Cognitive Dimensions of

Information Artefacts:
a tutorial
Thomas Green and Alan Blackwell
Version 1.2 October 1998
This document can be downloaded via links at
http://www.ndirect.co.uk/~thomas.green/workStuff/Papers/
Copyright T. R. G. Green and A. F. Blackwell, 1998
2
Preface to the tutorial
This tutorial, originally prepared for the BCS HCI Conference of 1998, will provide a concise
introduction to Green's "Cognitive Dimensions", with illustrative practical applications.
The Cognitive Dimensions Framework is one of the few theoretical approaches that provides a
practical usability tool for everyday analysts and designers. It is unique in stressing simple,
cognitively-relevant system properties such as viscosity and premature user commitment , and it
illustrates important trade-offs and design manoeuvres.
The tutorial is intended for HCI practitioners and consultants, especially those actively involved in
design of novel user interfaces, and for HCI researchers with an interest in system evaluation and
implementation. It provides:
a brief comparison of types of user activity, and how each is best supported by what
profile of dimensions;
many snapshot examples to illustrate the dimensions;
example analyses of commercially-available systems;
a discussion of standard design manoeuvres and their associated trade-offs;
a set of interactive, web-available miniature examples to illustrate design alternatives and
their effects;
a Further Reading section for those wishing to extend their knowledge.
Since this tutorial is practitioner-oriented we shall focus on the more immediately practical parts of
Greens framework. On completion, you will have a vocabulary in which to analyse systems, and a
checklist with which to consider how a system relates to its intended type of user activity; you will
be able to apply the ideas to both interactive systems and non-interactive (paper-based)
information representations; you will be able to choose design manoeuvres to ameliorate usability
problems, and you will be aware of the trade-offs that each manoeuvre entails.
Part 1 describes the framework. Part 2 applies it to some commercially-available examples, to show
what types of insight can be expected from it. Part 3 gives short descriptions of some interactive
toy examples, designed to illustrate differences between design decisions and how one cognitive
dimension can be traded against another; these examples are available over the web.
About the authors
Thomas Green has published extensively on cognitive and empirical approaches to HCI and has
co-authored several contemporary techniques (TAG, ERMIA, CDs, and watch out for Ontological
Sketch Models!). Ex-APU Cambridge, now Honorary Research Fellow at the University of Leeds.
Address: Computer-Based Learning Unit, University of Leeds, Leeds LS2 9JT, U.K.
[email protected]
http://www.ndirect.co.uk/~thomas.green
Alan Blackwell has a wide background in AI, cognitive science, and professional software
engineering. He was organiser of the first "Thinking with Diagrams" workshop and is currently
researching on visual programming languages. Address: MRC Applied Psychology Unit, 15
Chaucer Road, Cambridge CB2 2EF, U. K.
[email protected]
http://www.mrc-apu.cam.ac.uk/~alan.blackwell
3
CONTENTS
Preface to the tutorial 2
PART 1: THE FRAMEWORK 4
Introduction 5
Cognitive Dimensions: the full set 11
Viscosity 12
Hidden Dependencies 17
Premature Commitment and Enforced Lookahead 21
Abstractions, Abstraction Hunger, and the Abstraction Barrier 24
Secondary Notation 29
Visibility & Juxtaposability 34
The Remaining Dimensions 39
Cognitive Relevance, Trade-offs, and Design Manoeuvres 42
Further developments in cognitive dimensions 44
PART 2: REAL SYSTEMS 46
A Diary System 47
Two Visual Programming Languages 52
References (Pts 1 and 2) 61
PART 3: INTERACTIVE PRACTICE EXAMPLES 62
The Widgets 64
Form-Filling and Menu Choices 65
Controlling the Heating Controller 66
Number Games with Telephones 67
Tiling in Style 69
FURTHER READING 71
Index of examples 75
4
Part 1: The Framework
In this part we describe the cognitive dimensions approach its purpose, its framework, the individual
components (the dimensions themselves), and its relevance to design choices. Following parts address
applications to real systems (part 2) and interactive examples for class discussion and analysis (part 3).
5
Introduction
Background and Purpose
What are Cognitive Dimensions for?
Cognitive dimensions are a tool to aid non-HCI specialists in evaluating usability of information-based
artefacts (summative evaluation). Since they are addressed to non-specialists, they consciously aim for
broad-brush treatment rather than lengthy, detailed analysis. Their checklist approach helps to ensure
that serious problems are not overlooked.
They can also be used by designers to prompt possible improvements (formative evaluation). Many
design choices, from widely different domains, can be brought together and are then seen to be
standard design manoeuvres intended to ameliorate particular usability problems. Each manoeuvre
that improves one aspect will be associated with a likely trade-off cost on other aspects, and the
dimensional approach helps to elucidate the trade-offs.
Historically, the framework was introduced by Green (1989, 1991), who emphasised the generality of
this approach, applicable over all domains where information representation was used. A lengthy
analysis of one particular domain, that of visual programming languages, by Green and Petre (1996), is
the most detailed presentation of the framework currently published. For more details, see the section
entitled Further Reading.
What do they work on?
The cognitive dimensions approach applies to both interactive and static systems to anything, as long
as its an information artefact. Information artefacts are the tools we use to store, manipulate, and
display information. They comprise two classes:
interactive artefacts, such as word-processors, graphics packages, mobile telephones, radios,
telephones, central heating control systems, software environments, VCRs, ....
non-interactive artefacts, such as tables, graphs, music notation, programming languages, etc.
In this respect the approach is unlike most evaluation methods in HCI, which are solely aimed at
interactive devices, such as editors and the like. The focus of those approaches is on the process of
interaction. Therefore, they do not work well on non-interactive artefacts.
Discussion tools, not deep analyses
Cognitive dimensions are discussion tools. They give names to concepts that their users may only
have half-formulated, or may not have formulated at all.
Giving names to concepts (lexicalisation) is essential to serious thought and discussion. Once names are
given, discussants do not have to explain each idea every time they use it; they can refer to the idea
knowing that the other party will understand it. Paradigmatic examples help to make the concepts
concrete. Comparisons of trade-off costs become easier. Above all, a checklist can be constructed, to
make sure that an evaluative discussion has not overlooked important aspects.
Because these dimensions name concepts that are familiar, even if not well-formulated by most users,
the approach is easy to grasp. Learning to think fluently in these terms may take practice, but the same
is true of any conceptual system.
Cognitive dimensions will not give their users powerful mathematical tools for deep analysis. Such
analytical tools have plentiful uses, but they will never serve for broad-brush, non-specialist use. And
because the cognitive dimensions approach covers a wide range of very different dimensions, an
equally wide variety of mathematical analyses would be necessary to achieve the same coverage by
purely analytical techniques; there would need to be techniques drawn from parsing theory, from
6
analysis of information structures, from human reliability modelling, perceptual theory, and so on. The
result would be horrendous, a system that was too hard to learn, too slow to use, and too opaque to talk
about to non-specialists.
How do Cognitive Dimensions relate to more traditional approaches?
The cognitive dimensions analysis is a very different philosophy from such approaches as GOMS and
the Keystroke Level Model (Card et al., 1981) and Heuristic Evaluation (Neilsen and Molich, 1990)..
The Keystroke Level Model delivers predictions of overall time taken for a specified task. A detailed
task analysis is needed, usually based on GOMS, a considerable investment in time and expertise. The
value of task analysis has been much debated; users often display inventive unexpected approaches
that were not conceived by the task analysts.
Heuristic Evaluation identifies poor screen design, badly-phrased wording, and the like. Much quicker
than task analysis, it needs the services of up to three HCI experts viewing the displays of the fully-
implemeted artefact or of a reasonable mock-up. There is no theoretical foundation, but within its scope
it produces comprehensible and helpful suggestions. The approach has no way to consider effects of the
underlying information structure, which is a serious shortcoming.
The Cognitive Dimensions approach should be seen as complementary to other approaches. It is easy to
learn and quick to use; it identifies sub-devices and activity types, unlike other approaches; and it can
be applied at any stage of design, from idea to finished artefact. Because the ideas are so simple, they
make sense to ordinary users as well as to specialists.
Cognitive Dimensions traditional approaches
broad-brush highly detailed
quick to learn specialist training needed
quick to apply lengthy analysis
applicable at any stage of design requires full task analysis (GOMS/KLM) or fully-
implemented design or mock-up (Heuristic Eval)
differentiates user activity types all activity evaluated identically
multi-dimensional single dimension
vague precise metric
comprehensible to non-specialists only the metric is comprehensible - not the basis for it
Figure 1 Comparison of Cognitive dimensions and traditional evaluative approaches.
What does the approach deliver?
The CDs approach avoids any kind of simplified bug hunting or overall difficulty measure. Instead,
the information artefact is evaluated along 13 different dimensions, each of which is cognitively
relevant, giving a profile. The profile determines the suitability for various tasks.
The dimensions are not intrinsically good or bad. Different types of user activity are best supported by
different profiles.
Four types of activity
Designing a garden is not like designing a bridge. Garden design is very aesthetic, and part of the job is
exploring the options, changing ones mind, feeling ones way to a new vision. Bridge design is less
inspirational and very much more safety-critical. These are different types of user activity, and they are
likely to be supported by different kinds of software or drafting tools.
7
It would be nonsense to claim that all information artefacts should satisfy some single set of criteria. We
shall distinguish four main classes of activity, and as we explain the framework we shall develop a
preferred profile for each type of activity. The four types we distinguish are:
incrementation: adding a new card to a cardfile; adding a formula to a spreadsheet
transcription: copying book details to an index card; converting a formula into
spreadsheet terms
modification: changing the index terms in a library catalogue; changing the layout of a
spreadsheet; modifying the spreadsheet to compute a different problem
exploratory design typographic design; sketching; programming on the fly (hacking)
Figure 2 The four types of activity distinguished in this framework
These activity types are the pure and undisturbed cases. Sometimes the situation imposes extra
demands for example, the forms-based device illustrated in Figure 18 (page 37) is intended for
incrementation, but the poor design forces users to hut through their previous entries in a way that
approaches the problems of exploratory design.
Each activity is best supported by its own profile of CDs. The profiles will be gathered together as the
discussion proceeds. (If impatient, look ahead to Figure 19 on page 42.)
Why cognitive dimensions?
The approach deserves to be called cognitive because it focuses on usability aspects that make learning
and doing hard for mental (not physical) reasons. Button size, for instance, is a physical, not a
cognitive, issue although it has been included in certain versions of cognitive user modelling. On the
other hand, the degree to which the system makes users translate one operation in the head into a
multitude of individual actions is assuredly cognitive, since it relates to the users conception of an
operation (and ironically, this factor is not included in most versions of cognitive user modelling).
Aesthetic or emotive aspects of usability, important as they are, have no place in this framework. Nor
does the analysis of context-of-use, an aspect of usability whose importance is increasingly becoming
recognised. These omissions are not implicit assertions that they have no relevance. Our belief is quite
the contrary, but this particular framework does not lend itself readily to being extended in those
directions, which we believe need to be analysed in other ways.
More problematic to some is why we believe the term dimension to be appropriate. Comparison with
the dimensions of physics is knowingly invited by this choice, justified, in our opinion, by some
important similarities:
the cognitive dimensions are conceptual idealisations, just like velocity or potential energy
for our purposes, entities are fully described by their position on the various dimensional scales
the dimensions are conceptually independent, just as mass is conceptually independent of
length and time
for real entities, the dimensions are only pairwise independent, meaning that although any
given pair can be treated as independent, a change on one dimension usually entails a change
on some other dimension, in a way that we shall describe shortly.
An important corollary follows from the last point, namely that changing the structure of an artefact is a
matter of trading-off one disadvantage against another: just like physics, cussedness is constant.
Trade-offs and Pairwise Independence
Take a constant mass of gas, with 3 dimensions: temperature, pressure and volume.
If you heat the gas, the temperature rises, and it tries to expand. If the volume is held constant, it cant
expand, so the pressure increases. If the pressure is to be kept constant, the volume must increase.
8
So although pressure, temperature, and volume are conceptually independent, for physical objects they
are only pairwise independent. Within reason, any combination of pressure and temperature may be
obtained, but between them they determine the volume; and likewise for all other combinations.
Many of the cognitive dimensions are similarly pairwise independent: any two can be varied
independently, as long as some third dimension is allowed to change freely.
The structure of information artefacts
Before explaining the dimensions themselves we need to explain the model of information artefacts and
their construction and manipulation.
To present the framework it is necessary to introduce the concepts of notation, environment and
medium. To motivate these concepts we shall briefly present one example of a dimension, the most
convenient being viscosity.
A viscous system is one that is hard to modify. (This is over-simplistic at least two types of viscosity
can be distinguished but it will serve for the illustration.)
Text written with a word-processor is easy to modify; text written with a pen and paper is
much harder to modify. These are different environments.
Inserting a new paragraph into a novel is much easier than inserting a new page into a very
formal document with numbered paragraphs and manifold cross-references. These are different
information structures, or notations.
Changing a paragraph or even a word is much easier on paper, where the whole text is visible
and any part of it can quickly be reached, than when it is being dictated to a scribe, when only
the most recent word can easily be changed (sorry, Ill say that again). These are different
media.
The notation comprises the perceived marks or symbols that are combined to build an information
structure. The environment contains the operations or tools for manipulating those marks. The notation
is imposed upon a medium, which may be persistent, like paper, or evanescent, like sound.
The cognitive dimensions assess the process of building or modifying information structures. Since that
process is affected by all three of the notation, the environment, and the medium, it is necessary to take
them all into account. A notation cannot be said to possess viscosity; only a system, comprising all
three, can be viscous. (In what follows we shall assume that the medium is persistent, unless specifically
stated otherwise.)
The interpretation of this framework is immediately apparent if we restrict ourselves to non-interactive
artefacts, such as a programming language: the notation is the language itself, the environment is the
editing system, and the medium is normally the computer screen (but occasionally we may have to give
programming support by telephone, and then the differences between media become very marked).
Interaction languages as a form of notation
It may not be so easy at first to see how to apply the framework to interactive artefacts, such as using a
telephone. Consider first the information structure that is to be built; we can take it to be the number
that is to be dialled. The notation is the set of digits, possibly including the star and hash keys. The
environment may simply allow each number on the keypad to be pressed, or it may include a redial
last number feature; there is also an action (replacing the handset) that cancels everything done so far.
The medium is a persistent one, in that the telephone number is being constructed inside the memory of
the telephone system.
Thus, we model all information artefacts in terms of notations and their use. Non-interactive artefacts,
such as timetables, are directly expressed in the notation. Interactive artefacts are controlled by an
interaction language whose commands are the notation.
