learning and IR: Despite the popularity of
using neural networks for information retrieval, we see
only limited use of symbolic learning techniques for
IR. In Blosseville et al. ( 1992).the researchersuseddiscriminant analysis and a simple symbolic learning
technique for automatic text classification. Their symbolic learning process representedthe numeric classification results in terms of IF-THEN rules. Text classification involves the task of classifying documents
with respect to a set of two or more predefined classes
(Lewis, 1992). A number of systems were built based
on human categorization rules (a knowledge-basedsystem approach) (Rau & Jacobs, 1991). However, a
range of statistical techniques including probabilistic
models. factor analysis, regression,and nearest neighbor methods have been adopted (Blosseville et al.,
1992: Lewis, 1992: Masand, Gordon, & Waltz, 1992).
Fuhr et al. (1990) adopted regressions methods and
ID3 for their feature-basedautomatic indexing technique. Crawford, Fung, and their coworkers (Crawford
et al., 199I; Crawford & Fung, 1992: Fung & Crawford, 1990) have developed a probabilistic induction
technique called CONSTRUCTOR and have compared it with the popular CART algorithm (Breiman et
al., 1984). Their experiment showed that CONSTRUCTORs output is more interpretable than that
produced by CART, but CART can be applied to more
situations (e.g., real-valued training sets). Chen and
She ( 1994) adopted ID3 and the incremental IDSR algorithm for information retrieval. Both algorithms
were able to use user-supplied samples of desired documents to construct decision trees of important keywords which could represent the usersqueries. For a
test collection of about 1000 documents, both symbolic learning algorithms did a good job in identifying
the concepts (keywords) which best represent the set
of documents identified by users as relevant (positive
examples) and irrelevant (negative examples). More
testing, however, is underway to determine the
effectivenessof example-baseddocument retrieval using ID3 and IDSR.
Several recent works which involved using symbolic learning techniques in the related databaseareas
were also identified, especially in relational database
management systems (RDBMS). Cai, Cercone, and
SCIENCE-April
1995
Han ( 1991) and Han, Cai, and Cercone (1993) developed an attribute-oriented, tree-ascendingmethod for
extracting characteristic and classification rules from
relational databases.The technique relied on some existing conceptual tree for identifying higher-level, abstract concepts in the attributes. Ioannidis, Saulys, and
Whitsitt (1992) examined the idea of incorporating
machine learning algorithms (UNIMEM and COBWEB) into a database system for monitoring the
stream of incoming queries and generating hierarchies
with the most important concepts expressedin those
queries. The goal is for these hierarchies to provide
valuable input for dynamically modifying the physical
and logical designsof a database.Also related to database design, Borgida and Williamson (1985) proposed
the use of machine learning to represent exceptions in
databasesthat are based on semantic data models. Li
and McLeod ( 1989) used machine learning techniques
to handle object flavor evolution in object-oriented databases.
Genetic algorithms and IR: Our literature search revealed several implementations of genetic algorithms
in information retrieval. Gordon (1988) presented a
genetic algorithms-based approach for document indexing. Competing document descriptions (keywords)
are associatedwith a document and altered over time
by using genetic mutation and crossover operators. In
his design, a keyword representsa gene (a bit pattern),
a documents list of keywords representsindividuals (a
bit string), and a collection of documents initially
judged relevant by a user representsthe initial population. Based on a Jaccards score matching function
(fitness measure), the initial population evolved
through generations and eventually converged to an
optimal (improved) population-a set of keywords
which best described the documents. Gordon (1991)
further adopted a similar approach to document clustering. His experiment showed that after genetically redescribing the subject description of documents, descriptions of documents found co-relevant to a set of
queries will bunch together. Redescription improved
the relative density of co-relevant documents by
39.74% after 20 generations and 56.6 1% after 40 generations. Raghavanand Agarwal ( 1987) have also studied the genetic algorithms in connection with document clustering. Petry et al. (1993) applied genetic programming to a weighted information retrieval system.
In their research,a weighted Boolean query was modified to improve recall and precision. They found that
the form of the fitness function has a significant effect
upon performance. Yang and coworkers (Yang & Korlhage, 1993; Yang, Korlhage, & Rasmussen, 1993)
have developed adaptive retrieval methods based on
genetic algorithms and the vector space model using
relevance feedback. They reported the effect of adopting genetic algorithms in large databases,the impact of
genetic operators, and GAs parallel searchingcapability. Frieder and Siegelmann ( 1991) also reported a data
placement strategy for parallel information retrieval
systemsusing a genetic algorithms approach. Their results compared favorably with pseudo-optimal docu-
ment allocations. In Chen and Kim ( 1993), a GA-NN

hybrid system, called GANNET, was developedfor IR.
The system performed concept optimization for userselected documents using genetic algorithms. It then
used the optimized concepts to perform conceptexploration in a large network of related concepts through
the Hopfield net parallel relaxation procedure. A Jaccards score was also adopted to compute the fitness
of subject descriptions for information retrieval.
Following this overview, we present three sample implementations of neural networks, symbolic learning,
and genetic algorithms, respectively, for illustration
purposes. We hope that examining these implementations in the context of IR will encourage other researchers to appreciate these techniques and adopt them in
their own research.
Neural Networks for II?
Neural networks provide a convenient knowledge
representation for IR applications in which nodes typically represent IR objects such as keywords, authors, and
citations and bidirectional links represent their weighted
associations (of relevance). The learning property of
backpropagation networks and the parallel search property of the Hopfield network provide effective means for
identifying relevant information items in databases.
Variants of the backpropagation learning in IR can be
found elsewhere (Belew, 1989; Kwok, 1989). In this section, we review a Hopfield network implementation and
its associated parallel search property.
A Hopjield Network: KnowledgeRepresentationand

Procedure
The Hopfield net (Hopfield, 1982; Tank & Hopfield,
1987) was introduced as a neural net that can be used as
a content-addressable memory. Knowledge and information can be stored in single-layered interconnected
neurons (nodes) and weighted synapses (links) and can
be retrieved based on the networks parallel relaxation
method-nodes
are activated in parallel and are traversed until the network reaches a stable state (convergence). It had been used for various classification
tasks and global optimization (Lippmann, 1987; Simpson, 1990).
A variant of the Hopfield network for creating a network of related keywords developed by Chen (Chen &
Lynch, 1992; Chen et al., 1993) used an asymmetric similarity function to produce thesauri (or knowledge bases)
for different domain-specific databases. These automatic
thesauri were then integrated with some existing manually created thesauri for assisting concept exploration
and query refinement. A variant of the Hopfield parallel
relaxation procedure for network search (Chen et al.,
SCIENCE-April
1995
201
bias and B0is used to modify the shape of the SIG-
1993) and concept clustering (Chen et al., in press-b)had

been reported earlier.
The implementation reported below incorporated the
basic Hopfield net iteration and convergence ideas.
However, significant modification was also made to accommodate unique characteristics of information retrieval; for example, asymmetric link weights and the
continuous SIGMOID transformation function. With
the initial search terms provided by searchersand the association of keywords captured by the network, the Hopfield parallel relaxation algorithm activates neighboring
terms, combines weighted links, performs a transformation function (a SIGMOID function,f,), and determines
the outputs of newly activated nodes. The process repeats until node outputs remain unchanged with further
iterations. The node outputs then represent the concepts
that are strongly related to the initial search terms. A
sketch of the Hopfield net activation algorithm follows:
MOID function.
This formula shows the parallel relaxation property of the Hopfield net. At each iteration, all nodes
are activated at the same time. The weight computation scheme, net, = X7& &F,(t), is a unique characteristic of the Hopfield net algorithm. Based on parallel activation, each newly activated node derives its
new weight based on the summation of the products
of the weights assignedto its neighbors and their synapses.
(4) Convergence: The above process is repeated until
there is no change in terms of output between two
iterations, which is accomplished by checking:
n-1
c lP,(t+ l)-P,Cc,(Q 56
,=o
where 6 is the maximal allowable error (a small number). The final output represents the set of terms relevant to the starting keywords. Some default threshold values were selected for (0,, 19,).
(1) Assigning synaptic weights: For thesauri which were
generatedautomaticallyusing a similarity function

(e.g., the COSINE function) (Everitt. 1980), the re-
sultinglinks representprobabilistic,synapticweights
between any two concepts. For other external thesauri which contain only symbolic links (e.g., narrower term, synonymous term, broader term, etc.),
a user-guided procedure of assigning a probabilistic
weight to each symbolic link can be adopted (Chen
et al., 1993).
The training phase of the Hopfield net is completed when the weights have been computed or assigned. to represents the synaptic weight from
node i to nodej.
(2) Initialization with search terms: An initial set of
search terms is provided by searchers, which serves
as the input pattern. Each node in the network which
matches the search terms is initialized (at time 0) to
have a weight of 1.
p,(O)=x,,Osisn-
p,(t) is the output of node i at time t and xl, which

has a value between 0 and 1, indicates the input pattern for node i.
(3) Activation, weight computation, and iteration.
n-l
p,(t + 1) =f; [ C tb@i(t)],0 5j 5 n - 1
i=O
A Hopjield hTetwork Example
A sample sessionof the Hopfield net spreadingactivation is presented below. Three thesauri were incorporated in the experiment: a Public thesaurus (generated
automatically from 3000 articles extracted from DIALOG), the ACM Computing Review Classification System (ACM CRCS), and a portion of the Library of Congress Subject Headings (LCSH) in the computing area.
The links in the ACM CRCS and in the LCSH were assigned weights between 0 and 1. Several user subjects
(MIS graduate students) were also asked to reviewed selected articles and create their own folders for topics of
special interest to them. Notice that some keywords were
folder names assignedby the users (in the format of *.*);
for example, QUERY.OPT folder for query optimization topics; DBMS.AI folder for artificial intelligence
and databases topics; and KEVIN.HOT folder for
HOT (current) topics selectedby a user, Kevin. In the
example shown below, the searcherwas asked to identify
descriptors which were relevant to knowledge indexed
deductive search.The initial searchterms were: information retrieval, knowledge base, thesaurus, and
automatic indexing (as shown in the following interaction).
wheref, is the continuous SIGMOID transformation

function as shown below (Dalton & Deshmane,
199 1; Knight, 1990)
.t;(nq) =
1 + exp[ -l(neiO- 1
where n&j = Cy:d &p,(t), 0, serves as a threshold or
202
JOURNAL
OF THE AMERICAN
*--.--....-....--...*
Initialterms:
{*Suppliedbythesubject.
_-_---_-----1. (P L) INFORMATIONRETRIEVAL{*P:Public,
ACM,L:LCSH*)
)KNOWLEDGEBASE
2. (P
3. (P
)THESAURUS
4. (P L)AUTOMATICINDEXING
*.-...--..----------*
SCIENCE-April
1995
*)
A:
TABLE I.
Iteration
no.
0
2
3
4
Sample Hopfield net iterations.