9
Layers
Now consider a slightly more complicated example, building a program.
There are two layers. In one layer, the notation is the programming language, the environment is the
editor, and the medium is the computer memory, which is persistent, since the program is built up
there.
The other layer is the editing layer. Here, the notation is keystrokes and mouse actions, the environment
is the keypad, and the medium is transient the notation is translated into editor actions, such as
selecting a word or deleting a character in the program.
Each layer has its own dimensions. Viscosity of the program text is determined by the programming
language and by whether the editor has tools specifically designed for that language. Viscosity of the
editor layer is determined by how a user intention, such as selecting a given word, is expressed as a
sequence of actions, and whether it is easy to recover from a slip in the execution of those actions.
A serious analysis of layers is beyond the scope of what can be covered in this tutorial, but it is
necessary at times to be aware of the different layers involved or to risk analysing the wrong level of
interaction.
Sub-devices
The last component of the framework is the notion of a sub-device. A sub-device is a part of the main
system that can treated separately because it has different notation, environment, and medium.
The simplest sub-device might be a pad of paper, considered as part of an information system whose
main components are software. Its function might be to record notes about the main information
structure, such as list of identifiers not easily obtained from the main environment; or its function might
be to allow work to be sketched out before being transcribed to the main system not infrequently,
pencil and paper is a better system for exploratory design than the main system for which the design is
intended.
One particular type of sub-device must be mentioned here, even though it will not receive full
treatment until considerably later: we call it the abstraction manager. Any system that allows users to
create new terms, such as macros, needs some kind of manager to record the names and definitions of
the macros and perhaps to allow them to be edited. Even a humble telephone with an internal memory
of ten numbers needs a simple manager so that the user can associated a quick-dialling code with some
oft-used telephone number. The sub-device in these cases builds persistent structures out of the
notation of actions (the commands that the macro executes or the digits of the keypad), but it does not
execute the actions that is the job of the host device.
10
Summary of the framework
The Cognitive Dimensions framework gives a broad-brush usability profile, intended to provide
discussion tools rather than detailed metrics. The 13 or so dimensions can be applied to any type of
information device, whether interactive or not, allowing similarities to be detected in very different
types of device.
All forms of interaction are seen as building or modifying an information structure. Non-interactive
(e.g. paper-based) information, such as the structure of a timetable, is assessed in terms of the ease of
creating or changing it. The operation of interactive devices is seen as creating commands in an
interaction language; these commands are the information structures that are to be built or modified.
Usability depends on the structure of the notation and the tools provided by the environment. Thus, the
dimensions apply to the system as a whole, taking into account the tools provided as well as the
notation or the information structure, They also depend on what persists: in a conventional program,
the names of actions persist and can be edited, but in an interactive device the effects of actions persist
rather than their names.
Some forms of evaluation are geared towards bug-hunting or towards predicting time to perform a
task. The cognitive dimensions approach is not so simplistic. The dimensions are not intrinsically good
or bad; together they define a hyperspace, and for particular types of user activity, different regions of
that hyperspace are appropriate.
Four types of user activity are distinguished in this document:
incrementation adding further information without altering the structure in any way (e.g.
entering a new file card);
modification changing an existing structure, possibly without adding new content;
transcription copying content from one structure to another structure;
exploratory design combining incrementation and modification, with the further characteristic
that the desired end state is not known in advance.
Each type of user activity will be best served by a different set of values on the dimensions.
The dimensions are supposedly independent, in the sense that for a given system, it is possible to
redesign it to change its position on one given dimension A without affecting its position on another
dimension B. Independence is only pairwise, however almost certainly some other dimension C will
be affected (and vice versa, A can be changed without changing C, but something else, e.g. B, will have
to change). Redesign is therefore an exercise in choosing trade-offs and making compromises (see
below).
11
Cognitive Dimensions: the full set
This list of cognitive dimensions has been arrived at by exploring a variety of devices and looking for
important differences. There is no guarantee that it is complete or fixed.
In this document we shall only explore a subset, namely those that are less psychological in character.
Others, such as error-proneness and role-expressiveness, demand more understanding of cognitive
psychology than it would be convenient to present here.
Dimensions treated in this tutorial
Abstraction types and availability of abstraction mechanisms
Hidden dependencies important links between entities are not visible
Premature commitment constraints on the order of doing things
Secondary notation extra information in means other than formal syntax
Viscosity resistance to change
Visibility ability to view components easily
Dimensions not treated in this tutorial
Closeness of mapping closeness of representation to domain
Consistency similar semantics are expressed in similar syntactic forms
Diffuseness verbosity of language
Error-proneness notation invites mistakes
Hard mental operations high demand on cognitive resources
Progressive evaluation work-to-date can be checked at any time
Provisionality degree of commitment to actions or marks
Role-expressiveness the purpose of a component is readily inferred
12
Viscosity
Definition
Resistance to change: the cost of making small changes.
Repetition viscosity: a single goal-related operation on the information structure (one change 'in the
head') requires an undue number of individual actions,.
Knock-on viscosity: one change 'in the head' entails further actions to restore consistency
Thumbnail illustrations
Repetition viscosity: Manually changing US spelling to UK spelling throughout a long document
Knock-On viscosity: inserting a new figure into a document creates a need to update all later figure
numbers, plus their cross-references within the text; also the list of figures and the index
Explanation
Viscosity in hydrodynamics means resistance to local change deforming a viscous substance, e.g.
syrup, is harder than deforming a fluid one, e.g. water. It means the same in this context. As a working
definition, a viscous system is one where a single goal-related operation on the information structure
requires an undue number of individual actions, possibly not goal-related. The goal-related operation
is, of course, at the level that is appropriate for the notation we are considering. Viscosity becomes a
problem in opportunistic planning when the user/planner changes the plan; so it is changes at the plan
level that we must consider. This is important, since changes at a lower level can sometimes be pure
detail (inserting a comma, for instance), giving a misleading impression of low viscosity. So, in general,
viscosity is a function of the work required to add, to remove, or to replace a plan-level component of an informa-
tion structure.
Note that viscosity is most definitely a property of the system as a whole. Certain notations give the
potential for viscosity (e.g. figure numbers) but the environment may contain tools that alleviate it by
allowing aggregate operations (e.g. a smart word-processor with auto-number features). See below for
associated trade-offs.
Also note that viscosity may be very different for different operations, making it hard to give a
meaningful overall figure.
Cognitive relevance
Breaks train of thought, gets too costly to change or to take risks; encourages reflection and planning,
reduces potential for disastrous slips and mistakes.
Acceptable for transcription activity and for simple incrementation activity, but problematic for
exploratory design and for modification activity.
transcription incrementation modification exploration
viscosity acceptable
(because one
shouldnt need to
make changes)
acceptable harmful harmful
13
Cost Implications
It is easy for designers to view a device as a pure incrementation system and overlook the need for
restructuring. This is a serious and far too frequent mistake. So a computer-based filing system may readily
allow new categories to be added, but have very little provision for redefining categories and re-
categorising the objects within them.
The worst problems come when viscosity is combined with premature commitment. The user has to
make an early guess (thats the premature commitment) but then finds that correcting that guess is
time-consuming or costly (thats the viscosity).
Types and Examples
Repetition viscosity is a frequent problem in structures which exist in the users mind rather than being
recognised by the system. A collection of files making up one document in the users mind may need
to be edited to bring their typography into conformance, usually by editing each file individually,
because few systems have the ability to apply the same editing operations to each file in a list.
Knock-on viscosity is the other main type, frequently found in structures that have high inter-
dependencies, such as timetables:
9.00-10.00 10.00-11.00 11.00-12.00
Year 1 Mr Adams Ms Burke Ms Cooke
Year 2 Ms Cooke Mr Davis Mr Adams
Year 3 Ms Burke Mr Adams Mr Davis
Figure 3 A timetable showing which teacher is teaching which class. Timetables
typically contain high knock-on viscosity. If Mr Adamss Year 1 class has to be moved
to the 10-11 slot, everything else will have to be re-organised.
In practice, potent combinations of knock-on and repetition viscosity are the type to be feared, turning
up frequently in graphics structures. Although word-processors are steadily reducing the viscosity
problem for document generation, drawing packages are not making comparable progress. Editing a
drawing usually requires much tedious work, and frequently many similar alterations need to be made
to different parts of the picture; automation tools, desirable as they might be, are not yet commercially
available.
Figure 4 This tree-chart, showing part of the JavaScript object hierarchy, illustrates
both knock-on viscosity and potential repetition viscosity: if the window box is moved
to one side, all the lines connecting it to other boxes will have to be moved (knock-
on). In most drawing systems each line will have to be redrawn individually (repetition
viscosity).
14
Drawing a genealogical tree for any reasonably extensive family will well illustrate the difficulties.
Many programming languages exhibit pronounced viscosity. This is a serious problem, because
software is always changing during development and also after it is supposed to be finished. Many of
the trends in programming language development have therefore been aimed at reducing viscosity.
One of the most recent is object-oriented programming, which allows a single change to be made in a
parent class, thereby altering the behaviour of many different subclasses. Unfortunately this has the
side effect of introducing a new kind of viscosity - if the interface to a class changes, rather than its
behaviour, this often results in knock-on changes to its parent class, as well as to all its siblings, and the
contexts in which they are invoked (even though they may be only tenuously related to the original
change).
So far we have described non-interactive examples, but the viscosity concept is readily extended to
interactive systems. Think of using devices as uttering sentences in an interaction language. For non-
interactive devices (i.e. those where the names of actions persist and their execution is delayed, such as
programs and sheet music) viscosity obviously consists in the work required to reshape the actions. For
interactive devices, in which the actions are executed as soon as named (e.g. dialling a telephone,
finding a given radio programme) viscosity consists in the work required to get from one state of the
device to another state: for the telephone, after a slip of the hand has caused a wrong digit to be dialled,
the only remedy is to cancel the transaction and restart; for the radio, changing to a different
programme may be easy or difficult (see example below).
Other contexts
Hypertext structures, although easy to create, can become extremely viscous. Fischer (1988), describing
what appears to be the process of opportunistic design that takes place while creating a collaboratively-
generated hypertext structure, writes ... For many interesting areas, a structure is not given a priori
but evolves dynamically. Because little is known at the beginning, there is almost a constant need for
restructuring. [But] despite the fact that in many cases users could think of better structures, they stick
to inadequate structures, because the effort to change existing structures is too large.
Tailorable systems seem to create opportunities for viscosity at the organizational level. Examples of
these systems include customisable settings and macros for spreadsheets and word-processors, and
other systems allowing users to tailor their office environment. The problem here is that users lose the
ability to exchange tools and working environments; moreover, the arrival of a new release of the
software may mean that it has to be tailored all over again, possi bly with few records of what had been
done before.
Indeed, many organizations are repositories of information among many people (distributed
cognition) and the organizations themselves can be therefore be considered as information structures.
Changing a component of the information held collectively by the organisation may be a matter of
telling every single person separately (Repetition Viscosity), or it may, in a hierarchical or bureaucratic
organization, only need one person to be told. The levels of the hierarchy have the same role as the
abstractions in other systems.
Viscosity growth over time
Finally, diachronic aspects of viscosity. In many design systems, the viscosity of any one structure or
design tends to increase as it grows. In the context of CAD-E design, one workplace has reported that
their practice is to explicitly declare a point after which a design can no longer be changed in any
substantial way. At that point the design is said to have congealed, in their terminology, and
opportunistic design must be restricted to details. There is at present no reported data, to my
knowledge, describing in detail the time course of viscosity in different systems. The impression is that
certain programming languages, such as Basic, have low initial viscosity, but as programs grow their
viscosity increases steeply. Pascal starts higher but increases more slowly, because more is done in the
initial abstractions. Object-oriented systems are intended to preserve fluidity even better, by using late
15
binding and the inheritance mechanism to minimise the effort needed to modify a program. Some of
the investigations made under the banner of software lifecycle studies may prove to be revealing.
Workarounds, Remedies and Trade-offs
Before considering possible remedies, observe that viscosity is not always harmful. Reduced viscosity
may encourage hacking; increased viscosity may encourage deeper reflection and better learning.
Reduced viscosity may allow small slips to wreak disaster; for safety-critical systems, high viscosity
may be more desirable.
Nevertheless, viscosity is usually to be avoided, and some of the possible manoeuvres are:
$ The user decouples from the system: the workaround for unbearable viscosity (especially when combined
with a need for early or premature commitment) is to introduce a different medium with lower
viscosity to make a draft version, and then transfer it to the original target medium. Thus, the process is
now two-stage: an exploratory stage followed by a transcription stage.
Thus, before drawing in ink, an artist may make a sketch in pencil, which can be erased. A programmer
may work out a solution on paper before typing it.
Alternatively, in some circumstances users may decouple by adopting a different notation instead of a
different medium. The programmer may choose to develop a solution in a program design language
before transcribing it into the target language.
$ Introduce a new abstraction. This is the classic remedy. Got a problem with updating figure numbers? OK,
well create an AutoNumber feature. See below for the trade-off costs of abstractions.
$ Change the notation, usually by shifting to a relative information structure instead of an absolute one.
Consider a list of quick-dial codes for a mobile telephone:
1 Aunt Mary
2 Neighbours
3 Take-away
4 Office
5 Car rescue service
Figure 5 Modifying the order of entries in a telephone memory is usually a high-
viscosity operation
I want to be able to dial Car rescue service very quickly indeed, let us say. That means moving it to the
top of the list, because on this model of mobile phone the dialling codes are scrolled one by one until
the required target is visible. I want to move that one item while keeping the others in their existing
order. Almost certainly, on existing models, I will have to re-enter the whole list; but in a good outliner
system, I could perform exactly the same modification in a single action, by selecting the Car rescue
entry and dragging it upwards:
Figure 6 Order modification in an outliner is a low-viscosity operation. The highlighted
item (car rescue service) is being dragged upwards; when released, it will take its
place where the horizontal arrow points, and all lower items will be displaced
downwards.
16
Many other dimensions can affect viscosity, but it will obviously be easier to discuss such trade-offs
when we reach the dimensions concerned.
17
Hidden Dependencies
Definition
A hidden dependency is a relationship between two components such that one of them is dependent on
the other, but that the dependency is not fully visible. In particular, the one-way pointer, where A
points to B but B does not contain a back-pointer to A.
The search cost of a hidden dependency structure is a measure of the effort required to expose a typical
dependency. Search cost is a function of the length of the trail, the amount of branching, and the effort
required to follow each link.
Thumbnail illustration
HTML links: if your page is linked to someone elses, e.g. the HCI 98 site, how will you know if and
when that page is moved, changed, or deleted?
Many links are fossils out of date pointers to pages that have been deleted or moved. Because checking
links is slow, the search cost for testing integrity of a site is quite high, so fossils are likely to increase
over the years.
Explanation
A dependency can be hidden in two different ways, by being one-way or by being local. The HTML link
is a one-way pointer, so it is invisible from the other end I cannot find out what links to my page have
been created, unless I use a search engine. As it happens, HTML links are also local: they point to the
next page on a chain, but say nothing about what lies beyond.
Spreadsheet links are just the same in structure. Here their local nature is more important, because I
may well want to know which cells use the value in a given cell, whether directly or indirectly.
Cognitive relevance
Hidden dependencies slow up information finding. If the user wishes to check on a dependency and
has to find it by a lengthy search process, the job of building or manipulating an information structure
will take much longer; the risk of error will increase; and the train of thought will be broken, adding
again to the time and unreliability of the process. When dependencies form long trails and there may
be several branches at each step, the job of searching all the trails becomes very onerous, and users are
tempted to make guesses and hope.
Exploratory design can tolerate hidden dependencies for small tasks, but they are problematic for
modification activity. Transcription activity is probably not much affected except when mistakes are
made.
transcription incrementation modification exploration
hidden dependencies acceptable acceptable harmful acceptable for
small tasks
18
Cost Implications
Hard to estimate, but hidden dependencies are apparently responsible for a very high error frequency
in spreadsheets, possibly because the spreadsheet programmer hopes for the best rather than
scrupulously checking out all the consequences of a change.
Types and Examples
Dependency pointers can be one-way or symmetric, local or wide, and hidden or explicit.
One-way and symmetric pointers
A GOTO statement in a programming language is a one-way dependency, because at its target there is
no corresponding COME-FROM. A flowchart also contains GOTOs but they are represented
symmetrically; the arrow shows which commands can be the targets of GOTO commands. By using
indentation (and restricting the range of control constructs) textual languages can be given symmetric
pointers:
1: J = 1
2: IF J > 10 GOTO 6
3: J = J + 1
4: (compute)
5: GOTO 2
6: (continue)
for J=1 to 10
{
compute
}
(continue)
Figure 7 One-way pointers are shown in the code on left. The code on the right
contains implied pointers that are symmetric.
Local and distant pointers
A local pointer only leads to the immediate target; a distant pointer gives a view of the farther reaches
of the dependency. A file-structure displayed on a Unix platform by using the ls command shows
the children of a given directory, but it does not show the grandchildren, nor the grandparents. The
same file-structure displayed as a tree shows the entire structure.
Hidden and explicit pointers
An explicit pointer is one that is shown in the normal viewing state. A hidden pointer is not shown in
the normal viewing state, and indeed may not be shown at all, so that it has to be searched for.
Examples
Spreadsheets are a mass of one-way, local dependencies. Even in the formula view, as illustrated, we
cannot detect, by looking at B1, that it is used in F2. Some modern spreadsheets contain tools to analyse
dependencies, but early versions had none, so that altering a cell in someone elses spreadsheet was a
risky business.
The search cost for spreadsheets rises sharply with the size of the sheet and with the branching factor
(average number of data-supplying cells quoted, per formula).:
A B C D E F
1 item unit price units price tax total price
2 apples 0.84 2 =B2 * C2 0.29 =E2 * D2
Figure 8 Spreadsheets illustrate hidden dependencies.
Certain types of visual programming language make the data dependencies explicit, as shown, thereby
accepting different usability trade-offs:
19
2
0.84
*
*
+
0.29
units
price
tax
total
Figure 9 Data flow representations make the dependencies explicit.
Styles, object hierarchies, and other abstractions usually impose hidden dependencies on their users:
change a style definition, and who knows what else will change in your document.
Contents lists and similar lists obviously contain hidden dependencies. Every now and again we all
forget to update a contents list or its equivalent.
Surprisingly, contemporary car radios contain hidden dependencies: they contain a feature for traffic
information, so that as you are driving announcements can cut in over the programme you are listening
to, to warn of congestion or other problems. These announcements come from a different station, so
they may be louder or quieter (or in any case you may have been playing a tape). To avoid getting your
ears knocked off, there is a feature for pre-setting the maximum volume. An interesting problem for the
user, because pre-setting the maximum volume is (a) a problem in premature commitment, (b) has no
chance of progressive evaluation, and (c) is a hidden dependency -- when Im driving and vaguely
listening to a soothing bit of Radio 3 and suddenly a geezer starts banging on at about 100 dB about
traffic delays in some place Ive never heard of, I want shot of him NOW, not after 30 seconds of
fiddling around. The preset is a hidden dependency, because there is no link back from the instruments
present mode to the mode that allows me to change the maximum allowed volume.
Conventional programming languages are built upon hidden dependencies. Assignments made in one
statement are then used elsewhere; a simple form that usually causes few problems, because the search
cost is not too high (except in the notorious side-effect, where values were surreptitiously altered in
unexpected places). Type declarations can sometimes create hidden dependency problems.
Scripting languages have now become common adjuncts of advanced applications. Event-based
versions, such as Visual Basic, HyperCard and Toolbook, are notorious for hidden dependencies,
because they spawn little packets of code attached to various buttons and other graphics,
communicating by shared global variables. Figuring out which packet of code does what when becomes
a nightmare. Despite which, they are very popular for some purposes, because the trade-offs are
favourable for some types of activity see below.
Fossil deposits
Fossils, mentioned in connection with broken web links, are becoming a noticeable feature of operating
systems. Unix platforms use dot files for various purposes, typically to contain preference data for
editors and similar utilities; these are hidden dependencies, because various programs may expect to
find a particular dot file in place, and crash if it is not there. Users may find that their directories contain
a large number of dot files whose purpose is obscure, but which they hesitate to delete in case
something does need them. PC platforms under Windows have similar large deposits of fossils, notably
out-of-date DLLs.
In general information structures that cause a build-up of fossils are poor choices for long-term use,
since they will obviously suffer from mature disfluency that is, they will get harder to use as time
goes on.
Workarounds, Remedies and Trade-offs
$ Add cues to the notation: To avoid hidden dependencies in programming, a number of designs have been
proposed. Among these are the scrupulously neat languages like the Modula family, in which the
technique for encapsulating code into functions or procedures requires very explicit statements of what
identifiers are publicly visible through the wall.
20
$ Highlight different information: The dataflow-based visual programming language illustrated above
exposes the data dependencies as the central feature of the notation, a different paradigm that accepts
an entirely different set of trade-off positions from say, a C version of the same algorithm.
$ Provide tools: Modern spreadsheets can highlight all cells that contribute to the value in a particular cell,
or alternatively all cells that draw upon the value in a given cell. This has its uses, although it has to be
noted that the visibility is reduced. The dependency information for two different cells is not easy to
compare, for example.
Do not rush to eliminate hidden dependencies. Among the most popular programmatic creations have
been Basic, especially Visual Basic; spreadsheets; and the event-based family of HyperCard, Toolbook,
et al. All of these are high on hidden dependencies.
Evidently, hidden dependencies are quite acceptable in appropriate circumstances. The languages
mentioned are typically used for quick and dirty coding or for exploratory design. Both strategies mean
getting your ideas down quickly. A good designer can accept a high mental load, remembering what
bits depend on what, until the main structures are in place, but cannot accept constant fiddly
interruptions. So an incremental system, like a spreadsheet, is good for quickly assembling some
material and some formulas. During that design phase, because the links are only one-way, the
viscosity is lower; if the links were two-way, then every time a new addition was tried and then
corrected, some links would need to be adjusted, slowing up the creative process. Architects seem to
work the same way.
The problem of exploratory design comes in the later stages, when the structure tries to turn into a
congealed mess. Experienced designers probably know how far they can go without cleaning up. For
lesser mortals, environments that can document links automatically would be useful.
If hidden dependencies are to be reduced, consider the costs of the remedies. Remedy (i), where the
information is made explicit by cueing, as in the Modula family, is probably more suited to a system
designed for transcription activity than for exploratory activity. The remedy makes the notation much
more viscous. Remedy (ii), where the notation is focused upon the dependencies, creates a more diffuse
language, sometimes with its own problems of premature commitment (where shall I draw the next
operation?) etc. Remedy (iii) has poor juxtaposability.
21
Premature Commitment and Enforced Lookahead
Definition
Constraints on the order of doing things force the user to make a decision before the proper information
is available.
Thumbnail illustration
The amateur signwriter at work:
Explanation
The problems arise when the target notation contains many internal constraints or dependencies and
when the order constraints force the user to make a decision before full information is available
(premature commitment) or to look ahead in a way that is cogniviely expensive (enforced lookahead).
The amateur signwriters target notation contains a constraint relating the width of the sign-board and
the width of the wording, but until the sign was written he or she did not know how wide the wording
would be so they made a guess (premature commitment). The professional would paint the sign on a
different surface, measure it, and copy it to the real target (enforced lookahead, using an auxiliary
device).
Cognitive relevance
An obvious disruption to the process of building or changing an information structure.
Problematic in all contexts, except where the lookahead is not extensive. Probably one of the important
differences between novices and experts in studies of design is that experts recognise potential
problems much earlier, perhaps not from looking ahead but by familiarity with likely sources of
difficulty. Research has shown that expert designers frequently treat potential trouble-spots differently,
putting them aside for later treatment or attacking them early.
transcription incrementation modification exploration
premature commitment harmful harmful harmful harmful
Cost Implications
One origin of premature commitment is that the designers view of a natural sequence was at variance
with the users needs. A second possible origin is that the technical aspects of the design made it
necessary to deal with one kind of information before another type, regardless of which was more
22
salient to the user. Either way, the result is that users will have to make second or even a third attempt.
The cost implications will vary from trivial to unacceptable, depending on the circumstances of use.
Types and Examples
Telephone menu systems (If you wish to make a booking, press 1 ...if you wish to reserve a seat, press
2 ... etc) impose premature commitment on the long-suffering user, who has to decide whether to press
key 1 before knowing what other choices are available.
The 4-function calculator is another everyday
example: What keypresses are needed to compute
(1.2 + 3.4 - 5.6) / ((8.7 - 6.5) +
(4.3))
on the calculator illustrated?
Turning to software examples, every amateur has met the problem of choosing categories for a
database and then discovering that the world is full of exceptions. For example, my simple mailmerge
list of addresses allows 6 lines per address. Now and again I meet somebody whose address needs 7 or
more lines. What now? The first design of the database had to be based on guesswork, in other words I
was prematurely committed.
In an example that is closely related in its aetiology but far removed in scale, the structure of a
traditional (non-computerised) filing system is usually susceptible to premature commitment. If the
filing system is to hold documents that already exist, the problem is less acute, but if the documents do
not yet exist, hard choices have to be made (as they say). Deweys decimal classification was a brave
effort to categorise the world, whose effectiveness has been seriously marred by the later development
of whole new branches of knowledge which were unknown to him.
Old library buildings which have left no space for modern knowledge areas have an even worse
problem.Designed before the development of biochemistry or information technology, their science
shelves are bursting out all over.
Surreptitious order constraints
But even when the world of possible instances is reasonably well understood, the system can make
problems by denying users the opportunity to choose their own order of actions. Consider designing a
graphic editor for entity-relationship diagrams; most of us would probably provide a facility to define
new entities, and a facility to define new relationships between existing entities. The users choice of
actions appears to be free entities can be added, deleted, or modified at any time, and relationships
between existing entities can likewise be added, deleted, or edited at any time.
23
Figure 10 A box-and-line editor (e.g. for entity-relationship diagrams). These editors
frequently impose a type of order constraint: users can add boxes with no connectors,
but cannot add connectors except between existing boxes. The dashed line is an
example of a free connector, that could not be drawn.
Yet that system imposes an order constraint. If the user wants to say OK, this entity has to have a
1:many relationship with something, but I dont yet know what, there is no way to record that fact.
The design only permits relationships to be defined between pre-existing entities thereby making the
implementation much easier, no doubt, but imposing on the user he burden of remembering dangling
relationships until the associated entities have been defined.
Effect of medium
Premature commitment is more strongly affected than other dimensions by the choice of medium.
When marks are transient, as with auditory media, it is essential to help the user make good choices by
providing ample information. When the marks are persistent, as with paper, the user usually has time
to make a choice by juxtaposing them and choosing the best. Thus, even when the telephone menu is
too complex to be usable, the same information structure could be handled facilely and routinely if the
medium were paper or some other persistent medium.
Workarounds, Remedies and Trade-offs
$ Decoupling: As with viscosity, the obvious workaround is to use an intermediate stage to decouple from
the immediate problem. The amateur signwriter should paint the sign somewhere else, measure it, and
then allow adequate space.
$ Ameliorating: Premature commitment is less costly when viscosity is low, since bad guesses can be easily
corrected, so one remedy is to reduce the viscosity in the system.
$ Deconstrain: The ideal remedy is to remove constraints on the order of actions. This has been one of the
contributions of GUI interfaces.
24
Abstractions, Abstraction Hunger, and the Abstraction Barrier
Definition
An abstraction is a class of entities, or a grouping of elements to be treated as one entity, either to lower
the viscosity or to make the notation more like the users conceptual structure.
The abstraction barrier is determined by the minimum number of new abstractions that must be
mastered before using the system; if the system allows one user to add new abstractions that must then
be understood by subsequent users, that will further raise the abstraction barrier.
Abstraction-hungry systems can only be used by deploying user-defined abstractions. Abstraction-
tolerant systems permit but do not require user-defined abstractions. Abstraction-hating systems do not
allow users to define new abstractions (and typically contain few built-in abstractions).
Thumbnail illustrations
Styles in word processors: an example of abstraction-tolerance. The Z specification language: an
abstraction-hungry system with a high initial abstraction barrier. Spreadsheets: an abstraction-hating
system.
Explanation
The characteristic of an abstraction from our point of view here is that it changes the notation. Almost
invariably, the notation is changed by expansion a new term is added. In the specification language Z,
the off-the-shelf notation contains a large number of primitives, which can be assembled to form
abstract specifications of parts of a devices behaviour; the abstractions can then be assembled to form a
complete specification of the device.
It may be less apparent how abstractions change the notation for interactive devices. Think of
interacting with them as uttering sentences in an interaction notation. The sentences for setting all
level 1 headings to 24-point bold would include commands to find next heading, select it, set size to 24,
and set style to bold. By creating a style called Heading 1, an alternative notation is created select
Heading 1 and set size and style.
In the same way, the off-the-shelf notation for dialling telephone numbers comprises the digits 0-9. If
the telephone has a memory the notation can be changed to include E1 (meaning dial 0123-456-7890,
say), and so on.