Activations
Suggestedterms
INFORMATIONRETRIEVAL
KNOWLEDGEBASE
THESAURUS
AUTOMATICINDEXING
INDEXING
KEVIN.HOT
CLASSIFICATION
EXPERTSYSTEMS
ROSS.HOT
RECALL
INFORMATIONRETRIEVALSYSTEM
EVALUATION
SELLING-INFORMATION
STORAGEANDRETRIEVAL
SYSTEMS
1.00
1.00
1.00
1.00
0.65
0.56
0.50
0.50
0.44
0.50
0.26
0.15
..
Enter the numberof

system-suggestedterms
or '0'
toquit>>lO
{* The users suppliedatargetnumber
of relevant
terms.*)
Given these starting terms, the Hopfield net iterated

and converged after 11 iterations. The activated terms
after the first four iterations and their associatedlevels of
activation are shown in Table 1. Due to the damping
effect of the parallel search property (i.e., the farther
away from the initial searchterms, the weaker the activation), terms activated at later iterations had lower activation values and were less relevant to the initial search
terms in general.Fourteen terms were suggestedafter the
complete Hopfield net activation. Searchers could
browsethe system-suggested
list, selectterms of interest,
and then activate the Hopfield net again. The user-system interaction continued until the user decided to stop.
{* The system reported14
relevant
terms as shown
below. *}
1. ( )INDEXING
2. ( ) SELLING - INFORMATION STORAGE AND RETRIEVALSYSTEMS
3. ( )KEVIN.HOT
4. ( ) INFORMATIONRETRIEVALSYSTEMEVALUATION
5. ( )RECALL
6. ( )EXPERTSYSTEMS
7. ( )CLASSIFICATION
8. ( ) DBMS.AI
9. ( )ROSS.HOT
10. ( ) INFORMATION STORAGE AND RETRIEVAL SYSTEMS
11. ( ) INFORMATIONRETRIEVAL
12. ( )KNOWLEDGEBASE
13. ( )THESAURUS
14. ( )AUTOMATICINDEXING
Enter numbers [lto14]

or '0' toquit:
7,10-14
{*Theuserselectedtermshedeemedrelevant.
The system confirmed
the selections
displaythesource
foreachterm.
*}
1. (P
2.c
3.
(P
4.(P
5. (P
6-C
7.
8.
9.
10.
(P
(P
(P
(P
1, 2, 4, 5,
made and
1 INDEXING
L) SELLING - INFORMATION STORAGEANDRETRIEVALSYSTEMS
) INFORMATION RETRIEVAL SYSTEM EVALUATION
) RECALL
) CLASSIFICATION
L) INFORMATIONSTORAGEANDRETRIEVALSYSTEMS
L)INFORMATIONRETRIEVAL
)KNOWLEDGEBASE
)THESAURUS
L)AUTOMATICINDEXING
Enterthenumberof
system-suggestedtermsor
'0'
toquit>>
{* The uses decide to broaden the search by requesting
the Hopfield
network
to identify
30
newtermsbasedonthetermshehadselected.
*}
........
Enter number
35,36,38
........
[lto40]
or '0'
toquit:
3-7,
9, 33,
Enternumbers
[lto67]or
'0'toquite:O
{*Thesystemlistedhisfinalselections.*}
1.
2.
3.
4.
5.
7.
8.
9.
(P
(P
(P
(P
(P
(
(
(
(P
10.
11.
(P
(
12.
(P
6.
)PRECISION
L) INFORMATIONRETRIEVAL
)INDEXING
L)AUTOMATICINDEXING
)RECALL
L)AUTOMATICABSTRACTING
L)AUTOMATICCLASSIFICATION
L)AUTOMATICINFORMATIONRETRIEVAL
) INFORMATION RETRIEVAL SYSTEM EVALUATION
)THESAURUS
L)INFORMATIONSTORAGEANDRETRIEVALSYSTEMS
)KNOWLEDGEBASE
{* A total
of 12 terms were selected.
weresuggestedbytheHopfieldnetalgorithm.
Eight
terms
*}
In a more structured benchmark experiment, we

tested 30 sample queries using the Hopfield algorithm
in an attempt to understand the generalbehavior of the
algorithm. We tested five caseseach for queries with 1
term, 2 terms, 3 terms, 4 terms, 5 terms, and 10 terms, a
total of 30 cases.A few examplesof the queriesused,all
in the computing area, were: (1 term: Natural Language
Processing);(2 terms: Group Decision Support Systems,
Collaboration); (3 terms: SystemsAnalysis and Design,
Simulation and Modeling, Optimization); etc.
JOURNAL OF THE AMERICAN SOCIETY FOR INFORMATION SCIENCE-April
1995
203
TABLE 2. Results of Hoplield

Case
1
2
3
4
5
6
7
8
9
IO
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
Average
No. of
terms
2
2
2
2
2
3
3
3
3
3
4
4
4
4
4
5
5
5
5
5
IO
IO
10
10
10
5
networktesting
Query terms in
(P, A> L)
Suggestedterms in
NN: (P. A, L)
(1, 131)
(l,O, 1)
(1, 1,1)
(0.0,1)
(I,% 1)
(2, 1,O)
(12, I. 7)
(5,0, 16)
(ll,5. II)
(0, 0, 20)
(4,4, 19)
(19,2.3)
(LO, 2)
(LO, 0)
Cl, J)
(2. 1,2)
(16,0,8)
(330, 1)
(I. 2, 1)
(2, I,31
(1.3, 1)
(1.&a
(2.2,4)
(3,2,2)
(2. 3,2)
(1.3.4)
(I, 2, 1)
(1.4, 1)
(4.2,2)
(3,234)
(j,O, 1)
(5.0, 1)
t&O, 3)
(10, I,31
(8.0,4)
(9. l,5)
(8.2,3)
(3.1, 1.2, 1.9)
18
I5
14
11
17
21
19
20
15
27
19
22
18
16
9
17
19
24
19
18
16
15
11
23
12
17
25
24
21
28
18.8
(20,3,4)
(Il,5. II)
(11,O. 12)
(20,O. 18)
(4, 11,8)
(22, 1,s)
(20,2,2)
(l3,9,3)
(l7,4,4)
(ll,2,13)
(18,5.6)
(1&2,5)
(15, 8. 3)
(19,436)
(IO, 1. 12)
(2,O. 18)
(19,0.3)
(20,O. I)
(I l,O. 13)
(1%2, 10)
(16,0.8)
(19, 1.6)
(20, 2. 3)
(14.5,2.5,8.5)
For each query, we selected terms from different

knowledge sources,P for the Public KB, A for the
ACM CRCS, and L for the LCSH, as shown in Table
2. Some terms may have appeared in more than one
knowledge source. The three knowledge sources contained about 14,000 terms and 80,000 weighted links.
The results shown in Table 2 reveal the number of iterations, the computing times, and the sources of knowledgefor the query terms and the system-suggested
terms,
The reasonfor investigating the source of knowledge for
system-suggested
terms was to show the extent to which
the Hopfield algorithm branchedout and utilized knowledgefrom various knowledgesources.
Despite the variation in the number of starting terms,
the responsetimes increasedonly slightly when the number of starting terms was increased.The averageresponse
time was 24.5 secondsafter about an averageof about 19
iterations by the Hopfield network. The reason for this
was that the Hopfield net thresholds (0, and 0,) helped
prune the searchspace.However, more stringent thresholds may need to be adopted to achievereasonablerealtime responsefor large databases.
204
No. of iterations
NN
Times (seconds)
NN
21
14
18
10
26
18
22
24
16
29
31
34
29
23
10
II
31
33
32
6
27
27
23
33
30
34
32
36
25
31
24.5
Another important observation was that the Hopfield

net appearedto invoke the different knowledge sources
quite evenly. As shown in Table 2, for most queries the
Hopfield net (NN) almost always produced terms from
all three knowledgesources.Most terms suggestedby the
algorithm appeared relevant and many of them were
multiple links away from the initial search terms (conventional Hypertext browsing doesnot traversemultiple
links effectively). However, detailed user studies need to
be performed to examine the usefulnessof the algorithm
in search,especiallyfor large-scaleapplications.
Symbolic Learning for IR
Even though symbolic learning techniqueshave been

adopted frequently in various database,engineering,and
businessdomains, we seeonly lim ited use of such techniques in IR. For illustration purposes,we summarize
below a symbolic learning for IR implementation based
on the ID3 and IDSR algorithms (Chen & She, 1994).
1995
ID3/ID5R:
Knowledge Representation and Procedure
ID3 is a decision-tree building algorithm developed

by Quinlan (1979, 1983).It adoptsa divide-and-conquer
strategy for object classification. Its goal is to classify
mixed objects into their associatedclassesbasedthe objectsattribute values. In a decision tree, one can classify
a node as:
l
l
a leaf node that contains a classname; or

a non-leaf node (or decision node) that contains an attribute test.
Each training instance or object is representedas a list

of attribute-value pairs, which constitutes a conjunctive
description of that instance. The instance is labeledwith
the name of the class to which it belongs. Using the divide-and-conquer strategy, ID3 picks an attribute and
uses it to classify a list of objects based on their values
associatedwith this attribute. The subclasseswhich are
created by this division procedure are then further divided by picking other attributes. This processcontinues
until each subclassproduced contains only a single type
of object. To produce the simplest decision tree (a minimal tree) for classification purpose,ID3 adopts an information-theoretic approach which aims at minimizing
the expected number of tests to classify an object. An
entropy (a measure of uncertainty) concept is used to
help decide which attribute should be selectednext. In
general,an attribute which can help put objects in their
proper classestends to reduce more entropy and thus
should be selectedas a test node.
In IR, we can assumethat there exists a database(universe) of records (documents, tables, etc.). Records are
describedby attributes (keywords, primary keys, fields).
Each record in the databasethen belongsto only one of
two possibleclasses:
l
the positive class (+): consisting of records that are

desired; and
the negative class (-): consisting of records that are
undesired.
Different databaseusers may desire different sets of

documents due to their unique information needs,and
the set of documents desired by one user often constitutes only a small portion of the entire database.Enabling the system to identify this small set of positive
documents is therefore a challengingtask.
In our implementation, we maintained a list of all the
keywords that existed in the desireddocuments and used
this list to decide what attributes were crucial to describing documents in the positive class.The test at each nonleaf node of the decision tree determined the presenceor
absenceof a particular keyword: yes meant that the
test keyword existed in a document, and no meant
that the keyword did not exist in a document. Thus, ID3
created a binary classification tree. A sketch of the ID3

algorithm adopted follows:
(1) Compute entropyfor mixed classes:Initially searchers were requested to provide a set of positive and
negative documents. This set of documents servedas
the training examplesfor the ID3 algorithm. Entropy
was calculated by using the following function
(Quinlan, 1983):
entwv
= -pPo.%b
ppos - pneJog pneg
where ppOiand pnenrepresentedthe proportions of the

documents that were positive or negative, respectively.
(2) Select the best attribute based on entropy reduction.
For each untested attribute (keyword), the algorithm
computed an entropy value for its use when classifying mixed documents. Each branch of the decision
tree represented the existence or nonexistence of a
particular keyword. The keyword that reduced the
entropy most servedas the next decision node in the
tree. As a greedy algorithm, ID3 always aims at
maximizing local entropy reduction and never backtracks.
(3) Iterate until all documents are classified: Repeating
steps( 1) and (2) ID3 computed the entropy value of
each mixed class and identified the best attribute for
further classifying the class.The processwas continued until each class contained either all positive or
all negative documents.
Consideredas an incremental version of the ID3 algorithm, IDSR, developedby Utgoff (1989), is guaranteed
to build the same decision tree as ID3 for a given set of
training instances(Quinlan, 1993). In IDSR, a non-leaf
node contains an attribute test (sameas in ID3) and a set
of other non-test attributes, each with object counts for
the possiblevalues of the attribute. This additional nontest attribute and object count information at each noleaf node allows IDSR to update a decision tree without
rebuilding the entire tree. During the tree rebuilding process,an old test node may be replacedby a new attribute
or swappedwith other positions in the tree. As in ID3,
the tree-building process requires much less computation and time than other inductive learning methods, including neural networks and geneticalgorithms.
To create a robust and real-time inductive learning
system,a relevancefeedback schemewas introduced into
our implementation. Although the proposed inductive
learning algorithms require usersto provide examplesto
confirm their interests,it is inconceivable that userswill
be able to browse the entire databaseto identify such instances. An incremental, interactive feedback process,
therefore, was designedto allow usersto examine a few
documents at a time. In essence,our IDSR algorithm
was implemented such that it provided a few suggested
documents basedon the documents initially provided by
SCIENCE-April
1995
205
a user after examining a small portion of the database.
When a predetermined number of desired documents

had been found (say three, in our implementation), the
system presentedthese documents to the user immediately for evaluation (as desired or undesired).This iterative system-induction and user-feedbackprocesscontinued until the user decided to stop or the complete databasehad been traversed.
During the relevancefeedbackprocess,the newly confirmed documents, either desired or undesired, could be
used by IDSR to update the decision tree it previously

had constructed. It was shown that when more examples
are provided by the users and when the database is more
exhaustively searched,IDSR can significantly improve
its classification accuracy and searchperformance.
006
008
083
084
thesaurus,remote use, keyword (+)

retrieval, interface (+)
syntax checking, remote use, test, user (-)
interface, protocol, standardization (-)
Providing negativedocuments was optional. If a user

could not think of an example of a document which was
undesired,the systemby default automatically generated
one negative document which contained no keyword
identical to any that was present in the desired set. The
initial positive keyword list then consisted of all keywords from desired documents; that is, thesaurus, re-
mote use, keyword, retrieval, interface (in that order).

Therefore, the set of initial training instances can be rep-
resentedas:
Initial Training Instances
An ID3/ID5R Example
We created a small test databaseof 60 records. For
evaluation purposes,we were able to manually select a
small set of target desired documents (i.e., eight documents in the areas of information retrieval and keywording). The goal of the experiment was to present a
few documents at a time to our system and seewhether
the system would be able to identify them after the iterative relevance feedback process. The performance of our
IDSR-based system was also compared with that of the

more conventional ID3 algorithm, which used only an
initial set of desired documents to generatea query tree.
Sample entries in the literature databaseare shown below, where the first column representsthe document
number, and the remaining columns representdifferent
numbers of keywords (two to five) associatedwith the
document.
Y
n
n
n
Y
n
n
n
Y
n
Y
n
n
Y
n
n
n
Y
n
Y
(+I
(+I
t-1
(-)
If a document contained a particular keyword in the

keyword list, its attribute value was labeled y (yes),
otherwise the value was n (no).
Based on the set of
training instances,ID3 first computed the entropy value

when adopting thesaurus (the first keyword obtained
from the desired documents). It then computed the entropy values when adopting other positive keywords.
The thesauruskeyword producedthe most entropy reduction and was thus selected as the first decision node.
Following the same computation, retrieval was se-
lected as the next (and last) decision node. ID3 constructed the decision tree shown in Figure 1. In the figure,
for example, [2, l] means2 instanceswere in the negative
010
013
014
018
021
022
023
030
031
048
049
050
generic, keyword, reference

modeling, thesaurus, terrorism
modeling, simulation, thesaurus,terrorism
keyword, thesaurus
ID3, AI, NN
file, keyword
hierarchy, interface, index
carat, AI, expert, keyword, thesaurus
AI, protocol, thesaurus
keyword, retrieval
cross-reference,remote use, redundancy
expectations, market, maintenance, quel, interface
io7
149
152
177
iT, computerized, MIS

database,query, keyword
sort, indexing, merge, keyword
country, code, keyword, IS0
class and 1 instance was in the positive class. The deci-
Initially the user was able to identify the following

documents as desired (+) or undesired (-), respectively
(documents which the user had seenbefore):
206
JOURNAL
OF THE AMERICAN SOCIETY FOR INFORMATION
FIG. I. Initial treecreatedfor an IR example.
SCIENCE-April
1995
sion tree in Figure 1 can be represented as production

rules: ( 1) IF a document has thesaurus as a keyword
THEN it is desired (one +, the rightmost branch); (2) IF
a document does not have thesaurus as a keyword, but
has retrieval THEN it is also a desired document (one
+, the middle branch); (3) IF a document does not have
thesaurus or retrieval as a keyword THEN it is an
undesired document (two-, the leftmost branch).
Based on this decision tree, the system searched the
database for similar documents and identified three
more documents as presentedbelow:
013 modeling,thesaurus,terrorism(+)
014 modeling,simulation,thesaurus,terrorism(+)
018 keyword,thesaurus(+)
These documents were then presented to the user,

who provided feedback as to whether or not they were
desired. If the user confirmed that document 0 18 was desired but rejected documents 0 13 and 0 14, IDSR used
the new (contradictory) evidence to update its current
tree. The new training instances for IDSR were:
New Training instances
Y
Y
Y
n
n
n
n
Y
n
n
n
n
n
n
t-1
t-1
(+I
The system produced a new tree as shown in Figure 2.

This new tree looked different from the original one and
can be summarized by the following rules: (1) IF a document has keyword as a keyword THEN it is desired
(two +, the rightmost branch); (2) IF a document does
not have keyword as a keyword, but has retrieval
THEN it is also a desired document (one +, the middle
branch); (3) IF a document does not have keyword or
retrieval as a keyword THEN it is an undesired document (four -, the leftmost branch). The whole process
was repeated until the entire databasewas traversed. For
this particular example, the final decision tree was the
same as the one shown in Figure 2.
To determine how IDSR performed during the user
relevancefeedback processwe examined its recall at each
point of relevance feedback and compared its performance with that of ID3. ID3 used only the initial document feedback from the usersto construct a decision tree
and used the tree to search the database. IDSR, on the
other hand, collected new evidence during each iteration
and updated its trees accordingly. The recall measure
was defined as:
Recall =
We developed a test database of about 1000 documents from the 1992 COMPENDEX CD-ROM collection of computing literature. We then identified 10 research topics, each of which had between 5 and 20 relevant documents in the database (manually identified).
The testing was conducted by comparing the recall of the
ID3 algorithm and that of the IDSR incremental approach using the 10 researchtopics.
Detailed results of the experiment are presentedin Table 3. IDSR and ID3 achieved the same levels of performance for 5 of the 10 test cases(cases3 and 6-9). After
we examined these casescarefully, we found that the initial documents presentedfor these caseshad very precise
keywords assignedto them. New instancesprovided during relevance feedback were consistent with the initial
documents, thus IDSR did not revise its decision tree.
(At each interaction, IDSR searchedonly a portion of the
entire database.The trees constructed by ID3 remained
constant becauseID3 did not have any interaction with
its users. However, to compare its results with those of
the IDSR fairly, ID3s performance at each interaction
was computed based on the same documents visited by
IDSR. As more documents were examined, ID3s classification results may also have improved.)
For the other five test cases,IDSRs performance increasedgradually until it reached 93.1%. ID3 had been
able to reach 74.9%. These researchtopics tended to have
more diverse keywords in the initial documents provided. IDSR appearedto benefit from incremental query
tree revision based on the relevance feedback information provided by users. In all 10 cases,IDSR was able to
terminate in eight interactions. The responsetimes were
often less than a second for each decision-tree building
process.
In conclusion, the symbolic ID3 algorithm and its
IDSR variant both were shown to be promising techniques for inductive document retrieval. By using the en-
Number of relevant records retrieved

Total number of relevant records in database
JOURNAL
OF THE AMERICAN
d4,
FIG. 2. Updated tree after relevance feedback.
SCIENCE-April
1995
207
TABLE 3. Results of ID3 and ID5R testing.