Cognitive relevance
Abstractions make notations more concise and sometimes they make them fit the domain concepts
better. They are an investment in the future: by changing the notation now, it will be easier to use in the
future, or it will be more re-usable, or it will be easier to modify whats built in the notation because the
abstractions can reduce the viscosity.
Using abstractions is costly. First, they are hard. Thinking in abstract terms is difficult: it comes late to
children, it comes late to adults as they learn a new domain of knowledge, and it comes late within any
given discipline. Systems that enforce abstractions will be difficult to learn and may permanently
exclude some possible users. (We can see why they are hard from machine learning theory, where the
constant problem is to avoid either over-generalisation or over-specialisation.)
Second, they must be maintained (see the section on abstraction management). Creating and editing
them takes time and effort and has potential for mistakes.
25
Thus, abstractions are potentially useful for modification and transcription tasks. They are good for
incrementation tasks if they fit the domain well. For exploratory tasks, they are problematic. Highly
skilled designers can use them, but they are most likely to be exploring the abstractions needed as well as
exploring for a solution given the abstractions. Less skilled designers working with smaller problems
typically prefer to save the investment, even at the expense of using a more long-winded notation and
building structures that will be more viscous.
That is equally true of interactive and non-interactive systems. Less skilled word-processor users will
avoid styles, auto-formatting, macros, advanced (pattern-matching) editing, and other forms of
abstraction.These features could save effort, but they are error-prone and cognitively costly.
transcription incrementation modification exploration
abstraction hunger useful useful (?) useful harmful
Cost Implications
Programming languages often use many abstractions (procedures, data structures, modules, etc)
although spreadsheets use rather few. Everyday domestic devices tend to use rather few abstractions,
although telephones with built-in memory systems have a form of abstraction, some central heating
controllers use simple abstractions (for weekdays/weekends. holidays, etc), cars can or soon will be
personalised so that each driver can have the adjustments for seating, mirrors, radio stations etc.
adjusted individually at one go, and so on
Abstractions have a key role in the future of information devices, because they directly govern many
aspects of ease and because they require abstraction managers, which themselves are sub-devices with
their own usability characteristics.
Abstractions are proliferating, especially in personalised and embedded computers. The next
generation of car controls, including seat adjustments, in-car audio systems, etc., will include advanced
personalisation features: tell the car who you are (in any of various ways) and the seats and mirrors
will be adjusted, your own set of radio stations and preferences will become available, and so on. The
person is a new abstraction in this area. Digital broadcasting will bring many other new abstractions.
We shall repeatedly see tensions between the effort of learning to cope, and the convenience when the
system has been mastered; we shall have problems of preserving and exporting the definition of
Thomas or Alan, when we get a new car; we shall be frustrated by the overhead of having to modify
definitions as our personal characteristics change over time; and it may become almost impossible to
borrow a friends car in a hurry.
It is worrying that very little research has looked at the learnability, usability, and maintenance issues.
The difficult thing about abstractions is trying to get the scope right. An abstraction that describes a
particular set of items is not very useful, because it cannot be used for other similar items in the future.
This is an over-specialised abstraction. An abstraction that describes all possible items is not useful
either, because there is almost no action that one would want to specify for all possible items. This is an
over-generalised abstraction. Good abstractions depend on a well-chosen set of examples, and they also
depend on you knowing what you want to leave out. You often don't know what you want to leave out
at the time you encounter an abstraction barrier, and this results either in confusion or in bad
abstractions
Abstraction managers
If new abstractions can be created, some kind of sub-device must be provided to manage the user-
defined abstractions. This sub-device may or may not suffer from viscosity (it is easy to change an
abstraction); from hidden dependencies (will changing one abstraction impinge on others);
juxtaposability (when Im editing one abstraction, can I compare it others); etc.
26
Types and Examples
Since abstractions operate to change the notation, they have second-order effects with a potential to
vastly increase existing problems. They come in several types.
Persistence
Persistent abstractions, as their name implies, remain available until explicitly destroyed. Transient
abstractions automatically cease to exist, usually after one use.
Style sheets, macros, object hierarchies, grouping in drawing applications, and telephone memories
are persistent abstractions.
Transient abstractions include aggregates selected by pointing, e.g.:
Figure 11 A transient abstraction. Three files in a MacOS folder have been selected
by point-and-click, forming an aggregate that can be trashed, printed, opened, or
copied as one unit. This aggregate is transient, because the next MacOS command
will deselect the files.
Transient abstractions also include global find-and-replace commands, e.g.
Figure 12 Another transient abstraction. This search-and-replace command in Word
will change all hungry felines to fat cats.
Aggregates, definitions, and exemplars
Abstractions may aggregate several entities into a single unit, as the selected folders do, or they may
aggregate several actions into a macro definition. These different contents have rather different
properties (you can remove one entity from an aggregate of entities without affecting the rest: but just
try removing one action from a macro and see what you get!), but for broad-brush purposes we shall
ignore the differences.
Abstractions may also be defined as exemplars, as in many CAD programs where a sub-structure can
be encapsulated and stored in a library to be re-used as required. This allows further sub-varieties: if the
exemplar is changed, will all existing copies of it be updated? Or will the change only affect new
copies?
Word styles are automatic-update abstractions. Redefining one automatically changes all existing
instances of that style in that document (together with instances of other styles that inherit properties
27
from it). Note the potential for hidden dependency here if we cannot readily see which examples will
be changed, we may be in for nasty surprises.
Powerpoint templates, used to define the look of a presentation, are no-update abstractions. If the
template is changed, existing presentations are not automatically updated.
Sometimes, optional-update abstractions are available.
Figure 13 Optional update in a diary system: the time of one instance of the weekly
seminar is to be altered. A dialog gives options of which other instances to update.
Abstraction hunger and the abstraction barrier
Flowcharts are abstraction-hating. They only contain decision boxes and action boxes. Spreadsheets are
more various, but the elementary spreadsheet has no facilities for abstractions.
C and HyperTalk are both abstraction-tolerant, permitting themselves to be used exactly as they come,
but also allowing new abstractions of several kinds to be created. HyperTalk has a low initial level of
abstractions (i.e. a low abstraction barrier), and novices can start by learning how to use fields, then
how to use invisible fields, then how to use variables. C has a much higher minimum starting level, so
that the novice has many more new ideas to learn before being able to write much of a program.
Smalltalk has a high abstraction barrier because many classes must be understood before using it, and
moreover it is an abstraction-hungry language: to start writing a program, you first modify the class
hierarchy. The user must feed it new abstractions. This is a whole new ball-game. Every program
requires a new abstraction; the novice programmer cannot simply map problem entities onto domain
entities. (Of course, experienced programmers have a stock of abstractions they have already learnt,
ready to be applied in a new situation.)
The abstraction management problem has been mentioned above. By and large, abstraction management
subdevices are designed thoughtlessly and poorly. For instance, Word allows users to define new
styles, a very useful abstraction. The styles are held in a hierarchy, each inheriting all the characteristics
of its parent except those that explicitly over-ridden.
Two examples of problems with the Word 5.1 style manager: hidden dependencies and lack of
juxtaposability. (i) Modifying a style will potentially modify all its children i.e. there are
dependencies but these dependencies are completely hidden: the Modification dialog box shows the
parent of a style, but not its children, thus:
28
Figure 14 Hidden dependencies in the Word 5.1 style manager: if I modify the
definition of this style, I cannot tell which other styles will be affected.
(ii) The lack of juxtaposability is imposed by the same dialog box. The modification box illustrates the
style, but I cannot open two modification boxes at the same time. In consequence, almost-identical
styles proliferate.
Workarounds, Remedies and Trade-offs
$ Incremental abstractions: the best workaround is seemingly to provide users with a system that has a low
abstraction barrier but yet contains useful built-in abstractions as short-cuts, and tolerates the addition
of useful new abstractions. Novices can then start by doing operations the long but conceptually simple
way, and master alternative methods later.
The limits of this come when the alternatives create a morass of bewilderment. The latest, fully-
featured applications are always to be distrusted on this score.
$ Trade-offs from abstractions: abstractions are a standard way to reduce viscosity. They can also increase
the protection against error-proneness (for example, by declaring all identifiers, mistypings can be
detected at compilation rather than run-time). Well-chosen abstractions can also increase the
comprehensibility of a language. But remember that the abstractions may have to be chosen almost by
guesswork, so that premature commitment may be substantial and may not be easy to undo.
But abstractions will increase the problems for occasional users and for learners. Learning how to work
with abstractions is difficult, and understanding a structure that contains many abstractions is often
difficult. Moreover, abstraction-hungry systems necessarily take a long time to set up, suffer from a sort
of delayed gratification, because the appropriate abstractions must be defined before the users
immediate goals can be attacked.
We have already mentioned the problems of hidden dependencies and juxtaposability that bedevil so
many abstraction managers.
One reason why spreadsheets are so popular is that they are weak on abstractions. In fact, not
surprisingly, many potential end-users are repelled by abstraction-hungry systems.
One proposed workaround for the abstraction barrier is Programming By Example, in which you show
the system some items, and it decides what the abstraction should be. This tantalising idea doesn't seem
to have worked very well so far. The errors of excessively over-general or over-specialised abstraction
are even more likely to be committed by a machine with no recourse to common-sense sanity checks.
Even worse, you still have to do the hard work identifying the set of examples (including appropriate
negative examples) that will be used to define the abstraction. As a programming method,
programming by example has the potential to introduce some horrendous problems combining hidden
dependencies and premature commitment.
29
Secondary Notation
Definition
Extra information carried by other means than the official syntax.
Redundant recoding gives a separate and easier channel for information that is already present in the
official syntax. Escape from formalism allows extra information to be added, not present in the official
syntax.
Thumbnail illustration
Indentation in programs (redundant recoding); grouping of control knobs by function (redundant
recoding); annotation on diagrams (escape from formalism).
Explanation
The formal notation can sometimes be supplemented with informal additions which have no official
meaning, but which can be used to convey extra information to readers. Indentation or pretty-printing
is an essential part of a programming language, but it has no meaning to the compiler in most cases: it is
a form of secondary notation. Exactly the same pretty-printing idea is used to help make telephone
numbers easier to read, although interestingly enough, conventions differ from country to country - in
Britain we split numbers into threes and fours, but in France they split them into two-digit groups.
Cognitive relevance
Redundant recoding makes comprehension easier. Easier comprehension makes easier construction: all
research on the design process shows that creating information structures is a looping process in which
the part-finished structure is inspected and re-interpreted.
Highly relevant to exploratory and modification activities. Less relevant to transcription activities but
not to be ignored there, either;
Escape from formalism adds to the information available; typical comments will describe the origin, the
purpose, the design rationale, the reliability status (e.g., whether tested), or the intended future (still to
be revised). Such information is an essential commodity in the process of building or modifying an
information structure.
Secondary notation has an equally valid place in interactive systems; a control system in which
functionally-related controls are grouped together is likely to be easier to learn to induce fewer slips of
action too.
transcription incrementation modification exploration
secondary notation useful (?) v. useful harmful ?
Cost Implications
It is hard for me to imagine any situation that would not be improved by making a notation easier to
read, but die-hard C/predicate calculus/whatever users maintain that anyone can understand their pet
formalism. Attempts to create less impenetrable notations, such as visual programming languages, are
infantilisation of computer science (Dijkstra). (If you meet such a person, offer them an exam question
in C++ and ask whether they mind if you turn the lights right out and the music full up.)
30
Unfortunately designers often forget it. It is easy to work with implicit model of users in which they
only know, and only need to know, that information which is contained in the official syntax; or to
expect users to master all controls, however arbitrary and complicated.
Types and Examples
Redundant recoding
A typical Sheffield telephone number is printed as 0114 212 3456, indicating city, area (or service
provider), and individual number. Before the Sheffield numbers were changed to include an extra 2, it
would have been printed as 0114 123 456. When the extra 2 was added, there was some degree of
confusion: was the new digit part of the city code (giving 01142) or not? If Jez Rowbotham believes the
Sheffield code is 01142, and if Jez has been told the number as Sheffield 212 3456, what will happen
when he tries to ring that number? He will dial 01142 225 533(5) the final digit being ignored by the
telephone exchange. Whoever lives at 0114 222 5533 will get a wrong number.
(At home, TGs number is frequently rung by people who think we are a ski travel firm and try to
arrange ski-ing holidays. Can I speak to Maureen please? Sorry, were not the ski-travel firm. Well,
who are you then? I rung the right number, didnt I? Leeds 234 4321? Well, I think you probably rang
an extra 2, so youve got through to Leeds 223 4432. The code for Leeds is 0113, not 01132. Well, I
dunno, thats poor role-expressiveness, that is, they oughta improve the secondary notation, why dont
they learn about cognitive dimensions? Sorry to trouble you, bye now, and so on.)
You can get some idea of the importance of secondary notation in this context by starting to give British
people telephone numbers expressed in the French style:
0114 225 5335 becomes 0 11 42 25 53 35
At the same time, do the opposite with your French friends. You will find that your friends are
perplexed and probably irritated, because they cannot parse the new presentation into the units they
are used to.
The facade or front panel of an instrument (e.g. a car radio/cassette) gives designers a chance to use
secondary notation to good effect, by grouping related controls in the same area. This is simple enough
for an instrument with few facilities (see the telephone example on the web site) but for more complex
instruments it can be much harder to bring off well.
Functional grouping is likewise an important aspect of circuit notations (Figure 15). The wiring
diagram showing the layout of physical components is constrained by factors of size, heat emission and
so on, so that components have to be fitted together as best they can be; to make the functional
relationships clear, it is uaul to present a schematic showing the connections between idealised
components.
31
Figure 15 Secondary notation in circuit diagrams: the left hand shows a schematic in
which components have been organised to make their function apparent (redundant
recoding). The right hand diagram shows the layout on the circuit board.
The role of secondary notation is often ignored by the proponents of new formalisms and other
designs. In the traditional languages, like Basic, redundant information is routinely carried by other
means than the formal syntax. Sometimes there is even a visual rhyme, with similar fragments laid out
in similar fashion:
Vaccel = Thrust*COS(Angle)/Mass Gravity
Vveloc = Vveloc + Vaccel
Vdist = Vdist + Vveloc
Haccel = Thrust*SIN(Angle)/Mass
Hveloc = Hveloc + Haccel
Hdist = Hdist + Hveloc
Note the use of both white space and choice of statement order to get the desired effect. This allows the
reader to check and see that the differences are just as expected i.e. that the vertical component allows
for gravity, unlike the horizontal.
Escape from formalism
Now we turn to the use of secondary notation to carry additional information. A study on electronic
and paper calendar usage (Payne, 1993) showed that paper diaries had an advantage because their
users could use big letters to show that something was important, or could draw lines between entries.