Case
2
3
4
5
6
7
8
9
10
Avg. hits
Avg. recall
Int. 1
ID3/IDSR
Int. 2
ID3/ID5R
l/1
o/o
l/l
l/l
o/o
l/l
l/l
112
O/l
212
l/l
O/l
212
212
212
313
717
5/5
l/l
1.3/1.3
16.0/16.0
212
212.3
16.5/31.2
Int. 3
ID3/ID5R
Int. 4
ID3/IDSR
Int. 5
ID3/ID5R
213
516
l/4
4/4
l/3
315
619
012
313
11-2
o/2
515
313
313
w3
313
2.813.4
35.0/40. I
115
217
218
214
517
Int. 8
ID3/ID5R
717
IO/IO
717
5.1j6.2
66.3179.3
9/O
414
4.415.2
55.5164.1
for IR
A GeneticAlgorithm: KnowledgeRepresentation
and Procedure
Genetic algorithms (GAS) (Goldberg, 1989; Kohonen, 1989; Michalewicz, 1992) are problem-solving systems based on principles of evolution and heredity. A
GA maintains a population of individuals, P(t) = x,, . . . ,
x, at iteration t. Each individual representsa potential
solution to the problem at hand and is implemented as
some (possibly complex) data structure S. Each solution
x, is evaluated to give some measure offitness. Then a
new population at iteration t + 1 is formed by selecting
the fitter individuals. Some members of the new population undergo transformation by means of genetic operators to form new solutions. There are unary transformations m, (mutation type), which create new individuals
by a small changein a single individual and higher order
transformations c, (crossovertype), which create new individuals by combining parts from several(two or more)
individuals. For example, if parents are representedby a
five-dimensional vector (a,, a2, a3, a4, a5) and (b,, b2, b3,
b4,b,), then a crossoverof chromosomesafter the second
geneproduces offspring (a,, a2, b3, b4, b,) and (b,, b2,a3,
a4, as). The control parameters for genetic operators
Target
10
3110
I l/l I
7110
5.617.I
74.0/90.4
II
4
10
515
616
Often compared with the neural networks and the

symbolic learning methods, the self-adaptivenessproperty of genetic algorithms is also extremely appealingfor
IR applications.
208
Int. 7
ID3/ID5R
616
tropy concept in selecting keywords, both algorithms

were able to createminimal and understandabledecision
trees efficiently. However, IDSRs incremental learning
and relevancefeedbackcapabilities made it more robust
and appealingfor large-scale,real-time IR applications.
Genetic Algorithms
Int. 6
ID3/ID5R
5.617.2
74.019I .3
5.717.4
74.9193.I
6
6
5
8
12
10
8.2
(probability of crossoverand mutation) need to be carefully selectedto provide better performance. The intuition behind the crossover operation is information exchangebetweendifferent potential solutions. After some
number of generationsthe program converges-the best
individual hopefully representsthe optimum solution.
Michalewicz (1992) provided an excellent algorithmic
discussion of GAS. Goldberg (1989, 1994) presented a
good summary of many recent GA applications in biology, computer science,engineering,operations research,
physical sciences,and social sciences.
Genetic algorithms use a vocabulary borrowed from
natural genetics in that they talk about genes (or bits),
chromosomes (individuals or bit strings), and population (of individuals). Populations evolve through generations. Our geneticalgorithm was executedin the following steps:
(1) Initializepopulation and evaluatefitness: To initialize a population, we neededfirst to decide the number of genesfor each individual and the total number
of chromosomes @opsise)in the initial population.
When adopting GAS in IR, each gene (bit) in the
chromosome (bit string) represents a certain keyword or concept. The loci (locations of a certain
gene) decide the existence (1, ON) or nonexistence
(0, OFF) of a concept. A chromosome therefore represents a document that consists of multiple concepts. The initial population contains a set of documents which were judged relevant by a searcher
through relevance feedback.The goal of a GA was to
find an optimal set of documents which best
matched the searchersneeds (expressedin terms of
underlying keywords or concepts). An evaluation
function for the@ness of each chromosome was selected basedon Jaccardsscore matching function as
used by Gordon(1988)for documentindexing.The
Jaccards score between two sets,X and Y, was computed as:
SCIENCE-April
1995
or converged (i.e., showed no improvement in the

overall fitness of the population).
#(Xfl Y)/#(XU Y)
where #(S) indicated the cardinality of set S. The
Jaccards score is a common measure of association
in information retrieval (van Rijsbergen, 1979).
(2) Reproduction (selection): Reproduction is the selection of a new population with respect to the probability distribution based on the fitness values. Fitter
individuals have better chancesof being selectedfor
reproduction (Michalewicz, 1992). A roulette wheel
with slots (F) sized according to the total fitness of
the population was defined as follows:
pop3ize
F = C Jitness(VJ
,=I
wherejfitnc~ss(
Vj) indicated the fitness value of chromosome V, according to the Jaccardsscore.
Each chromosome had a certain number of slots
proportional to its fitness value. The selection process was based on spinning the wheel popsize
times-each time we selecteda single chromosome
for a new population. Obviously, some chromosomes were selected more than once. This is in accordance with the genetic inheritance: the best chromosomesget more copies, the averagestay even, and
the worst die off.
(3) Recombination (crossover and mutation): We were
then ready to apply the first recombination operator,
crossover,to the individuals in the new population.
The probability of crossover,pr, gave us the expected
number pc X popsize of chromosomes which should
undergo the crossover operation. For each chromosome, we generated a random number r between 0
and I; if r < pr, then the chromosome was selected
for crossover. We then mated selectedpairs of chromosomes randomly: for each pair of coupled chromosomes we generateda random number pas from
the range of (1.. .m - l), where m was the total
number of genesin a chromosome. The numberpos
indicated the position ofthe crossing point. The coupled chromosomes exchanged genes at the crossing
point as described earlier.
The next recombination operator, mutation, was
performed on a bit-by-bit basis. The probability of
mutation, pm, gave us the expected number of mutated bits p,,, X m X popsize. Every bit in all chromosomes of the whole population had an equal chance
to undergo mutation, that is, change from 0 to 1 or
vice versa. For each chromosome in the crossovered
population, and for each bit within the chromosome,
we generated a random number r from the range of
(0. . 1); if r < pm, we mutated the bit. Typical pc selected ranged between 0.7 and 0.9 and pm ranged between 0.0 1 and 0.03.
(4) Convergence: Following reproduction, crossover,
and mutation, the new population was ready for its
next generation. The rest of the evolutions were simply cyclic repetitions of the above steps until the system reacheda predetermined number of generations
JOURNAL
A GA Example
We present a sample session, implementation details,

and some benchmark testing results below. In our system, a keyword represented a gene (bit) in GAS; a userselecteddocument representeda chromosome (individual); and a set of user-selecteddocuments represented
the initial population.
The keywords used in the set of user-selecteddocuments were first identified to representthe underlying bit
strings for the initial population. Each bit representedthe
same unique keyword throughout the complete GA process.When a keyword was presentin a document, the bit
was set to 1, otherwise it was 0. Each document could
then be representedin terms of a sequenceof 0s and 1s.
The keywords of five user-selecteddocuments are presentedbelow. The set of unique conceptspresentin these
sample documents is also summarized-33 keywords
(genes)in total. As in the Hopfield network example,
some concepts were folder names assignedby the users
(in the format of * .*); for example, QUERY.OPT folder
for query optimization topics.
We computed the fitness of each document basedon
its relevance to the documents in the user-selectedset.
Higher Jaccards score (a value between 0 and 1) indicated stronger relevance between two documents. For
document 0, we computed five different Jaccardsscores
between document 0 and documents 0, 1, 2, 3, and 4,
respectively (shown below). An averagefitness was then
computed for document 0 (0.28774). The same procedure was applied to other documents to compute their
fitness. A document which included more concepts
sharedby other documents had a higher Jaccardsscore.
Jaccards Score of DOCO and DOCO = 1.OOOOOO
Jaccards Score of DOCO and DOCl = 0.120000
Jaccards Score of DOCO and DOC2 = 0.120000
Average Fitness (JaccardsScore) of Document0 : 0.28774
If a user provided documents that are closely related,

the averagefitness for the complete document set was
high. If the user-selecteddocuments were only loosely
related, their overall fitness was low. Generally, GAS did
a goodjob optimizing a document set which was initially
low in fitness. Using the previous example, the overall
Jaccards score increased over generations. The optimized population contained only one single chromosome, with an average fitness value of 0.45 12 1. The optimized chromosome contained six relevant keywords
OF THE AMERICAN SOCIETY FOR INFORMATION
SCIENCE-April
1995
209
Input Documents and Keywpords
DOCO DATA
DOCl
DOC2
DOC3
DOC4
RETRIEVAL,
DATABASE,
Total Set gf Concepts

DATA RETRIEVAL. DATABASE, COMPUTER. NETIMPROVEMENTS,
INFORMATION
REWORKS,
TRIEVAL, METHOD, NETWORK, MULTIPLE, QUERY,
RELATION, RELATIONAL, RETRIEVAL, QUERIES. RELATIONAL DATABASES. RELATIONAL DATABASE. US,
CARAT.DAT, GQP.DAT. ORUS.DAT, QUERY.OPT. INFORMATION. INFORMATION STORAGE, INDEXING,
STORAGE, KEVIN.HOT, ARTIFICIAL INTELLIGENCE,
INFORMATION RETRIEVAL SYSTEMS, NATURAL LANGUAGE PROCESSING, DBMSAI, FUZZY SET THEORY,
PERFORMANCE, RETRIEVAL SYSTEMS, STAIRS.
-
Initial Genetic Pattern sf Chromosome in Population __
chromosome
111111111111111111110000000000000
000010000001000100001111100000000
000010000000000101000010011110000
000000000001100100000010101001110
000000000001000100000010101000001
Average Fitness = 0.389 I
fitness
[0.287744]
[0.4 I 16921
[0.367556]
[0.427473]
[0.451212]
which best describedthe initial set of documents. Using

these optimized keywords, an information retrieval
system could proceed to suggestrelevant documents to
users.The user-GA interaction continued until a search
was completed or the user decidedto stop.
Table 4 summarizes the results of a benchmark testing. In the testing we randomly retrieved five test casesof
1-document, 2-document. 3-document, 4-document, 5document, and 1O-document examples, respectively.
from the 3000-document DIALOG-extracted database
discussedearlier. There were 30 test casesin total. For
each test case, an initial fitness based on the Jaccards
score was computed. For l-document and 2-document
210
chromosome
000000000001000100000010101000001
000000000001000100000010101000001
000000000001000100000010101000001
000000000001000100000010101000001
000000000001000100000010101000001
Average Fitness = 0.45 12
COMPUTER
NETWORKS, IMPROVEMENTS, INFORMATION

RETRIEVAL, METHOD, NETWORK, MULTIPLE,
QUERY, RELATION, RELATIONAL, RETRIEVAL,
QUERIES, RELATIONAL DATABASES, RELATIONAL DATABASE, US, CARAT.DAT, GQP.DAT,
ORUSDAT, QUERY.OPT
INFORMATION, INFORMATION RETRIEVAL,
INFORMATION
STORAGE, INDEXING, RETRIEVAL, STORAGE, US, KEVIN.HOT
ARTIFICIAL INTELLIGENCE, INFORMATION RETRIEVAL SYSTEMS, INFORMATION RETRIEVAL,
INDEXING, NATURAL LANGUAGE PROCESSING, US, DBMS.AI, GQP.DAT
FUZZY SET THEORY, INFORMATION
RETRIEVAL SYSTEMS, INDEXING,
PERFORMANCE, RETRIEVAL SYSTEMS, RETRIEVAL,
QUERIES, US, KEVIN.HOT
INFORMATION RETRIEVAL SYSTEMS. INDEXING, RETRIEVAL, STAIRS, US, KEVIN.HOT
Optimized Chromosomes in the Population ~
fitness
[0.45 121]
[0.45121]
[0.45121]
[0.45 12I]
[0.45121]
Derived Conceptsfrom Optimized Population-
RETRIEVAL, US, INDEXING, KEVIN.HOT, INFORMATION RETRIEVAL SYSTEMS, STAIRS.
test cases,their initial fitness tended to be higher due to

the smaller sample size (seecolumn 2 of Table 4). In Table 4 we also report performance measuresin terms of
Jaccard scoresfor the GA processes,the CPU times, and
the averageimprovements in fitness.
Using the GA optimization process, our system
achievedan averagefitness improvement from 5.38%to
17.7%. This improvement was slightly worse than the
performance improvement for indexing reported by
Gordon (1988). An interesting observation was that
when more initial documents were present, the initial
fitness tended to be lower, which allowed the system to
do a better job in improving the precisenessof the initial
keywords and in identifying other relevant documents.
As shown in Table 4, fitness improvement increasedas a
function of the number of initial documents. This finding also suggestedthat when initial user-supplieddocuments are fuzzy and not well articulated, GAS may be
able to make a more significant contribution in suggesting other relevant documents. This could be quite important for complex information retrieval sessionsduring which searchersneed help in query articulation and
searchrefinement.
The number of documents suggestedby GANNET after the first GA processwas between 9 and 13, with an
averageof about 11documents. The CPU times required
of the GA process also was quite reasonable,with an averageof 0.168 seconds.The responsetimes were significantly better than the Hopfield net activation. In conclusion, by using reproduction and the genetic operators,
GAS provided an interesting system-aidedway of analyzing usersintermediate search results and suggesting
other potentially relevant documents.
Conclusion
and Future Directions
Information retrieval research has been advancing

very quickly over the past few decades.Researchershave
experimented with techniques ranging from probabilistic models and the vector spacemodel to the knowledgebased approach and the recent machine learning tech-
SCIENCE-April
1995
TABLE 4. Results of genetic algorithms testing.

Impr.
@)
CPU
bc.)
I.0
1.0
1.0
1.0
1.0
I.0
0.0
0.0
0.0
0.0
0.0
0.0
0.067
0.05
0.067
0.05
0.067
0.06
7
25
7
9
5
10.6
0.5139
0.5833
0.6111
0.6486
0.7857
avg.
0.5139
0.5833
0.6111
0.6486
0.7857
0.6285
0.0
0.0
0.0
0.0
0.0
0.0
0.083
0.1
0.083
0.067
0.083
0.08
10
8
5
10
16
9.8
1
2
3
4
5
3 dots.
0.384 1
0.4157
0.4286
0.5032
0.5899
avg.
0.3984
0.4360
0.46 11
0.5215
0.6349
0.4904
3.72
4.88
7.1
3.6
7.6
5.38
0.023
0.1
0.1
0.133
0.083
0.088
8
5
13
5
16
9.4
1
2
3
4
5
4 dots.
0.2898
0.3078
0.3194
0.3319
0.4409
avg.
0.3010
0.3142
0.3495
0.3442
0.5060
0.3629
3.8
2.1
9.4
3.7
14.7
6.74
0.117
0.1
0.283
0.25
0.25
0.2
22
15
5
II
10
12.6
1
2
3
4
5
0.3048
0.3068
0.3194
0.4655
0.6181
avg.
0.3370
0.3267
0.3575
0.567 1
0.7171
0.4610
10.5
6.4
11.9
21.8
16.0
13.32
0.4
0.15
0.52
0.3
0.12
0.298
12
7
5
21
21
13.2
IO dots.
0.2489
0.2038
0.2016
0.4997
0.3727
avg.
0.2824
0.2282
0.2343
0.6201
0.4540
0.3638
13.5
12.9
16.2
24.1
21.8
17.7
0.32
0.35
0.47
0.13
0.13
0.28
18
8
6
5
I1
9.6
All
avg.
0.5511
0.168
10.87
No.
Init. score
1
2
3
4
5
I dot.
1.0
1.0
1.0
1.0
1.0
avg.
1
2
3
4
5
2 dots.
5 dots.
I
2
3
4
5
GA score
niques. At each stage,significant insights regardinghow

to design more useful and intelligent information retrieval systemshave been gained.
In this article, we presentedan extensivereview of IR
research that was based mainly on machine learning
techniques. Connectionist modeling and learning, in
particular, has attracted considerableattention due to its
strong resemblanceto someexisting IR models and techniques. Symbolic machine learning and genetic algorithms, two popular candidatesfor adaptive learning in
other applications, on the other hand, have been used
only rarely. However, thesenewer techniques have been
found to exhibit promising inductive learning capabilities for selectedIR applications.
7.19
Dots.
selected
For researcherswho are interestedin examining these

techniques, this study has discussedan algorithmic approach and knowledge representationsappropriate for
IR. We feel that the proper selectionof knowledgerepresentation and the adaptation of machine learning algorithms in the IR context are essentialto the successful
use of such techniques. For example, in IR a keyword
could representa node in the Hopfield net, a singlebit in
a geneticalgorithm, or a decision node in ID3 and IDSR.
Similarly, the paralkl relaxation searchof the Hopfield
net, the entropy reduction schemein ID3, and the Darwinian seIectionof genetic algorithms all need to be carefully studied and modified in the unique IR context.
Despite some initially successful application of se-
1995
211
among the most advanced full-text retrieval systems

and feature content-based IR and learning capabilities. PixTex calls its indexing process learning.
The system automatically extracts patterns from binary data (texts or images) and associates(or learns)
the storage location of the data based on neural network technology (the exact form and algorithm are not
clear due to the lack of publications on and the proprietary nature of the product). By automatically storing
visual scene or textual contents in terms of Huffman
codes. the system can then retrieve other similar scene
objects or texts during IR. Veritys Topic claims to use
fuzzy logic in its design of conceptual searching for
intelligent document retrieval systems. It allows users to create and reuse hierarchical, weighted query
trees (thus becoming part of the corporate menwry),
which produce rank-ordered documents. It also appears to have some similarity search capability (e.g.,
find me all documents like this one). However. like
PixTex. no algorithmic detail can be obtained. Despite
the lack of implementation detail, we believe that with
the extensive indexing capabilities provided by such
full-text software, a simple user relevance feedback
component and inductive machine learning algorithms. similar to the ones discussed in this research,
could be incorporated to help identify what userswant,
based on the concepts (keywords) learned from the
sample documents. As more researchers and practitioners recognize the need for concept-based and intelligent IR, application of machine learning algorithms presents unique challenges and opportunities.
lected machine learning techniques for IR, there are numerous research directions that need to be pursued before we can develop a robust solution to intelligent information retrieval. We briefly review several important
researchdirections below:
l
Limitations of learning techniquesfor IR: The performance of the inductive learning techniques relies
strongly on the examples provided (as in any other statistical and classification techniques) (Weiss & Kulikowski, 199 1). In IR, these examples may include userprovided queries and documents collected during relevance feedback. The importance of sample size has
been stressedheavily, even in the probabilistic models
(Fuhr & Buckley, 1991: Fuhr & Pfeifer, 1994). In reality, user-provided relevance feedback information may
be limited in quantity and noisy (i.e., contradictory or
incorrect), which may have adverse effects for the IR or
indexing tasks. Some learning techniques such as the
neural networks approach have documented noise-resistant capability, but empirical evidence and research
need to be performed to verify this characteristic in the
context of IR and indexing. In our preliminary investigation, all three machine learning algorithms performed satisfactorily for small document samples. but
the effect of the sample size needsto be examined more
carefully.
For large-scale real-life applications, neural networks and, to some extent, genetic algorithms, may
suffer from requiring extensive computation time and
lack of interpretable results. Symbolic learning, on the
other hand. efficiently produces simple production
rules or decision-tree representations. The effectsofthe
representations on the cognition of searchers in the
real-life retrieval environments (e.g., usersacceptance
of the analytical results provided by an intelligent system) remain to be determined.
Applicahilit~~ to the fkll-text retrieval environment: In
addition to extensive IR research conducted in probabilistic models, knowledge-based systems, and machine learning, significant efforts have also been made
by many commercial companies in pursuit of more
effective and intelligent information retrieval systems. In an attempt to understand the potential role
of machine learning in commercial full-text retrieval
systems, we examined several major full-text retrieval
software packages on the market, including: BRS/
SEARCH, BASIS/Plus, PixTex,3 and Topic.4
Most full-text retrieval software has been designed
to handle large volumes of text by indexing every word
(a!d its position). This allows users to perform proximity search, morphological search (using prefix,
suffix, or wildcards), and thesaurus search. BRS/
SEARCH and BASIS/plus are typical of this type of
software. PixTex and Topic, on the other hand, are
Vendedby BRSSoftwareProducts,McLean,VA, USA.