The electronic forms presented all entries in the same font and had no facilities for adding graphics, so
they were unable to express enough information, even though the official information was the same.
(See figure below.)
32
Figure 16 Secondary notation in the paper-based diary (above) can be rich and
expressive; much less is available in a computer-based diary (below). In hand-held
electronic organisers, possibilities are even more limited.
The programming equivalent is the comment. Programmers need to map the program onto the domain
and tell a story about it; and they need to document sources (The next part follows the method of X
and Y) or doubts and intentions (This part is untested, but Ill check it out soon). Amazingly, many
notations intended for non-expert end-users, such as many visual programming languages and even
spreadsheets, have little or no provision for comments.
important
scar
scar
regular
event will
NOT take
place
different
handwriting
and colours
distinguish
different
users
the scars
show that
this day
was hard to
plan - avoid
further
changes!
33
Workarounds, Remedies and Trade-offs
The main workarounds/remedies are:
$ Decoupling: print out a hard copy and attack it with a pencil.
$ Enriched resources: provide tools in the system that allow components to be labelled and described and
their relationships made explicit. Hendry and Green (1993) describe a technique for doing this with
spreadsheets, called CogMap, the cognitive mapper.
Extensive secondary notation creates additional problems of viscosity; if the structure is changed, the
secondary notation goes out of date. If the environment contains tools to maintain its accuracy, such as
indentation facilities, well and good; but it is hard to imagine tools to support escape from formalism,
by its very nature.
34
Visibility & Juxtaposability
Definition
Visibility: ability to view components easily.
Juxtaposability: ability to place any two components side by side.
Thumbnail illustrations
Visibility: searching a telephone directory for the name of the subscriber who has a specified telephone
number.
Juxtaposability: trying to compare two statistical graphs on different pages of a book.
Explanation
In the telephone number case, the indexing facilities of a system do not provide the necessary
information, the design assumption being that users will rarely want to know which person is number
XX. As a real life case that is not actually true: in the UK there is a useful who-rang-this-number-last
feature, but sometimes one hasno idea about the identity of a mystery number; and in the telephone
memory there may be numbers with no attached names, so that one would like to know who they are
or which one is the Car Rescue Service, preferably without having to ring each number to find out.
Visibility problems also turn up when the trail of steps leading to a target is longer than desirable,
making navigation onerous and possibly error-prone; seeking a particular page on the web, for example
(exacerbated here by the typical slow response of the web). In these cases, users building or modifying
or querying an information structure are frustrated by the extra work.
Many systems restrict viewing to one component at a time, losing juxtaposability. Typically these
systems present information in a window, and are restricted in how many windows can be viewed
simultaneously.
Cognitive relevance
The relevance of visibility is immediately apparent in the simple case of searching for a single item of
information. Less apparent is that a complex structure is not grasped by inspecting one detail at a time,
but by scanning a large sweep (for example, programmers do not read a program one line at a time, but
by scanning to and fro). Poor visibility inhibits scanning, a fact too often overlooked.
The relevance of juxtaposability is even more often overlooked. Grasping the difference between two
similar items needs close comparison, as in the example of comparing graphs in a book. Consistent
treatment of similar data can only be achieved by making sure the second instance is treated the same
way as the first. And, more subtly but just as important, the solution to a problem is often found by
finding a solution to a similar problem and then building a modified version; or the new solution is
compared to an old one to make sure nothing has been left out. The absence of juxtaposed, side-by-side
viewing amounts to a psychological claim that every problem is solved independently of all other
problems.
Transcription and incrementation only require visibility and juxtaposability for error-checking.
Modification makes heavy demands on juxtaposability, to ensure consistency of treatment. Exploratory
design uses juxtaposition as a seed for problem-solving.
35
transcription incrementation modification exploration
visibility / juxtaposability not vital not vital important important
Cost Implications
Mobile computers and wearable computers are going to make us all too familiar with tiny windows in
which juxtaposability has been sacrificed.
Types and Examples
Visibility
Paper-based artefacts, with some centuries of typographical development behind them, are quite good
for many purposes, but when the information load becomes too heavy visibility can drop dramatically.
Consulting a nation-wide railway timetable will quickly illustrate the problems.
Contemporary designs of domestic information-based devices face a recurrent difficulty that will be
familiar to many. These devices allow many parameters to be changed, and it is necessary to provide
default settings for some or all of them. In-car radios need to have a default initial volume setting for
normal programs, and a separate default initial volume setting for traffic announcements (because these
are broadcast from a different station, they might otherwise be much louder). Fax machines need to
store information about user name, the call-back number, the time and date, etc. Cameras are going the
same way. To access the default settings on these devices, the typical method is to provide a two-level
menu system in which the top level covers major groups of functions and the second level provides
more detailed items within each group.
The menu structure only requires one display item at a time, plus two keys to scroll to and fro through
the menu and one further key to accept the current item, so this approach avoids both a clutter of
information on the display, and a proliferation of special purpose keys. The trade-off is the poor
visibility. Confirming that the fax machines call-back number or the car radios initial traffic
announcement volume is correctly set up takes several actions, enough to discourage many users.
To check the call-back number on the British Telecom Image 350 fax machine,
perfectly typical of the menu approach (but rather nicely designed and accompanied by
good documentation), the following steps are required (quoted from the instruction
manual):
Press the SET UP button.
Press the < or > buttons until the display shows USER SET-UP.
Press the (tick-mark) button; display shows LINE SET-UP.
Press the < or > buttons until the display shows USER DATA.
Press the button; display shows SENDER ID.
Press the < or > buttons until the display shows CALL BACK TEL NO.
Press the button; display shows the currently set call back number [and
allows it to be edited].
The precise number of moves will depend on the state the machine is in, but can go to a
dozen or more button-presses.
Figure 17 Visibility problems in a typical personal information artefact mean that
checking stored values is slow and cognitively costly.
Just as important, the user has to remember how to get to the right part of the control tree. The fax machine
described is provided with quite sensible, meaningful item labels; but in-car radios, typically, have
much terser and more opaque labels, such as DSC (thats where the default volume settings are not an
36
easy guess for the uninitiated: it means Direct Software Control) or TIM (nothing to do with time
that one means Traffic Information Memo).
Interestingly, the usual documentation of these devices gives a second-level example of visibility
problems. Typically the description is purely procedural, like the one above, which means that to find
the information needed the user has to trawl through many pages of very similar procedures. It would
be very straightforward to present the menu structure directly, providing a more memorable
representation that is more quickly searched. For the fax machine mentioned above, the tree would start
like this
set resolution
normal
fine
superfine
photo
set contrast
normal
light original
dark original
print-outs
(sub-menu items)
directory set-up
(etc)
Event-based scripting languages create well-known visibility problems. For example, Visual Basic
disperses little packets of code (scripts or methods) in many different places through the stack. For a
single script, the local visibility is good; the code is usually associated closely with the object that uses it.
But the result of the dispersal is that overall visibility is low. Extraordinary debugging difficulties were
chronicled by Eisenstadt (1993), caused by the many hidden dependencies and the poor visibility.
Juxtaposability
Problems with juxtaposability are more pervasive than simple visibility problems.
The old-fashioned alarm clock dial simultaneously shows current time and alarm time, with a different
knob to control each one. Very simple. The digital version usually has two modes, one showing current
time, the other showing alarm time, so that the user cannot compare them. Not too much of a problem,
but a mild nuisance, and typical of failure to appreciate the need for juxtaposition of information.
Entering data into a form-based system is a frequent source of problems; the user wants to make sure
that the layout is consistent from one to the next, yet there is no way to see an earlier entry while adding
a new entry. The EPSRC Electronic Proposal Form
1
(Version 3.0, as of Feb. 1998) is a splendid example.
One section requires names of, and some information about, co-investigators; a subsequent section
requires further information about each co-investigator, to be given in the same order yet you can no
longer see the previous section.

1
Available at http://www.epsrc.ac.uk/resgrant/resform/index.htm
37
Figure 18 Lack of juxtaposability in a forms-based system. In a previous section, I
entered the names of two proposed co-investigators, A. Einstein and C. Darwin. Now
I have to enter personal information about those people. Since there is no link by
name, the information has to maintain the order in which the names were entered
either Einstein first or Darwin first but I am not allowed to see the previous form.
The EPSRC form is clearly designed with transcription activity in mind: the user is expected to have
sorted out all the information neatly before starting to enter it. Real life is different. The personal
information is gained from scraps of paper, old e-mails to be found somewhere in the file-system, or
hasty telephone calls, all the while rushing to get the work completed before the mail goes (or some
other deadline). Transcription activity has no place in this scenario of opportunistic incrementation,
which is in fact more like exploratory design than relaxed transcription.
Often, the question may not be whether consistency has been maintained, but whether information has
already been entered. What entries have been made in the memories of a mobile telephone? Do I need
to add Aunt Sheila, or is she there already?
Spreadsheets, with their two layers (data and formulas), have asymmetric visibility and juxtaposability.
When the sheet as a whole is displaying data, only the formula in one cell can be made visible at any
one time. Obviously it will be difficult to compare that formula to any others. When the sheet is
displaying formulas, the formula comparisons are easy but now it is no longer possible to see the
values computed by the formulas, which loses much of the point of the spreadsheet model.
The advent of windowing systems brought great advances in juxtaposability for all jobs where more
than one task was performed concurrently. A notable improvement for many users was being able to
run terminal emulations with windows, using one window for each thread under say X-windows and
making a whole new way of working possible; in contrast, on the standard Unix platform without
windows, all the threads had to be merged into one stream, making multiple concurrent sessions
extremely hard if not impossible to manage.
38
Workarounds, Remedies and Trade-offs
$ Working Memory: When side-by-side viewing is not supported by the environment, users have two
choices. The first is to rely on working memory, frequently refreshed by revisiting the two items being
compared in turn.
$ External Memory: Or else they can make a hard copy of one component to study. This second strategy
amounts to decoupling themselves from the given environment and proving themselves with a new
environment in which side-by-side viewing is possible.
$ Adding a browser: Serious visibility problems can be remedied by the provision of browsers, such as
Smalltalk-80s class browser, and the provision of alternative views for instance, some of
HyperCards problems have been eased in other systems by providing a contents list of all the cards in
the stack.
Visibility trades off against clutter, which has not been discussed here, and against abstraction level,
since providing a browser for an alternative view is creating a new abstraction.
39
The Remaining Dimensions
The remaining dimensions are more psychological in character. We shall do no more than sketch them
here.
Closeness of Mapping
Definition
Closeness of representation to domain
Thumbnail illustrations:
A close mapping: the visual programming language LabVIEW, designed for use by electronics
engineers, is closely modelled on an actual circuit diagram, minimising the number of new concepts
that need be learnt. A distant mapping: in the first version of Microsoft Word, the only way to count the
characters in a file was to save the file to disc whereupon it told you how long the file was.
Consistency
Definition:
similar semantics are expressed in similar syntactic forms
Thumbnail illustration:
Many domestic information artefacts are now menu-driven, such as in-car audio sets, with keys to
move through the menus. In a consistent version, the same keys move up or down, step to the next
sibling, select the current item, or return to the menu display. But frequently there are slight differences
from item to item.
Note: Consistency usually affects learnability rather than usability (but see error-proneness). The work
of Cramer (1990) suggests that its influence is less than that of mnemonic labelling, which is not
considered here because it is not a structural property.
Diffuseness
Definition
verbosity of language
Thumbnail illustration:
COBOL is a verbose language:
MULTIPLY A BY B GIVING C
Forth is a terse language: the command to print a newline is
.
(i.e. a single full stop).
Cognitive importance:
The effect of over-verbosity is probably slight in most situations, but it should be kept in mind that
studies on working memory have shown that mental arithmetic performance is better for children
whose native language has short names for numbers rather than long names: the implication is that
having to keep long-winded descriptions in mind will limit users who are engaged in activities that
demand working memory, such as exploratory design.
40
Terseness can also be a problem, because it makes it easier to make mistakes (see error-proneness) an
awkward trade-off.
Error-proneness
Definition
notation invites mistakes
Thumbnail illustrations:
(i) Poor discriminability: the language Forth uses full stop and comma to mean different things; Fortran
uses the identifiers I and O, regularly confused with one and zero.
(ii) Inadequate syntax-checking: Prolog has no declarations of identifiers, so mistypings cause problems
detectable only at run-time.
(iii) Bad dialogue design: inconsistencies in dialogues invite errors, e.g. always using ENTER as a
default except in one case.
(iv) Memory overload (caused by premature commitment, etc).
Cognitive importance:
Obvious.
Hard mental operations
Definition
high demand on cognitive resources
Thumbnail illustration:
Threading a maze new branches to remember are spawned at every choice point. For certain physical
mazes there are memoryless algorithms, but following circuit diagrams, auditing spreadsheets, and
even finding files in big directory structures are maze isomorphs and are correspondingly hard.
Cognitive importance:
Obvious.
Progressive evaluation
Definition
work-to-date can be checked at any time
Thumbnail illustration
Spreadsheets recompute at every opportunity, so that the user can develop a solution bit by bit.
Cognitive importance:
Novices need frequent checks on work-to-date. Even experienced users will need checks during
difficult times, such as working under stress of frequent interruptions.
41
Provisionality
Definition
degree of commitment to actions or marks
Thumbnail illustration:
Pencils are used by architects, typographers and other designers to make faint blurry marks, meaning
something more or less like this goes more or less here, as well as precise hard marks.
Cognitive importance:
Reduces premature commitment.
Role-expressiveness
Definition
the purpose of a component (or an action or a symbol) is readily inferred
Thumbnail illustration:
This dimension describes the ease with which the notation can be broken into its component parts,
when being read, and the ease of picking out the relationships between those parts. An experienced
electrical engineer can look at a radio circuit diagram and quickly pick out the parts that deal with
different stages of the process (detecting the signal, first amplification stage, etc.).
Cognitive importance:
Although some notations and devices seem to give less trouble, not much research has been reported,
and this remains one of the hardest dimensions to clarify, needing further research on parsing and
comprehension.
42
Cognitive Relevance, Trade-offs, and Design Manoeuvres
Each cognitive dimension above has been characterised as particularly relevant to certain user activities.
Some of the trade-off relationships and some of the classic design manoeuvres have also been noted.
This section tabulates what has been said above.