2 Vendedby Information DimensionsInc.. Dublin, OH, USA.
3Vendedby ExcaliburTechnologiesCorp., McLean,VA, USA.
4 Vendedby Verity, Inc., Mountain View, CA, USA.
212
JOURNAL
OF THE AMERICAN
We believe this researchhas shed light on the feasibility and usefulnessof the newer, AI-based machine learning algorithms for IR. However, more extensive and systematic studies of various system parameters and for
large-scale,real-life applications are needed. We hope by
incorporating into IR inductive learning capabilities,
which are complementary to the prevailing full-text, keyword-based, probabilistic, or knowledge-based techniques, we will be able to advance the design of adaptive
and intelligent information retrieval systems.
Acknowledgments
This project was supported mainly by NSF Grant

#IRI-92 11418, 1992- 1994 (NSF/CISE, Division of Information, Robotics, and Intelligent Systems).
References
Appelt, D. (1985,August).The role of usermodellingin languagegenerationand communicationplanning.In UserModelling Panel, Proceedings of the Ninth International Joint Conferenceon Arrificial Intelligence, (pp. 1298-1302).Los Altos,CA: MorganKaufmann Pub-
lishers,Inc.
Belew,R. K. (1989.June).Adaptiveinformation retrieval.In Proceed-
ings ofthc Tne[fih Annual InternationaliiCM/SIGIR
SCIENCE-April
1995
Conferenceon
Reseurch and Development in Information Retrieval (pp. 1l-20).

NY, NY: ACM Press.
Blair, D. C., & Maron, M. E. (1985). An evaluation of retrieval
effectivenessfor a full-text document-retrieval system. Communications of theACM, 28,289-299.
Blosseville, M. J., Hebrail, G., Monteil, M. G., & Penot, N. (1992,
June). Automatic document classification: Natural language processing, statistical analysis, and expert system techniques used together. In Proceedingsof the Ftifieenth Annual International ACM/
SIGIR Conferenceon Researchand Developmentin Information Retrieval(pp. 5 l-57). NY, NY: ACM Press.
Booker, L. B., Goldberg, D. E., & Holland, J. H. (1990). Classifier systems and geneticalgorithms. In J. G. Carbonell (Ed.), Machine learning, paradigms and methods (pp. 235-282). Cambridge, MA: The
MIT Press.
Bookstein, A., & Swanson,D. R. (1975). Probabilistic models for automatic indexing. Journal ofthe American Societyfor Information Science,26.45-50.
Borgida, A., & Williamson, K. E. (1985, August). Accommodating exceptions in a database,and refining the schema by learning from
them. In Proceedings of the 11th International VLDB Conference
(pp. 72-8 1). Saratoga,NY: VLDB Endowment.
Brajnik, G., Guida, G., & Tasso, C. (1988). IR-NLI II: Applying manmachine interaction and artificial intelligence concepts to information retrieval. In Proceedingsof the Eleventh Annual International
ACM/SIGIR Conferenceon Researchand Development in Information Retrieval(pp. 387-399). NY, NY: ACM Press.
Brauen, T. L. (197 1). Document vector modification. In G. Salton
(Ed.), The Smart retrieval system-experiments in automatic documentprocessing(pp. 456-484). Englewood Cliffs, NJ: Prentice-Hall.
Breiman, L., Friedman, J., Olshen, R., & Stone, C. (1984). Clussification and regressiontree. Monterey, CA: Wadsworth.
Buckland, M. K., & FIorian, D. (1991). Expertise,task complexity, and
artificial intelligence: A conceptual framework. Journal of the Americun SocietJ,,forInformation Science,42, 635-643.
Cai, Y., Cercone, N., &Han, J. (1991). Attribute-oriented induction in
relational databases.In G. Piatetsky-Shapiro& W. J. Frawley (Eds.),
Knowledge discovery in databases(pp. 2 13-228). Cambridge, MA:
The MIT Press.
Carbonell, J. G., Michalski, R. S., & Mitchell, T. M. (1983). An overview of machine learning. In R. S. Michalski, J. G. Carbonell, &
T. M. Mitchell (Eds.), Machine Learning, An Artificial Intelligence
Approach (pp. 3-23). Palo Alto, CA: Tioga.
Chen, H., Basu, K., & Ng, T. (in press-a).An algorithmic approach to
concept exploration in a large knowledge network (automatic thesaurus consultation): Symbolic branch-and-bound vs. connectionist
Hoplield net activation. Journal of the American Society for Information Science.
Chen, H., Buntin, P.. She, L., Sutjahjo, S., Sommer, C., & Neely, D. (in
press). Expert prediction, symbolic learning, and neural networks:
An experiment on greyhound racing. IEEE Expert.
Chen, H., & Dhar, V. (1987, July). Reducing indeterminism in consultation: a cognitive model of user/librarian interaction. In Proceedings of the 6th National Conferenceon Artificial Intelligence (AAAI87) (pp. 285-289). Los Altos, CA: Morgan Kaufmann Publishers,
Inc.
Chen, H., & Dhar. V. ( 1990). User misconceptions of online information retrieval systems.International Journal of Man-Machine Studies, 32, 673-692.
Chen, H., & Dhar, V. ( I99 I). Cognitive processas a basis for intelligent
retrieval systems design. Information Processingand Management,
27,405-432.
Chen, H., Hsu, P., Orwig, R., Hoopes, L., & Nunamaker, J. F. (1994b).
Automatic concept classification of text from electronic meetings.
Communications of the ACM, 37, 56-73.
Chen, H., & Kim, J. (1993). GANNET: Information retrieval usinggenetics algorithms and neural networks. (Working Paper, CMI-WPS).
Center for Management of Information, College of Business and

Public Administration, University of Arizona.
Chen, H., & Lynch. K. J. (1992). Automatic construction of networks
of concepts characterizing document databases.IEEE Transactions
on Systems,Man and Cybernetics,22, 885-902.
Chen, H., Lynch, K. J., Basu, K., & Ng, T. (1993). Generating, integrating, and activating thesauri for concept-based document retrieval. IEEE Expert (special serieson Artificial Intelligence in TextBased Information Systems),8, 25-34.
Chen, H., & She, L. (1994, January). Inductive query by examples
(IQBE): A machine learning approach. In Proceedingsof the 27th
Annual Hawaii International Conference on System Sciences
(HICSS-27). Information Sharing and Knowledge Discovery Track.
Los Alamitos, CA: IEEE Computer Society Press.
Chiaramella, Y., & Defude, B. (1987). A prototype of an intelligent
system for information retrieval: IOTA. Information Processingand
Management, 23, 285-303.
Cohen, P. R., & Kjeldsen, R. (1987). Information retrieval by constrained spreadingactivation in semantic networks. Information Processingand Management, 23, 255-268.
Crawford, S. L., Fung, R., Appelbaum, L. A., & Tong, R. M. (1991).
Classification trees for information retrieval. In Proceedingsof the
8th International Workshop on Machine Learning (pp. 245-249).
San Mateo, CA: Morgan Kaufmann.
Crawford, S. L., & Fung, R. M. (1992). An analysis of two probabilistic
model induction techniques. Statistics and Computing, 2, 83-90.
Croft, W. B.. & Thompson. R. H. (1987). 13R:A new approach to the
design of document retrieval systems.Journal of the American Societyfor Information Science,38, 389-404.
Dalton, J., & Deshmane, A. (199 I). Artificial neural networks. IEEE
Potentials, 10. 33-36.
Daniels, P. J. ( 1986).The user modelling function of an intelligent interface for document retrieval systems.In B. C. Brookes (Ed.), Intelligent information systemsfor the information society. Amsterdam:
Elsevier.
Derthick, M. (1988). Mundane reasoning by parallel constraint satisfaction. Ph.D. thesis,Carnegie Mellon University, Pittsburgh, PA.
Doszkocs, T. E., Reggia, J., & Lin, X. (1990). Connectionist models
and information retrieval. Annual Review of Information Science
and Technology(ARIST), 25,209-260.
Ever&, B. (1980). Cluster analysis (2nd ed.). London: Heinemann.
Fisher, D. H., & McKusick, K. B. (1989, August). An empirical comparison of ID3 and backpropagation. In Proceedingsof the Eleventh
International Joint Conferenceon Artificial Intelligence (IJCAI-89)
(pp. 788-793). San Mateo, CA: Morgan Kaufmann Publishers,Inc.
Fogel, D. B. ( 1994).An introduction to simulated evolutionary optimization. IEEE Transactions on Neural Networks, 5, 3-14.
Fogel, L. J. (1962). Autonomous automata. Industrial Research,4. (pp.
14-19)
Fogel, L. J. (1964). On the organization of intellect. Doctoral dissertation, UCLA, Los Angeles,CA.
Fox, E. A. (1987). Development of the CODER system: A testbed for
artificial intelligence methods in information retrieval. Information
ProcessingandManagement, 23, 341-366.
Frawley, W. J., Pietetsky-Shapiro,G., & Matheus, C. J. (1991). Knowledgediscovery in databases:An overview. In G. Piatetsky-Shapiro&
W. J. Frawley (Eds.), Knowledge discovery in databases(pp. I-30).
Cambridge, MA: The MIT Press.
Freund, J. E. (197 I). Mathematical statistics. Englewood Cliffs, NJ:
Prentice-Hall.
Frieder, O., & Siegelmann,H. T. (199 1, October). On the allocation of
documents in multiprocessor information retrieval systems.In Proceedingsof the Fourteenth Annual International ACM/SIGIR Conference on Researchand Development in Information Retrieval (pp.
230-239). NY, NY: ACM Press.
Fuhr, N., & Buckley, C. (199 1). A probabilistic learning approach for
1995
213
document indexing. ACM Transactions on Information Systems, 9,