Cognitive relevance
Here we collect the comments made under each dimension. The columns give desirable profiles of
dimensions for each of the types of user activity mentioned above.
transcription incrementation modification exploration
viscosity acceptable acceptable harmful harmful
hidden dependencies acceptable acceptable harmful acceptable for small
tasks
premature
commitment
harmful harmful harmful harmful
abstraction barrier harmful harmful harmful harmful
abstraction hunger useful useful (?) useful harmful
secondary notation useful (?) v. useful v. harmful
visibility /
juxtaposability
not vital not vital important important
Figure 19 Table of user activities
Notable Trade-offs
The issue of trade-offs is very complex. Only some of the more obvious ones have been listed here,
especially those that form part of typical design manoeuvres, such as exchanging viscosity for
abstractions. Very little research has been reported, to our knowledge, on these issues. The following
diagram illustrates the relationships that have been mentioned above.
43
visibility
secondary notation
viscosity
abstraction usage
learnability
hidden dependencies
premature commitment
juxtaposability
Figure 20 A line indicates a trade-off relationship identified in the text above. Not all of
these trade-offs are available in all cases, and the diagram does not pretend to
identify all possible relationships.
Design Manoeuvres
Some of the classic design manoeuvres identified above are listed in the following table.
AIM MANOEUVRE AT THIS COST
to reduce viscosity: add abstractions (so that one
power command can change
many instances
increases need for lookahead (to
get the right abstractions);
raises the abstraction barrier;
may also increase hidden
dependencies among the
abstractions
to improve comprehensibility: allow secondary notation let
users choose where to place,
things, how to use white space,
what font and colour to use;
allow commenting
increases viscosity (because
layout, colour etc are not
usually well catered for by the
environment)
to make premature commitment
less expensive:
reduce viscosity (so that users
can easily correct their first
guess)
see above, re viscosity
to remove need for lookahead: remove internal dependencies
in the notation;
or allow users to choose an
easier order to make decisions
may make notation diffuse, or
increase errors
allow free order needs a
cleverer system
to improve visibility: add abstractions (so that the
notation becomes less diffuse)
see above re abstractions
44
Further developments in cognitive dimensions
Although the cognitive dimensions framework succeeds in its aim of being broad-brush, it pays by
being too unspecific, making it hard to metricate the dimensions or even to define exactly what they
mean. Several projects have set out to improve the situation.
Metrication
Yang et al. (1997) offer benchmarks for navigable static representations. These authors realised that for
practical use benchmarks with real numbers attached would be more useful for some purposes than
mere discussion tools, so they set out to define metrics for some of the dimensions. They sensibly
restricted their aims to those dimensions that they thought would be of most interest for their purposes
of designing and improving visual programming languages, and their first decision was to concentrate
on the static part of a program representation rather than its dynamics, arguing that obtaining a usable
static representation was an essential first step.
Eighteen benchmarks are defined in their paper. Examples are:
Visibility or hiddenness of dependencies:
D1: (sources of dependencies explicitly depicted) / (sources of dependencies in system)
D2: The worst-case number of steps required to navigate to the display of dependency
information
Visibility of program structure:
PS1: Does the representation explicitly show how the parts of the program logically fit together?
(Yes/No)
These benchmarks, though admittedly somewhat on the crude side, were applied by the authors to
different languages and helped them find overlooked usability problems and encouraged them to
redesign the languages. They encourage others to find similar domain-specific benchmarks for their
own purposes.
Formalisation
The vagueness of the cognitive dimensions has stimulated more than one attempt to define satisfactory
formal representations.
ERMIA
Green (1991), much developed by Green and Benyon (1996), offered an extended entity-relationship
model of information artefacts in which certain of the cognitive dimensions could fairly easily be
identified and in some cases metricated. Viscosity lends itself quite well to this approach. An ERMIA
(entity-relationship modelling for information artefacts) model of conventional music notation looks
like this:
45
Piece
Key signature
Note Symbol Height
Note Name
Key
Pitch
1
M
1
1
1
1
1
1
1
1
1
1
1
1
1
1
indicating that the piece of music contains many note symbols, each of which has a height on the staff
(the black border indicates a perceptual cue) and a note name. To transpose the key of a piece, every
note must be rewritten. The approach even offers a symbolic computation of the cost of change:
Transpose Staff Notation
write Key' change key signature
foreach Note: change all notes
read Note
compute Note'
write Note'
Total actions: #(write actions) = #(Notes)
Working memory storage:
1 place-holder , to keep track of notes in the original document
For this structure, V = O(n), where n is the number of note symbols. (The value of n can run into the
thousands.) Note that an alternative representation for music, known as tonic sol-fa, reduces the
viscosity down to a single action!
Formalisation through System Modelling
Roast and Siddiqi (1996 et al.) take a very different approach, using system abstractions to lead to closer
definitions of the precise meaning of various cognitive dimensions. Detailed consideration of their
careful work would take up too much space here, unfortunately, but some flavour of it can be given by
looking at the route taken by Roast (1998) in examing premature commitment.
Roasts approach provides interpretations based on
the goals and sub-goals which users may achieve,
the user inputs which enable users to satisfy different goals,
the objects central to characterising the goals to be achieved.
Roast describes premature commitment as the user having to satisfy the secondary goal prior to
achievingthe primary goal. To develop a formal estimation of premature commitment he focuses on
three facets of the concept: the characterisation of how the secondary goal is unavoidable; the
relationship between the secondary and primary goals which indicates the inappropriateness of having
to commit to the secondary goal first, and the initial conditions. This leads to a new distinction between
weak and strong premature commitment, and thence to a detailed formal analysis of the relationship
between premature commitment and viscosity. Finally, in a small case study, the formal analysis is
applied to an editor used for the specification language Z, and is shown to conform to the difficulties
experienced by certain users.
46
Part 2: Real Systems
In this part we shall work the other way round: instead of listing the dimensions and mentioning
examples, we shall start with genuine applications and apply the cognitive dimensions approach to
them.
It must be understood that doing a proper analysis, even in a broad-brush style, requires close
familiarity with the device or notation. Since that degree of familiarity can hardly be achieved merely
by reading about the device, the analyses given here will be rather sketchy.
The examples illustrate very different systems an on-line diary and a pair of visual programming
languages, showing how to go about applying the framework in these different cases.
47
A Diary System
On-line diaries and calendars, meeting organisers, time managers, and so on are one of the popular
applications of information devices. The example we shall analyse is Now Up-to-Date, a product of
Now Software, henceforth abbreviated NUD.
NUD is a complex calendar application designed to have lots of features yet be easy to use. Before
applying individual dimensions we will outline its operation.
Events are the main type of data. Events can be dated (holiday, appointment); undated (e.g. to-dos);
and things like banners and graphics and notes, which are undated so one might hesitate to call them
events, but thats what the device calls them.
Attributes of events include
name
type
category (see below)
priority
reminder (has a default)
start time
duration
whether to carry forward
Not all attributes apply in every case e.g. a note doesnt have a start time. Events may also be repeating
see below.
Categories describe events. Every event is in one and only one category: users can create new categories
ad lib. For instance, a business category would contain work-related appointments and to-dos; there
might be a holidays category, etc. One category must always be the default category; new events with
unspecified category are defaulted to that category.
Sets determine which categories are visible; only one set is visible at once. They can contain any number
of categories and can overlap as much as desired. (E.g. you might have a work set containing the
business and holidays category, and a leisure set containing the holidays category plus the category of
meetings of hobby clubs, and another set for your spouse/partner using the same system ...)
Several different views are available, some by time unit (year, month, day) and some as lists of events.
The most relevant ones are the month view, the day view, and the list view.
Figure 21 Month view in NUD.
48
Figure 22 The day view in NUD.
Figure 23 The list view in NUD. The user has considerable control over what
categories are included in a list.
Cognitive dimensions: Main device
Viscosity
While in the month view, Modifying attributes (start time, etc) of the events that have been set up is
very easy but requires typing in the time. Moving events from day to day is just drag-and-drop.
Inserting a new event in the month view potentially causes both knock-on and repetition. There is no
way to insert an event between existing events and have them all close up slightly to make space for the
newcomer, and so the user has to adjust all the existing events manually (this is the knock-on viscosity).
In the month view, each event has to be separately opened and have its start and end time retyped (this
is the repetition).
Figure 24 Defining a new event in NUD. To change the scheduled times of an existing
event, this dialog box is re-opened and edited.
49
Fortunately, the day view is less demanding. Times are mapped onto spatial layout in a simple table, in
which the start and finish times can be dragged up and down with the cursor:
Figure 25 Left, modifying a days schedule: coffee time is being extended. Right,
rearranging the schedule by dragging discuss life upwards. Note that overlapping
events are shown side by side in the day view.
Although there is a small degree of repetition viscosity when inserting a new event into a busy
schedule, the alternatives would probably have been too complicated.
Viscosity does, nevertheless, cause problems when a number of days are to be updated in the same
way. If a set of appointments is to be carried over from day to day all week, or if something means that
all appointments on each day for the next few weeks have to be put back half and hour, it will be a
tedious job to enter all the changes. There is no facility, for instance, for copying a complete day and
pasting it into the following day.
Hidden dependencies
There are very few dependencies in this system. Each event is independently recorded, and events can
be allowed to overlap; thus changing one event will have no effect on others. But it should be noted that
there is also no representation of dependencies that may exist in the users life. One of the examples
given by Payne (1993) is that of organising a conference, an activity spawning lots of sub-activities; if
the conference is cancelled, all the associated dates and appointments can also be cancelled.
The system does, however, give a representation for events that repeat (birthdays, weekly seminars,
etc), and there are some potential dependencies for these events, since different occurrences are
necessarily linked. If the start or finish time of an individual event is to be altered, NUD offers a dialog
box allowing the user to change all the events or only the currently selected event. But if an event is
changed from a type with a start time to one that has no fixed time (such as a Note), then all the other
repetitions of that event are silently changed, and their start times are lost.
Figure 26 Defining a repeating event in NUD.
50
Premature commitment
In the main display, there is a modicum of premature commitment when defining the type of an event,
since it must be classed as one of the set of types at the moment of creation; however, its type is very
easily changed, so this is unimportant.
Abstraction Barrier and Abstraction Hunger
The abstractions available are categories and sets, described above, and repetitions of events. Every
event belongs to just one category; sets may comprise any number of categories. Only one set is
displayed at one time. The definition and editing of categories and sets is through an abstraction
management sub-device, dealt with separately below.
The defaults allow the user to work in ignorance of category and set concepts until such time as they are
needed so there is no initial abstraction barrier. Thus, NUD is an abstraction-tolerant device.
Secondary Notation
Secondary notation is available through variations in colour, font and style. One use is to distinguish
classes of event. Each set includes any number of categories, and categories may belong to any number
of sets; and each set can have its own way to display each category. Thus a set called Holiday events
could display all the events put into the Lolling around category in blue, and all the events in the
Seeing the sights category in red; while another set might include only one of those categories, and
display it in green.
Any individual event can also be assigned its own over-riding display attributes, so the user could
choose to show important events in big bold red letters if desired, regardless of the default attributes.
Showing provisional events or cancellations is not so easy they have to be treated as pseudo-events.
There is no escape from the formalism.
Visibility and Juxtaposability
Visibility for single events and classes is well handled. The day and month views illustrated above deal
with most events; a list view can also be obtained, in which all events of a given set are listed together
without the visual framework of the calendar:
Figure 27 The list view in NUD allows timelines to be separated by displaying only
the events of one category.
Juxtaposing days, weeks or months is possible by opening a duplicate window. This expedient is not
immediately apparent and for a long time this user believed that there was no way to juxtapose weeks.
Abstraction management sub-device
The chief abstractions available are the repetition and the set of categories. Not much can be said about
repetitions, but building and modifying sets is more interesting. The main display for constructing sets
51
of categories looks like the figure below. The list shows which categories are included in the set and
what text attributes to use when displaying those events.
Figure 28 The dialog for managing sets of categories of events in NUD.
Viscosity
Viscosity is low: categories are readily introduced or removed.
Hidden dependencies
Few dependencies are involved.
Premature commitment
Building a set necessitates a certain amount of premature commitment for most users, since they do not
know in advance whether their categorisation of their life events is stable or accurate. Fortunately,
editing the set is not difficult.
Abstraction Barrier and Abstraction Hunger
NUD cannot be used without at least one category, to use as a default, and at least one set, again to use
as a default. Although new users do not have to operate on these defaults, they cannot ignore them
completely. There is, therefore, a real, though not enormous, abstraction barrier. The problems of new
users would have been easier if better names had been found, indicating more clearly the relationship
that every event exactly one category and every category at least one set.
Secondary Notation
There is no secondary notation for sets and categories.
Visibility and Juxtaposability
Visibility is good for single set definitions, as shown above, but there is no juxtaposition possible; only
one set definition can be displayed at one time.
Conclusions
NUD is a good system for incrementation or transcription activity. Modification of existing schedules is
less good.
52
Two Visual Programming Languages
Information-based devices are getting more sophisticated. At the rate were going, our central heating
controls will soon have macro languages and our video systems will have compilers and link loading.
Maybe thats a slight exaggeration, but at the very least, we need to give increasingly complex
instructions to devices like, for example, telephone-based query systems. How are users going to
communicate what they want to happen?
Visual languages have been a frequent suggestion. Often the reasons given are based on rather suspect
folk psychology, such as graphical languages must be better because they use more of the brain or
that tired, tired slogan A picture is worth ... (see Blackwell, 1996, for a report on the folk psychology
and its relationship to research findings).
But folk psychology aside, graphical languages have some degree of promise. They dont look like
programming, which is a help: no grey slabs of small print, no tricky syntactic niceties. They are in
disguise, the acceptable face of programming.
Box-and-wire diagrams are the commonest suggestion, though not necessarily the best: each box
contains an instruction, and the wires either carry data from one box to the next (the data-flow
paradigm) or indicate the sequence of operations, as in a flow-chart. We have seen box-and-wire
languages proposed or implemented for many types of system, including
scheduling systems
network configuration systems
component assembly environments
Java generators
project management tools
telephone menu programming systems
How usable are box-and-wire systems? There are many variations. Apparently-small notational
differences in box-and-wire languages can lead to extreme differences in the programs. Green and Petre
(1996) reported a usability analysis of two commercial box-and-wire systems, Prograph and LabView2,
and how they compared to the old-fashioned Basic language. This section is based on that work. (Only
the highlights can be mentioned here; read the full paper before implementing your own box-and-wire
system.)
Important note: both these languages are excellent for their purposes. In no way are the following
remarks to be taken as a complete review. The two languages have adopted different solutions to the
design issues and the trade-offs, and our purpose is merely to illustrate some of the consequences of
exchanging one kind of usability for another kind. The first author, TG, has made much use of both
languages and can warmly recommend them for their appropriate types of use.
Illustrating the notations
We cannot explain the details of these languages here, but we can give an idea of the important
differences.