223-248.
Fuhr, N., Hartmann. S., Knorz, G., Lustig. G.. Schwantner, M., &
Tzeras, K. (1990. July-August). AIR/X-a rule-basedmultistage indexing system for large subject fields. In Proceedingsof the Eighth
National Conferenceon Artificial Intelligence (AAAI-90) (pp. 78%
795). Boston, MA.
Fuhr, N., & Pfeifer. U. (1994). Probabilistic information retrieval as
a combination of abstraction, inductive learning. and probabilistic
assumptions. ACM Transactions on Information Systems, 12. 92115.
Fung, R., & Crawford. S. L. (1990. July-August). Constructor: A system for the induction of probabilistic models. In Proceedingsqfthe
8th National Conference on ArtiJicial Intelligence (,4AAI-90) (pp.
762-769). Boston. MA.
Gallant, S. I. (1988). Connectionist expert system. Communications qf
theACM. 31, 152-169.
Goldberg, D. E. ( 1989).Geneticalgorithms in search,optimization, and
machine learning. Reading, MA: Addison-Wesley.
Goldberg. D. E. (1994). Genetic and evolutionary algorithms come of
age.Communications ofthe ACM. 37. 113-I 19.
Gordon, M. (1988). Probabilistic and genetic algorithms for document
retrieval. G,mmllnicafions ofthe ACM, 31. 1208-12 18.
Gordon, M. D. ( 199I). User-baseddocument clustering by redescribing
subject descriptions with a genetic algorithm. Journal of the American Societyfiw Infijrmation Science,42. 3 1 l-322.
Greene, D. P., & Smith, S. F. (1992). COGIN: Symbolic induction with
genetic algorithms. In Proceedingsqfthe Tenth Nutional Conftirence
on Artificial Intelligence fAAAI-92) (pp. 1 1I - 1 16). Cambridge. MA:
The MIT Press.
Hall, L. O., & Romaniuk. S. G. (1990, July-August). A hybrid connectionist, symbolic learning system. In Proceedingsofthe Eighth Nutional Cor&wzce ondrt$cial Intelligence IAAAI-9Oj (pp. 783-788).
Cambridge, MA: The MIT Press.
Han, J.. Cai. Y., & Cercone. N. (1993). Data-driven discovery ofquantitative rules in relational databases.IEEE Transactions on Know+
edgeand Datu Engineering, 5. 29-40.
Harp. S.. Samad, T., & Guha, A. (1989). Towards the genetic synthesis
of neural networks. In Proceedingsofthe Third International Conference on GeneticAlgorithms. San Mateo, CA: Morgan Kaufmann.
Hayes-Roth, F., & Jacobstein,N. (1994). The state ofknowledge-based
systems.Communications c?fiheACM. 37, 27-39.
Holland, J. H. (1975). Aduptation in naturalandart~ficialsystems. Ann
Arbor. MI: University of Michigan Press.
Hopfield, J. J. (1982). Neural network and physical systems with collective computational abilities. Proceedingsofthe National Academy
ofScience. LTSA,78(8) (pp. 2554-2558).
Humphreys. B. L.. & Lindberg, D. A. (1989, November). Building the
unified medical language system. In Proceedingsof the Thirteenth
Annual S?fmposium on Computer Applications in Medical Care.
Washington. DC: IEEE Computer Society Press.
Ide, E. (197 I). New experiments in relevance feedback. In G. Salton
(Ed.). The Smart retrieval system-experiments in automatic document procex.sinl:(pp.337-354). Englewood Cliffs, NJ: Prentice-Hall.
Ide, E., & Salton, G. (197 1). Interactive searchstrategiesand dynamic
file organization in information retrieval. In G. Salton (Ed.). The
Smart retrieval sy.s/em-esperimmts in automatic document processinR(pp.373-393). Englewood Cliffs, NJ: Prentice-Hall.
Ioannidis, Y. E., Saulys: T.. & Whitsitt. A. J. (1992). Conceptual learning in database design. ACM Transactions on Information SJJstems,
10,265-293.
Kitano. H. (I 990. July-August). Empirical studieson the speedof convergenceof neural network training using genetic algorithms. In Proceedingsofthe Eighth National Conferenceon Artificial Intelligence
(AAAI-9Oj (pp. 789-795). Cambridge. MA: The MIT Press.
Knight, K. (1990). Connectionist ideas and algorithms. Commumcations efthe ACM, 33, 59-74.
214
Kohonen, T. (1989). Se&organization and associative memory (3rd

ed.). Berlin: Springer-Verlag.
Koza. J. R. (1992). Geneticprogramming: On theprogrammingo~con7puters LJJ~
means of natural selection. Cambridge, MA: The MIT
Press.
Kwok, K. L. (1989, June). A neural network for probabilistic information retrieval. In Proceedings of the Twelfth Annual Internationul
AC,V/SIGIR Conferenceon Researchand Development in Information Retrieval(pp. 2 I-30). NY, NY: ACM Press.
Lebowitz, M. (1987). Concept learning in a rich input domain: Generalization-basedmemory. In J. G. Carbonell, R. S. Michalski, &T. M.
Mitchell (Eds.), ?MachineIeurning, an artificial intelligence approach
(vol. II) (pp. 193-2 14.463-482). Los Altos, CA: Morgan Kaufmann.
Lewis. D. D. (199 I). Learning in intelligent information retrieval. In
Proceedingsqf the 8th International workshop on machine learning
(pp. 235-239). Los Altos, CA: Morgan Kaufmann.
Lewis, D. D. (1992, June). An evaluation of phrasal and clustered representations on a text categorization task. In Proceedingsqfthe Fif
teenth Annual International ACjLf/SIGIR Corzferenceon Research
and Development in Information Retrieval (pp. 37-50). NY, NY:
ACM Press.
Li. Q.. & McLeod. D. (1989). Object flavor evolution through learning
in an object-oriented databasesystem.In L. Kerschberg(Ed.), Expert
Databa.seSystems, Proceedingsfrom the SecondInternational Con,f&ence (pp. 469-495). Menlo Park. CA: Benjamin/Cummings.
Lin, X., Soergel.D.. & Marchionini. G. (1991. October). A self-organizing semantic map for information retrieval. In Proceedingsofthe
Fourteenth Annual International ACM/SIGIR Conference on Rc
search und Development in I&rmation Retrieval (pp. 262-269).
Chicago, IL.
Lindberg. D. A., & Humphreys, B. L. (1990, November). The UMLS
knowjledgesources:Tools for building better user interface. In Proceedingsofihe Fourteenth Annual Symposium on Computer Applications in Medical Care. Los Alamitos, CA: Institute of Electrical
and Electronics Engineers.
Lippmann, R. P. (1987). An introduction to computing with neural
networks. IEEE Acoustics Speechand Signal ProcessingMagazine,
4, 4-22.
MacLeod, K. J.. & Robertson, W. (1991). A neural algorithm for document clustering. I&rmation Processing& Management, 27, 337346.
Maron, M. E., & Kuhns. J. L. (1960). On relevance, probabilistic indexing and information retrieval. Journal qfthe ACM, 7, 2 16-243.
Martin, B. K.. & Rada, R. (1987). Building a relational data basefor a
physician document index. Medical Informatics, 12, 187-20 1.
Masand, B.. Gordon. L., & Waltz, D. (1992, June). Classifying news
stories using memory-based reasoning. In Proceedings qf the Fif
teenth ilnnual Internationul.4CM/SIGIR Conferenceon Research
andDevelopment in IF@mation Retrieval(pp. 59-65). Copenhagen.
Denmark.
McCray, A. T.. & Hole, W. T. (1990). The scope and structure of the
first version of the UMLS semantic network. In Proceedingsof the
Fourteenth Annual Symposium on Computer Applications in Medical Care. Los Alamitos. CA: Institute of Electrical and Electronics
Engineers.
Michalewicz. Z. ( 1992). Gerzeticalgorithms + data structwes = evolrrtion programs. Berlin: Springer-Verlag.
Michalski. R. S. (1983). A theory and methodology of inductive Iearning. In R. S. Michalski. J. G. Carbonell. and T. M. Mitchell (Eds.).
Machine learning, an arti&ial intelligence approach (pp. 83-l 34).
Palo Alto. CA: Tioga.
Mitchell, T. M. (1982). Generalization as search.ArtificialIntelligence.
18.203-226.
Monarch, I.. & Carbonell, J. G. (I 987). CoalSORT: A knowledge-based
interface. IEEE Expert, 39-53.
Montana, D. J., & Davis, L. (1989, August). Training feedforward neural networks using genetic algorithms. In Proceedingsqfthe Eleventh
1995
International Joint Conferenceon Artificial Intelligence (IJCAI-89)

(pp. 762-767). San Mateo, CA: Morgan Kaufmann Publishers, Inc.
Montgomery, D. D. (1976). Design and analysis ofexperiments. New
York: Wiley.
Mooney, R., Shavlik, J., Towel], G., & Gove, A. (1989, August). An
experimental comparison of symbolic and connectionist learning algorithms. In Proceedingsofthe Eleventh International Joint Conferenceon Art$cial Intelligence (IJCAI-89) (pp. 775-780). San Mateo,
CA: Morgan Kaufmann Publishers.
Parsaye,K., Chignell, M., Khoshafian, S., & Wong, H. (1989). Intelligent databases.New York: Wiley.
Pen-y.F., Buckles, B., Prabhu. D., & Kraft, D. (1993). Fuzzy information retrieval using genetic algorithms and relevance feedback. In
Proceedingsof the .4SIS Annual Meeting (pp. 122- 125) Medford,
NJ: ASIS.
Piatetsky-Shapiro, G. (1989). Workshop on knowledge discovery in
real databases.In International Joint ConferenceofArtificial Intelligence.San Mateo, CA: Morgan Kaufmann Publishers.
Pollitt, S. (1987). Cansearch:An expert systemsapproach to document
retrieval. Inltirmation Processingand Management, 23, 1 I9- 138.
Quinlan. J. R. (I 979). Discovering rules by induction from large collections of examples. In D. Michie (Ed.), Expert systems in the microelectronic age (pp. 168-201). Edinburgh: Edinburgh University
Press.
Quinlan, J. R. (1983). Learning efficient classification proceduresand
their application to chessend games.In R. S. Michalski, J. G. Carbonell, & T. M. Mitchell (Eds.), Machine learning, an artijcial intelligenceapproach(pp. 463-482). Palo Alto, CA: Tioga.
Quinlan, J. R. (1986). Induction of decision trees. Machine Learning,
1,81-106.
Quinlan, J. R. (1993). C4.5: Programs,for machine learning. Los Altos,
CA: Morgan Kaufmann.
Rada, R., Mili. H., Bicknell, E., & Blettner. M. (1989). Development
and application of a metric on semantic nets. IEEE Transactionson
Systems,Man, und Cvbernetics, 19. 17-30.
Raghavan, V. V., & Agarwal. B. (1987. July). Optimal determination
of user-oriented clusters: An application for the reproductive plan.
In Proceedingsof the Second International Conferenceon Genetic
Algorithms and thrirrlpplications (pp. 24 l-246). Hillsdale. NJ: Lawrence Erlbaum Associates.
Rau, L. F., & Jacobs, P. S. (199 1. October). Creating segmenteddatabasesfrom free text for text retrieval. In Proceedings ~fthe Fourteenth Annual International ACM/SIGIR Conferenceon Research
and Development in Irzformation Retrieval (pp. 337-346). NY, NY:
ACM Press.
Rich. E. (I 979. August). Building and exploiting user models. In International Joint ConferenceofArt$cial Intelligence (pp. 720-722).
Tokyo, Japan.
Rich, E. (1979b) User modeling via stereotypes.Cognitive Science,3,
329-354.
Rich, E. (1983). Users are individuals: Individualizing user models. International Journal qfMan-Machine Studies, IS, 199-2 14.
Roberston, S. E., & Sparck Jones, K. (1976). Relevanceweighting of
search terms. Journal qfthe American Society for Itzformation Science,27. 129-146.
Rocchio, J. J. (197 1). Relevance feedback in information retrieval. In
G. Salton (Ed.), The Smark retrieval svstem-experiments in automatic document processing (pp. 3 13-323). Englewood Cliffs, NJ:
Prentice-Hall.
Rose, D. E.. & Belew, R. K. (1991). A connectionist and symbolic hybrid for improving legal research.International Journal ofMan-Machinestudies, 35, 1-33.
Rumelhart, D. E., Hinton, G. E., & McClelland, J. L. (1986). A general
framework for parallel distributed processing.In D. E. Rumelhart, &
J. L. McClelland, &the PDP ResearchGroup (Eds.), Paralleldistributed processing(pp. 45-76). Cambridge, MA: The MIT Press.
Rumelhart, D. E., Hinton, G. E., & Williams, R. J. (1986). Learning
internal representationsby error propagation. In D. E. Rumelhart,

J. L. McClelland, &the PDP ResearchGroup (Eds.), Paralleldistributedprocessing(pp. 3 18-362). Cambridge, MA: The MIT Press.
Rumelhart, D. E., Widrow, B., & Lehr, M. A. (1994). The basic ideas
in neural networks. Communications ofthe ACM, 37, 87-92.
Salton, G. (1989). Automatic text processing.Reading, MA: AddisonWesley.
Shastri, L. (199 1). Why semantic networks? In J. F. Sowa (Ed.), Principles of semantic networks: Explorations in the representation of
knowledge(pp. 109- 136). San Mateo, CA: Morgan Kaufmann.
Simon, H. ( 1991).Artificial intelligence: Where has it been, and where
is it going?IEEE Transactionson Knowledgeand Data Engineering,
3, 128-136.
Simpson, P. K. (1990). ArtiJicial neural sytems: Foundations, parudigms, applications, and implementations. New York: McGrawHill.
Sleeman,D. (1985). UMFE: A user modeling front-end subsystem.International Journal ofMan-Machine Studies. 23, 7 l-88.
Smith. P. J., Shute, S. J., Galdes, D., & Chignell, M. H. (1989). Knowledge-basedsearch tactics for an intelligent intermediary system.
ACM Transactions on Information Systems, 7, 246-270.
Sparck Jones, K. ( I99 I). The role of artificial intelligence in information retrieval. Journal qfthe American Societyfor Information Science,42, 558-565.
Stepp, R. E., & Michalski, R. S. (1987). Conceptual clustering: Inventing goal-oriented classificationsofstructured objects. In J. G. Carbonell et al. (Eds.),Machine learning, an artificial intelligenceapproach
(Vol. II) (pp. 472-498, 463-482). Los Altos, CA: Morgan Kaufmann.
Swat-tout. W. (1985. August). Explanation and the role of the user
model: how much will it help? In User Modelling Panel, Proceedings
of the Ninth International Joint Conferenceon Artificial Intelligence
(pp. 1298- 1302).Los Altos, CA: Morgan Kaufmann Publishers,Inc.
Tank, D. W.. & Hopfield. J. J. (1987). Collective computation in neuronlike circuits. Scient$cAmerican, 257, 104-I 14.
Touretzky, D., & Hinton, G. E. (1988). A distributed connectionist
production system. CognitiveScience, 12, 423-466.
Turtle, H.. & Croft, W. B. (1990, September). Inference networks for
document retrieval. In Proceedingsofthe 13th Annual International
ACMBIGIR Conferenceon Researchand Developmentin Information Retrieval (pp. l-24). Brussels,Belgium. NY, NY: ACM Press.
Turtle, H., & Croft. W. B. (I 99 I). Evaluation of an inference networkbased retrieval model. .4CM Transactions on In/brmation S.ystems,
9, 187-222.
Tzeras, K., & Hartmann, S. (1993, June-July). Automatic indexing
based on Bayesian inference networks. In Proceedingsef the 16th
Annual International ACM/SIGIR Conferenceon Researchand Development in Injtirmution Retrieval (pp. 22-34). NY, NY: ACM
Press.
Utgoff. P. E. ( 1989). Incremental induction of decision trees.Machine
Learning, 4, 161-186.
van Rijsbergen, C. J. (1979). Injtirmation retrieval (2nd ed.). London:
Butterworths.
Vickery. A., & Brooks, H. M. (1987). PLEXUS-the expert system for
referral. Information Processingand Management, 23, 99- 117.
Weiss, S. M., & Kapouleas, I. (1989. August). Anempirical comparison
of pattern recognition, neural nets, and machine learning classification methods. In Proceedings of the Eleventh International Joint
Conferenceon Artificial Intelligence (IJCAI-89) (pp. 78 l-787). San
Mateo. CA: Morgan Kaufmann Publishers,Inc.
Weiss, S. M., & Kulikowski, C. A. (199 1). Computer systemsthat learn.
Classification and prediction methods ,fiom statistics, neural netwlorks,machine learning, and expert systems.San Mateo, CA: Morgan Kaufmann.
Widrow, B., Rumelhart, D. E., & Lehr, M. A. (I 994). Neural networks:
Applications in industry, business,and science.Communications of
theACM, 37,93-105.
1995
215
Wilkinson, R., & Hingston, P. (I 99 1. October). Using the Cosine measure in neural network for document retrieval. In Proceedingsojthe
Fourteenth Annual International ACM/SIGIR Corzferenceon Research and Development in Illformation Retrieval (pp. 202-2 IO).
Chicago, IL.
Wilkinson, R.. Hingston. P., & Osborn, T. (1992). Incorporating the
vector spacemodel in a neural network used for document retrieval.
Library Hi Tech, IO. 69-75.
Yang, J., & Korfhage, R. R. (I 993, April). Effectsofquery term weights
modification in document retrieval: A study basedon a genetic algorithm. In Proceedingsof the Second Annttal Symposium on Docu-
216
ment Analysis und Iqfixmation Retrieval (pp. 27 l-285). Las Vegas,

NV: University of Nevada.
Yang, J., Korfhage, R. R.. & Rasmussen.E. (1993, November). Query
improvement in information retrieval using genetic algorithms: A
report on the experiments of the TREC Project. In Te,xt Retrieval
Corzfirence(TREC-I) (pp. 3 l-58). Washington, DC: NIST.
Yu. C. T.. & Salton, G. (1976). Precision weighting: An effective automatic indexing method. Journal qfthe ACM, 23, 76-88.
Zissos. A. Y.. & Witten. 1. H. (1985). User modeling for a computer
coach: A casestudy. International Journal c$Man-Machine Studies. 23, 729-750.
1995

Machine Learning For Information Retrieval

Uploaded by

Copyright:

Available Formats

Machine Learning For Information Retrieval

Uploaded by

Document Information

Original Description:

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

Machine Learning For Information Retrieval

Uploaded by

Copyright:

Available Formats

Machine Learning for Information Retrieval: Neural

Networks, Symbolic Learning, and Genetic Algorithms

In the past few decades,the availability of cheap and

SOCIETY FOR INFORMATION

In classical information retrieval models, relevance

cinctly identify relevant documents and reject irrelevant

SOCIETY FOR INFORMATION

. queries. Fuhr et al. proposed a third, feature-oriented

view knowledge-based information retrieval, and then

SOCIETY FOR INFORMATION

makes inferences about the goals of the user and thus

and access conditions of biomedical databases of all

SOCIETY FOR INFORMATION

Learning Systems: Neural Networks, Symbolic

rely on complex mathematical models, stringent assumptions, or special underlying distributions.

behaviors, have shown unique capabilities for analyzing

SOCIETY FOR INFORMATION

Stepp and Michalskis CLUSTER/RD algorithm

starts out with a random set of weights. The network

Over the past years there have been several studies

Neural networks and IR: Neural networks computing,

JOURNAL OF THE AMERICAN SOCIETY FOR INFORMATION

SOCIETY FOR INFORMATION

LCSH). Instead of using a three-layer design, Chens

ment allocations. In Chen and Kim ( 1993), a GA-NN

A Hopjield Network: KnowledgeRepresentationand

JOURNAL OF THE AMERICAN SOCIETY FOR INFORMATION

bias and B0is used to modify the shape of the SIG-

1993) and concept clustering (Chen et al., in press-b)had

(1) Assigning synaptic weights: For thesauri which were

generatedautomaticallyusing a similarity function

p,(t) is the output of node i at time t and xl, which

A Hopjield hTetwork Example

wheref, is the continuous SIGMOID transformation

SOCIETY FOR INFORMATION

Sample Hopfield net iterations.

Enter the numberof

Given these starting terms, the Hopfield net iterated

Enter numbers [lto14]

In a more structured benchmark experiment, we

JOURNAL OF THE AMERICAN SOCIETY FOR INFORMATION SCIENCE-April

TABLE 2. Results of Hoplield

(3.1, 1.2, 1.9)

For each query, we selected terms from different

Another important observation was that the Hopfield

Symbolic Learning for IR

Even though symbolic learning techniqueshave been

JOURNAL OF THE AMERICAN SOCIETY FOR INFORMATION SCIENCE-April

Knowledge Representation and Procedure

ID3 is a decision-tree building algorithm developed

a leaf node that contains a classname; or

Each training instance or object is representedas a list

the positive class (+): consisting of records that are

Different databaseusers may desire different sets of

created a binary classification tree. A sketch of the ID3

ppos - pneJog pneg

where ppOiand pnenrepresentedthe proportions of the

JOURNAL OF THE AMERICAN SOCIETY FOR INFORMATION