For comparative purposes, here are three versions of the same program. This program, which computes
the flight path of a rocket, was derived from algorithms used by Curtis et al. for research into the
comprehensibility of flowcharts; Green and Petre re-used them for their research into modern visual

2
Produced and trademarked by Pictorius Inc. and National Instruments Inc., respectively.
53
programming languages. First, we show the Basic version, the LabView equivalent, and lastly the
Prograph version.
Mass = 10000
Fuel = 50
Force = 400000
Gravity = 32
WHILE Vdist >= 0
IF Tim = 11 THEN Angle = .3941
IF Tim > 100 THEN Force = 0 ELSE Mass = Mass Fuel
Vaccel = Force*COS(Angle)/Mass Gravity
Vveloc = Vveloc + Vaccel
Vdist = Vdist + Vveloc
Haccel = Force*SIN(Angle)/Mass
Hveloc = Hveloc + Haccel
Hdist = Hdist + Hveloc
PRINT Tim, Vdist, Hdist
Tim = Tim + 1
WEND
STOP
Figure 29 Rocket program in Basic
The LabView notation uses boxes for operations and wires for the data. Conditionals are represented as
boxes lying on top of each other, only one of which can be seen at any one time. Loops are represented
by a surrounding box with a thick wall.
54
Figure 30 The rocket program in LabVIEW.
The Prograph notation likewise uses boxes to indicate operations and wires to transmit data, but puts
each conditional into its own subroutine with it own window. The result looks like this:
55
Figure 31 The rocket program in Prograph.
Viscosity
Green and Petre made a straw-test comparison of viscosity by modifying each of the programs in the
same way, putting in a simple correction for air resistance. Because they wanted to measure the time to
edit the program but not the time taken to solve the problem, the modification was worked out in
56
advance, and they timed an experienced user modifying the original program (running in the standard
environment), working from a print-out of the modification required.
Inserting the material into LabView took a surprisingly long time because all the boxes had to be jiggled
about and many of the wires had to be rebuilt. Prograph was able to absorb the extra code to deal with
air resistance with little difficulty certainly less than evidenced by LabView. This was because the
new code was placed in new windows, and relatively little change had to be made to existing layouts.
(Notice, however, that this causes poor visibility and juxtaposability, as mentioned below.) For Basic,
the problem is just one of typing a few more lines. Overall time taken was as follows: LabView, 508.3
seconds; Prograph, 193.6 seconds; Basic, 63.3 seconds - an astonishing ratio of 8:1 between extremes.
These are differences of a whole order of magnitude, and if the programs were larger we would expect
them to increase proportionately (i.e. to increase faster for LabView, slower for Basic).
508.3
193.6
63.3
0
100
200
300
400
500
600
LabView Prograph Basic
Figure 32 Times in seconds to make equivalent editing modifications
Viscosity is obviously a major issue.
Hidden dependencies
An important strength of the visual approach is that many of the hidden dependencies of textual
languages are made explicit. Data in a textual language like Basic is transmitted via assignments and
use-statements, thus:
x = 1
... (possibly many pages of code here...)
y = x + 3
The same information in a LabView program is shown thus:
At least one type of hidden dependency is thereby brought into the open. The trade-off in this case,
obviously enough, is the use of screen space.
Premature commitment
Green and Petre noted several types of premature commitment in the construction of these programs:
commitment to layout, commitment to connections, and commitmwent to choice of construct.
57
Commitment to layout
Obviously, the visual programmer has to make the first mark somewhere on the virtual page. As the
program takes shape it may become clear that the first mark was unfortunately placed, or that
subsequent marks were ill-placed with respect to it and each other. Although there is always some way
to adjust the layout, the viscosity may be too high for comfort. That is certainly the case with complex
LabView programs.
The difficulty is better illustrated than described. Consider building a program to solve quadratic
equations, according to the usual formula
x = (-b (b
2
- 4ac) / 2a
The sequence of screenshots below I taken from an actual session in which a LabView user built the
program.
1. " ... Start with minus b..." :
2. " ... I'll need b-squared too ..." :
3. " ... Turn that into b-squared minus 4ac ... " : (Note: the wire bobbin in the next frame is the
cursor of the connection-making tool.)
58
4. " --- Oops, that's going to be 4ac minus b-squared ... try moving that 4ac chunk down and
reconnecting to the minus box ..." :
5. " ... OK, now I need to do plus-or-minus that lot ..." :
6. "That's root-b-squared-minus-4ac. But hey, What A Mess! and I still haven't used the -b, let
alone done the bottom line of the formula... How do I tidy it up?"
Commitment to connections
The 2-dimensional layout of VPLs requires subroutines to be associated with their callers by ports with
a definite placing. In both our target languages, one easily finds that the data terminals are not arranged
in the order one wants them for a particular purpose. For example, in Prograph, the order of terminals
is usually chosen to minimise wire crossings, but the lookahead to get this right is quite extensive, so
even in the simple rocket program some unnecessary crossings occur. The next figure shows the visual
spaghetti created by not looking ahead this example is drawn from practically the first method
window opened in the first authors file of Prograph programs. The point here is not that it is possible
to make a horrible-looking program; thats true in any language. The point is that it needs a lot of look-
ahead to avoid making a mess. Worse, because of the high viscosity of these languages, it takes too
much time to clear it up, so the mess gets left.
Whitley and Blackwell (submitted) report the astonishing fact that one of their respondents in a survey
on LabView claimed to employ a junior just to tidy up the layout of the code:
`One respondent solved the problem by hiring extra help: "Recently, I hired a real nitpicker
to clean up my diagrams while they are in progress. That saves me a lot of time and my
customers get even neater diagrams than I would have done."'
That quotation tells us all we need to know about premature commitment and layout viscosity in box-
and-wire diagrams.
59
Figure 33 Visual spaghetti. To avoid this, the programmer has to look ahead. This
example is in Prograph, but the same problem occurs in LabView or any other box-
and-wire language.
Commitment to choice of construct
When a programmer chooses a syntactic construct and then adds code to the construct, it may turn out
subsequently that the construct is the wrong one (e.g. while should be changed to for). Therefore,
programmers may prefer to postpone commitment to a particular construct until the code has been
further developed.
Early versions of LabView suffered the problem that once a control construct had been selected, it was a
tiresome chore to change it. More recent versions of LabView solve the premature commitment
problem partly by postponing commitment, as well as by making it easier to change the decision
(specifically, it is now possible to draw a loop around existing code, just as a for-statement can be
inserted into a textual program.)
Abstraction Barrier and Abstraction Hunger
The abstraction barrier (the number of new abstractions to be understood by the novice), although
probably more than a simple spreadsheet system, is obviously less than say C++. Differences in this
respect depend more on the choice of primitives than on the requirements of the box-and-wire
notational structure itself.
Both languages have facilities for creating new operational abstractions. LabView allows the user to
choose to create new subroutines (called virtual instruments) but does not insist upon it: this language
is therefore abstraction-tolerant. Prograph, being object-oriented, requires some new abstractions to be
created for most purposes, making it abstraction-hungry.
Layout abstractions are not in evidence; these are no easy features for group operations when re-
arranging the components of a program, and much time can therefore be wasted.
We shall not discuss the two different abstraction managers here.
Secondary Notation
Judging by these visual programming languages, designers have not thought hard enough about the
need to escape from the formalism. Neither language has good facilities for commenting, for example.
They each have the ability to attach a comment to a single item, which may seem adequate at first sight,
but remember that a text language gives power to comment on a section of code, not just a single
operation:
60
// -- compute the vertical components -- //
Vaccel = Force*COS(Angle)/Mass Gravity
Vveloc = Vveloc + Vaccel
Vdist = Vdist + Vveloc
// -- compute the horizontal components -- //
Haccel = Force*SIN(Angle)/Mass
Hveloc = Hveloc + Haccel
Hdist = Hdist + Hveloc
Looking back to the LabView and Prograph programs, how would those comments be adapted?
Nor can these visual languages achieve the same degree of control over spatial layout that Basic
achieves over textual layout. The Basic program has been organised to bring related operations into
close proximity (which is what makes those comments so effective), but in the visual languages the
severity of the constraints on laying out the boxes to avoid clutter and crossings makes it impossible to
achieve the same results.
In short, secondary notation is a commodity in short supply in this particular design of visual language.
The notation itself uses up the available parameters, leaving nothing over for expressing anything else.
Bertin (1965) emphasises that in graphical notations the number of available variations is always
limited, and they must be allocated wisely: in these notations, it seems, too many of the variations have
been allocated to the notation itself.
Visibility and Juxtaposability
The visibility of data flow in the LabView language is excellent. Prograph dataflow is much less visible,
the large number of small windows obscuring the flow and hindering search for specific items. The
effect of the design decision to use separate windows in Prograph becomes all too apparent when one
looks at the proliferation of windows.
Where LabView in turn meets trouble, however, is in the juxtaposition of related control branches. The
two branches of a conditional cannot both be viewed at once, and although a single mouse click will
take you to the next branch, the effect on comprehensibility is very worrying.
Figure 34 A LabView conditional from the rocket program, showing both arms. In the
LabView environment, only one arm is visible on screen at any one time.
Conclusions
The most striking features of these comparisons are on the one hand, the extraordinarily high viscosity
of box-and-wire notations and the very poor facilities for secondary notation, and on the other hand, the
remarkable freedom from hidden dependencies. These give box-and-wire programs a very different feel
from textual languages, and indicate some targets for would-be improvers to aim for.
61
References (Pts 1 and 2)
(For references not found here, such as Roast (1998), see the Further Reading section.)
Bertin, J. (1981). Graphics and Graphic Information Processing. Berlin: Walter de Gruyter.
Blackwell, A. F. (1996b). Metacognitive Theories of Visual Programming: What do we think we are
doing? In Proceedings IEEE Symposium on Visual Languages, pp. 240-246.
Card, S. K., Moran, T. P. and Newell, A. (1983) The Psychology of Human-Computer Interaction . Lawrence
Erlbaum.
Cramer, M. (1990) Structure and mnemonics in computer and command languages. International Journal
of Man Machine Studies, 32 (6), 707-722.
Eisenstadt, M. (1993) Why HyperTalk debugging is more painful than it ought to be. In J. L. Alty, D.
Diaper and S. Guest (Eds.) People and Computers VII: Proc. HCI 93 Conference. Cambridge:
Cambridge University Press.
Fischer, G. (1988) Panel on Critical assessment of hypertext systems. Proc. CHI 88 ACM Conf. on
Human Factors in Computing Systems, p 224. New York: ACM.
Neilsen, J. and Molich, R. (1990) Heuristic evaluation of user interfaces. In Empowering people CHI90
conference proceedings. ACM Press, New York.
Payne, S. J. (1993) Understanding calendar use. Human-Computer Interaction, 8, 83-100.
Payne, S. J. and Green, T. R. G. (1986) Task-Action Grammars: a model of the mental representation of
task languages. Human-Computer Interaction, 2, 93-133.
Whitley, K. N. and Blackwell, A. F. (submitted for publication). Visual programming in the wild: a
survey of LabVIEW programmers. Available on request as Technical Report CS-98-03, Computer
Science Department, Vanderbilt University, Nashville, TN 37235, USA.
62
Part 3: Interactive Practice Examples
This section is devoted to practice exercises in the use of cognitive dimensions.
The exercises are intended for group discussion. Toy interactive applications have been created in pairs
to illustrate contrasting design choices; the group is invited determine what trade-offs in the
dimensions have been caused by the design difference. Thus there are two telephone systems, each with
its own advantages; two central heating controls, each with its own advantages, and so on. Toy non-
interactive applications notations for displaying information may also be included at a later stage.
The cognitive dimensions framework cannot be successfully applied until one is familiar with the
device or notation being studied. For this exercise, it would be best to be familiar with both of the
devices in a pair, to make it easier to see how advantages in one trade-off against advantages in the
other. The group should therefore start by exploring each of the two versions and devising tasks that
highlight their differences, looking especially for tasks that find an advantage for one of them. They
should then identify the major design difference, apply the cognitive dimensions framework, and
decide whether the trade-offs discovered fit the patterns described in earlier parts of this tutorial. A
final and important step is to think of real-life examples of each alternative, preferably from commercial
applications, and to decide whether they support the analysis made by the group.
Toy applications (widgets) have been used for several reasons:
a single design choice can be isolated and its consequences compared in alternative versions;
they are small enough to be delivered over the web, and will continue to be available for use on
other occasions for example, if attendees wish to pass on to others what have learned, these
examples can be re-used;
exploring and analysing a full-scale application would take too long for class use.
Web delivery
The widget examples are accessible via the URL
http://www.ndirect.co.uk/~thomas.green/devices/
Most of them are written using JavaScript, and these ones work with Netscape Navigator 3 on Macs,
PCs, and Suns. Make sure that JavaScript has been enabled (open Options -> Languages and check the
JavaScript box).
There is no guarantee that the JavaScript widgets work with any version of Internet Explorer or any
other browser.
One widget pair has been written in Java. At the time of writing there is some technical difficulty in
making this widget work with any browser except Suns applet viewer. Let us hope that is resolved
before you read this ......
63
Aims
The aims are:
to identify differences in their cognitive dimensions;
and to observe how the design choice has traded off some dimensions against others.
As a result of this exercise, you may decide that one widget is better than another for certain types of
activity, while the other is better for different types of activity.
What to do
Find the widgets;
Become familiar with each;
Locate any sub-devices;
Devise tasks that highlight their differences, looking especially for tasks that illustrate competing
advantages;
Note the effects of the medium (persistent or transient?) and consider the most likely types of
activity that users will engage in (exploratory creation / transcription / modification /
incrementation);
Apply the cognitive dimensions framework to each of them and to their sub-devices, if present (this
may prompt you to think of further tasks);
Note what the trade-off relationships are;
Try to think of further real-life examples of each version.
Remember!
The aim of this exercise is not to decide that one widget is better than the other. It is very unusual for
one design to be better than another for every conceivable purpose, so dont be simple-minded. Dont just
say this one is obviously better (are you sure its better for everything? ): instead, look for the cognitive
dimensions trade-offs.
The trade-offs may be easy to spot in these small examples, but they will be less obvious in a larger
example, so practice is beneficial.
Find Generalisations
An important part of the exercise is to think of further real-life examples of the same principles.
64
The Widgets
65
Form-Filling and Menu Choices
Users often have to make a set of choices that are inter-dependent, such that each choice influences the
acceptable options for all other choices. A notably repulsive case is answering questions over the
telephone while a clerk fills out a form; too late you realise that you should have said No at an earlier
stage...
That abstract scenario has been realised in these widgets as choosing courses for a meal. You have to
make your choices under the following constraint:
No two courses should contain the same ingredient. E.g. you should only have fish in one course. Each
vegetable counts as a unique ingredient, so if you have asparagus in one course you must avoid it in
subsequent courses.
At Disobliging Daves, the waiter gives you the choices for each course independently. At Cheerful
Charlies, you fill out a form for the whole meal at one go.
Disobliging Daves
Snap your fingers
to summon the
waiter. He, she or it
says:
When you press
OK, the waiter
offers you a
selection:
Click on your
choice, then press
the Ive chosen
button.
Cheerful Charlies:
The whole menu is on one form. Each course is a set of alternatives:
Just set up the form as you want it thats all you have to do.
66
Controlling the Heating Controller
Once a set of choices has been made for one occasion, users may have to make the same choices for
several other occasions: and not infrequently, for some of those occasions, the choices will be identical
or only slightly different, so the user has to play the same moves again.
That abstract scenario has been realised as setting the domestic heating controls for each day of the
week. The heating system to be set has a clock to control three heating periods, each with its own ON
and OFF times, so you can set the heating to be (say) on from 7 am to 8 am, then again from 6 pm to 8
pm, and once more from 10 pm to 11 pm. (If the on and off times for a period are the same, the heater
does not switch on for that period.)
One controller, the Balmoral, just sets each of the times. The other controller, the Alhambra (based on a
real system, the Horstmann Channel Plus), has an added facility, a Copy function. If (say) Tuesday is
selected, then the effect of Copy is to make all Tuesdays on-off times identical to Mondays times and
then to select the next day, Wednesday. Pressing Copy 6 times sets all the on-off times to be the same,
right through the week.
The Balmoral Model
In this model, Plus and Minus control the
setting of the selected item (in the
illustration here its the 06:00, shown
shaded; its highlighted on the screen).
The Advance button selects Period -1-On,
Period-1-Off, ditto for Periods 2 and 3,
then selects next day: at that point, the
day is shown highlighted with no times
visible (as shown in the diagram of the
Alhambra system). Pressing Plus will
then move immediately to the following
day.
The Alhambra Model:
As for Balmoral, except that while the
day is selected (as shown here), pressing
the Copy button will set all that days on-
off times to those of the preceding day
and will then select the next day.
67
Number Games with Telephones
Telephones and telephone numbers have been mentioned several times in this tutorial. Here is one to
analyse, based on the British Telecom Relate 80 handset, which has a memory that can store up to ten
numbers.
There is only one version of this widget just compare its functioning with and without the memory.
Making a call
Lift the phone by pressing the top left radio button;
Then press the number keys in the appropriate order.
Thats all you have to do.
Hang up by pressing the top right-hand radio button.
Using the memory
(The following are essentially the manufacturers instructions.)
To store a number:
pick up the handset
press STORE
key in the telephone number you wish to store
press STORE again
press one of the keypad buttons (0 - 9) (depending on where you wish to store it)
the number is now stored
To dial a stored number:
pick up the handset
press RECALL
press one of the keypad buttons for the number you want to dial
68
The Memory Label
The BT Relate 80 comes with a holder for a piece of card on which
you can note the stored numbers. The manufacturers instructions
say:
Write the name of the person whose number is
stored on the memory label.
(Use a pencil, so it can be changed if necessary.)
(How does the recommendation to use a pencil affect the cognitive
dimensions?)
For the interactive widget, of course, you should type in the stored
numbers.
69
Tiling in Style
The telephone example illustrated one form of re-use. Here is another, from a different domain in which
the medium has different characteristics, and the probable user activities are different.
This is a an editor for use when defining tile floors. In one scenario we have envisaged, the editor is to
be used by a Tile Consultant. A customer will supply a sketch of the desired design or maybe give an
verbal informal specification, and the Tile Consultant will define a suitable design.
In an alternative scenario, the customer is a direct end-user of the tile editor, exploring possibilities
without any form of preliminary sketch.
The Greenwich model
In this model, the user defines any number of tile shapes, then arranges these shapes to match a sketch
that has been supplied by the customer.
Each shape of tile must be defined as a new abstraction before it can be used in the design; thereafter,
that shape can be freely re-used.
Redefining a shape changes all the instances that have already been used.
customer sketch
instance of
square
70
The Harrogate Model
An alternative version of the editor allows tiles to be defined using direct manipulation, so that no
abstraction needs to be created. Each tile is drawn (and modified) individually.
customers
sketch
a tile being drawn
add or subtract a
vertex
71
Further Reading
Following is a partial list of publications wholly or partly related to research on cognitive dimensions.
Buckingham Shum, S. and Hammond, N. (1994) Argumentation-based design rationale: what use at
what cost? International Journal of Human-Computer Studies 40 (4), 603-652.
http://kmi.open.ac.uk/~simonb/DR.html
Includes a cognitive dimensions analysis of visual design rationale tools
Gilmore, D. J. (1991) Visibility: a dimensional analysis. In D. Diaper and N. V. Hammond (Eds.) People
and Computers VI. Cambridge University Press.
Green, T. R. G. (1989). Cognitive dimensions of notations. In People and Computers V, A Sutcliffe and L
Macaulay (Ed.) Cambridge University Press: Cambridge., pp. 443-460.
The original paper
Green, T. R. G. (1990) The cognitive dimension of viscosity: a sticky problem for HCI. In D. Diaper, D.
Gilmore, G. Cockton and B. Shackel (Eds.) Human-Computer Interaction INTERACT 90. Elsevier.
Green, T. R. G. (1991) Describing information artefacts with cognitive dimensions and structure maps.
In D. Diaper and N. V. Hammond (Eds.) Proceedings of HCI91: Usability Now, Annual Conference of
BCS Human-Computer Interaction Group. Cambridge University Press.
Green. T. R. G., Winder, R., Gilmore, D. J., Davies, S. P. and Hendry, D. (1992) Designing a cognitive
browser for object-oriented programming. Artificial Intelligence and Simulation of Behaviour Quarterly,
Issue 81, 17-20.
Green, T. R. G. (1996) The visual vision and human cognition. Invited talk at Visual Languages '96,
Boulder, Colorado. Abstract only appears in: Proceedings of 1996 IEEE Symposium on Visual
Languages. Eds: W. Citrin and M. Burnett. Los Alamitos, CA: IEEE Society Press, 1996.
http://www.ndirect.co.uk/~thomas.green/workStuff/VL96Talk/VLTalk.html
Adapts much of the content of Green and Petre (1996) for web presentation
Green, T. R. G. & Petre, M. (1996) Usability analysis of visual programming environments: a 'cognitive
dimensions' framework. J. Visual Languages and Computing, 7, 131-174.
gzip (180K): ftp://ftp.mrc-apu.cam.ac.uk/pub/personal/thomas.green/VPEusability.ps.gz
uncompressed (2.2 Mb): ftp://ftp.mrc-apu.cam.ac.uk/pub/personal/tg/VPEusability.ps
A lengthy analysis of selected visual programming languages: the most extensive presentation so
far in print
Green, T. R. G. & Benyon, D. (1996) The skull beneath the skin: entity-relationship models of
information artifacts. International Journal of Human-Computer Studies, 44(6) 801-828
gzip (80K): ftp://ftp.mrc-apu.cam.ac.uk/pub/personal/tg/Skull_v2.ps.gz
uncompressed (305K): ftp://ftp.mrc-apu.cam.ac.uk/pub/personal/tg/Skull_v2.ps
PDF (121K): http://www.dcs.napier.ac.uk/~dbenyon/IJHCSpaper.pdf
72
Goes a little way to formalising the dimensions in terms of extended entity-relationship models
Hendry, D. G. and Green, T. R. G. (1994) Creating, comprehending, and explaining spreadsheets: a
cognitive interpretation of what discretionary users think of the spreadsheet model. Int. J. Human-
Computer Studies, 40(6), 1033-1065.
Includes a cognitive dimensions analysis of spreadsheet pros and cons
Hendry, D. G. and Green, T. R. G. (1993) CogMap: a visual description language for spreadsheets.
Journal of Visual Languages and Computing, 4(1), 35-54 (special issue on cognitive aspects of visual
languages)
How to improve spreadsheets, according to cognitive dimensions analysis (particularly with
regard to role-expressiveness, not explored here)
Hendry, D. G. and Harper, D. J. (1997) An informal information-seeking environment. Journal of the
American Society for Information Science, 48(11):1036-1048.
Considers information seeking as opportunistic design of queries, to which cognitive dimensions
can be applied, and relates usability of systems to the ease of forming and modifying the queries.
Lavery, D., Cockton, G., and Atkinson, M. (1996). Cognitive Dimensions: Usability Evaluation
Materials, Technical Report TR-1996-17, University of Glasgow.
http://www.dcs.gla.ac.uk/~darryn/research/publications/TR-1996-17/
Part of a study comparing effectiveness of different evaluation methods; need t contact the author
for full details
Modugno, F. M., Green, T. R. G. and Myers, B. (1994) Visual programming in a visual domain: a case
study of cognitive dimensions. In G. Cockton, S. W. Draper and G. R. S. Weir (Eds) People and
Computers IX: Proc. BCS HCI Conference. Cambridge University Press.
Analyses a programming-by-demonstration environment, leading to a redesign
Petre, M. and Green, T. R. G. (1990) Where to draw the line with text: some claims by logic designers
about graphics in notation. In D. Diaper, D. Gilmore, G. Cockton and B. Shackel (Eds.) Human-
Computer Interaction INTERACT 90. Elsevier.
Petre, M. and Green, T. R. G. (1992) Requirements of graphical notations for professional users:
electronics CAD systems as a case study. Le Travail Humain, 55(1), 47-70
Roast, C.R. and Siddiqi, J. I. (1996) Formally assessing software modifiability. In C. R. Roast and J. I.
Siddiqi, editors. BCS-FACS Workshop on Formal Aspects of the Human Computer Interface, Sheffield
Hallam University, 10-12 September 1996, Electronic Workshops in Computing. Springer-Verlag,
1996.
PS: http://homepages.shu.ac.uk/%7Ecmscrr/PAPERS/fahciOnLine.ps
Roast, C.R. (1998) Modelling premature commitment in information artifacts.. To appear in EHCI.
Proceedings (ed: L. Nigay).
http://homepages.shu.ac.uk/%7Ecmscrr/PAPERS/premcom.ps
Roast, C.R. and M.B. zcan. (1997) Cognitive dimensions applied to modifiability within an integrated
prototyping environment. In R.Osborn and B.Khazaei, editors, The Psychology of Programming,
PPIG97 Workshop Proceedings, pages 17-30. Computing Research Centre, Sheffield Hallam
University,
PS: http://homepages.shu.ac.uk/%7Ecmscrr/PAPERS/ppig97.ps
Roast, C.R. and Siddiqi, J. I. (1996) The formal examination of cognitive dimensions. In A.Blandford
and H.Thimbleby, editors, HCI96 Industry Day & Adjunct Proceedings, pages 150-156, 1996.
Roast, C.R. and Siddiqi, J. I. (1996) Relating knock-on viscosity to software modifiability. Proceedings of
OZCHI 96, Hamilton, New Zealand, pages 222-227, 1996
PS: http://homepages.shu.ac.uk/%7Ecmscrr/PAPERS/ozchi96.ps
73
Shum, S. (1991). Cognitive dimensions of design rationale. In D. Diaper and N. V. Hammond, (Eds.)
People and Computers VI: Proceedings of HCI'91, 331-344. Cambridge: Cambridge University Press.
http://kmi.open.ac.uk/~simonb/DR.html
Simos, M. & Blackwell, A. F. (1998). Pruning the tree of trees: The evaluation of notations for domain
modelling. In J. Domingue & P. Mulholland (Eds.), Proceedings of the 10th Annual Meeting of the
Psychology of Programming Interest Group, pp. 92-99.
This discusses the more abstract requirements of notations used in the field of Domain Modelling,
and considers how theories of metaphor in notation, as well as Green's Cognitive Dimensions, can
be used to analyse them.
Stacey, M. K. (1995) Distorting design: unevenness as a cognitive dimension of design tools. In G.
Allen, J. Wilkinson & P. Wright (eds.), Adjunct Proceedings of HCI'95. Huddersfield: University of
Huddersfield School of Computing and Mathematics, 1995.
http://www.mk.dmu.ac.uk/~mstacey/Documents/uneven.htm
Proposes a new dimension
Tweedie, L. (1995) Interactive visualisation artefacts: how can abstractions inform design? In Kirby, M.
A. R., Dix, A. J. and Finlay, J. E. (Eds.), People and Computers X, Proc. HCI95 Conference. Cambridge
University Press.
Proposes an approach derived from cognitive dimensions specialised for interactive visualisation
tools
Wilde, N. P. (1996) Using cognitive dimensions in the classroom as a discussion tool for visual
language design. Electronic Proceedings CHI 96, Eds R. Bilger, S. Guest, M. Tauber.
http://www.acm.org/sigs/sigchi/chi96/proceedings/shortpap/Wilde/wn_txt.html
How real are the cognitive dimensions? Do they help teaching?
Wood, C. C. (1993) A cognitive dimensional analysis of idea sketches. Cognitive Science Research Paper
275, School of Cognitive and Computing Sciences, University of Sussex.
ftp://ftp.cogs.susx.ac.uk/pub/reports/csrp/csrp275.ps.Z
The cognitive dimensions framework was used to interpret a study of collaborative idea sketching,
giving a cohesive and comprehensive characterisation of the pertinent properties of idea sketches,
as gathered from interviews and observation with 128 academics.
Yang, S., Burnett, M. M., DeKoven, E. and Zloof, M. (1997) Representation design benchmarks: a
design-time aid for VPL navigable static representations. Journal of Visual Languages and Computing, 8
(5/6), 563-599.
Describes simple but effective metrics for selected dimensions, leading to redesign of two different
languages.
74
75
Index of examples
alarm clock, 36
benchmarks, 44
C, 27
CAD, 26
calendar, 31
calendars, 47
Cameras, 35
circuit diagram, 41
circuit diagrams, 40
circuit notations, 30
comment, 32
Fax machines, 35
Flowcharts, 27
folk psychology, 52
formalisms, 31
form-based system, 36
front panel, 30
global find-and-replace, 26
grouping, 26
HyperTalk, 27
in-car audio, 39
In-car radios, 35
macro, 26
meeting organisers, 47
menu structure, 35
menus, 39
object hierarchies, 26
Pencils, 41
Powerpoint templates, 27
Programming By Example, 28
scripting languages, 36
Smalltalk, 27
Spreadsheets, 27, 37, 40
Style sheets, 26
telephone memories, 26
telephone number, 30, 34
visual programming language, 39
Word styles, 26