See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/220534159

Real and Virtual Spaces: Mapping from Spatial Cognition to Hypertext

Article · January 1990

DOI: 10.1080/09558543.1990.12031179 · Source: DBLP

CITATIONS READS
70 170

1 author:

Simon Buckingham Shum

University of Technology Sydney
290 PUBLICATIONS   5,889 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Scaling the Provision of Personalised Learning Support Actions to Large Student Cohorts View project

Mapping the curriculum View project

All content following this page was uploaded by Simon Buckingham Shum on 20 May 2014.

The user has requested enhancement of the downloaded file.

 Real and Virtual Spaces:
Mapping from Spatial Cognition to
Hypertext

Simon Shum

Technical Report EPC-1993-115

Reproduced by permission from

Hypermedia (now published as The New Review
of Hypermedia and Multimedia),
Volume 2 (2), 1990,
Taylor Graham Publishing, 500 Chesham
House, 150 Regent St., London W1R 5FA.

Copyright © Taylor Graham Publishing 1990.

Rank Xerox Research Centre

Cambridge Laboratory
61 Regent Street
Cambridge CB2 1AB

Tel:+44 1223 341500

Fax:+44 1223 341510
 Published in: Hypermedia (1990), 2(2), 133-158. Taylor Graham, London

Real and Virtual Spaces:

Mapping from Spatial Cognition to Hypertext
Simon Shum

Human-Computer Interaction Group

Department of Psychology
University of York
York YO1 5DD
UK
[email protected]

Introduction 1
Potential applications domains 2
Spatial Cognition Research And Hypertext 2
Introducing cognitive maps 2
The content of cognitive maps 4
Locational information 4
Attributional information 6
Acquiring spatial knowledge 7
Accessing users’ cognitive maps 7
Cashing The Ideas: What Does Spatial Imagery Buy Us? 8
Euclidean and virtual-space metaphors 8
Thinking in three dimensions 10
Defining a space with dimensions 11
The layered-space model 12
Scope of the layered-space model 18
Virtual realities 18
Conclusions 19
Acknowledgements 19
References 20
 Real and Virtual Spaces:
Mapping from Spatial Cognition to Hypertext
Simon Shum

Human-Computer Interaction Group †

Department of Psychology
University of York
York YO1 5DD
UK
[email protected]

ABSTRACT

Parallels are frequently drawn between navigating through everyday spatial environments and
information systems, hypertexts being a particular case in point. This paper examines the
cognitive mapping theory often borrowed implicitly from spatial cognition research, which
has a bearing on the appropriateness of using spatial imagery in hypertext. Conceptual
differences between euclidean and virtual spaces are identified, and ways considered in which
to make information spaces more coherent. A demonstration hypertext browser is described,
incorporating some of the cognitive principles discussed.

INTRODUCTION

In recent years, a number of researchers have made use of the analogy that navigating
information spaces is like navigating in the real world. Essentially, it is argued that
maintaining a sense of direction within a large, information system is similar to problems
faced in orienting within real world environments. [5] Hypertext, in particular, is often
viewed in spatial terms, its networked nature making it particularly prone to spatial language
in describing user activity, e.g. jumping from node to node, getting lost or disoriented,
descending a hierarchy. Akscyn and McCracken [6] describe their KMS hypertext system
metaphor as “a universe of connected spaces through which users rapidly travel, like pilots
navigating spacecraft in the real universe.” Taken to its logical conclusion, the use of three
dimensional (3D) imagery within the user interface has intuitive appeal, especially with the
availability of powerful graphical interfaces. Certainly, some researchers (e.g. Poltrock et al
[7]) see 3D displays as holding considerable promise as a future medium for human-computer
interaction. Durham [8] also reports a trend in software products using 3D spatial
representations for abstract data, whilst Jones [9] has extensively reviewed the move towards
graphical representation within knowledge engineering.

The term ‘metaphor’ is used widely to mean a number of different things. ‘Spatial metaphor’
may be invoked to refer to a 3D display, a general similarity such as ‘traversing hypertext
links is like going down tunnels,’ or a well developed interface which uses commands and
displays all relating to the metaphorical world (e.g. Hammond and Allinson’s Travel
metaphor [10]). ‘Conceptual model’ is perhaps a more precise term for the use of abstract 2D
and 3D spaces, unless presented as ‘airspace,’ for instance, with accompanying commands
(take off; change altitude) and objects (aircraft; airports). Conceptual models are, of course,
no less capable of creating compelling imagery within the user’s mind, especially when
coupled with graphical displays. For clarity, therefore, the terms model, and metaphor will be

† This work was carried out as doctoral research whilst at Rank Xerox EuroPARC.

1
used with care, whilst recognising the close relationships between them. The term ‘spatial
imagery’ is useful as a general descriptor, and reflects something of the subjective experience
which characterises the use of interfaces simulating aspects of real space.

The aim of this paper is to introduce some of the issues in the spatial cognition literature
which would appear, at least initially, to have relevance to the design of the cognitive
interface to information systems, hypertext being of particular interest.

The field of spatial cognition has grown over the last thirty years as geographers, architects,
anthropologists, urban planners, and psychologists have attempted, from their different
perspectives, to characterise those attributes of the spatial environments in which we live
which make them such a rich source of experience. This highly multidisciplinary literature is
not unlike Human-Computer Interaction in some respects, dealing at many levels with the
interaction between humans and the environments they create for themselves.

In the following analysis, cognitive attributes of space at both the macro and micro level are
examined, that is, the familiarisation with, and navigation of large spaces, and our use of local
space for organising information around us. Attention focusses on questions such as: How are
virtual spaces like real, euclidean spaces (i.e.. occupying a space with X, Y, Z dimensions)—
indeed, do we want them to be? To what extent can research into spatial orientation and
memory be extrapolated to hypertext? An understanding of the cognitive issues involved
allows us to make more informed judgements over the use of spatial imagery.

Potential applications domains

When considering how knowledge of spatial cognitive mapping could support hypertext
design, one would expect, or at least hope for interesting perspectives on those most
commonly cited problems in the hypertext literature of maintaining orientation, knowing
what information is available, knowing how to get there, and remembering where visited
information is (‘Where am I? What’s out there? How do I get there? Where was it?’). These
tasks are of relevance to all domains concerned with allowing users to familiarise themselves
with a new body of knowledge. Users may have no idea of how much complexity to expect,
or what an appropriate structure is; alternately, they may bring with them various
preconceptions in these respects. This is the case whether we are talking about students
learning coursework, employees browsing the company tree, or tourists seeking information.
The following analysis of spatial imagery affects issues of orientation, clarity of display, and
encoding of position, all of which impact the design of the sorts of systems listed above.
Consideration of the effect of 3D representation on navigation has a direct bearing on the
design of hypertext databases.

SPATIAL COGNITION RESEARCH A ND HYPERTEXT

This section focusses on the cognition associated with the acquisition and development of
cognitive maps. The term ‘hypertext’ is used simply to mean a system which allows the
traversal of explicit links— differences between hypertext models (c.f. Kahn and Myerowitz
[11]) will not be addressed. Having introduced some spatial cognition concepts, attention
then turns to the implications of enhancing the ‘spatiality’ of hypertext further.

Introducing cognitive maps

Neisser [12] views cognitive maps as ‘orienting schema,’ mental representations which
actively seek and integrate spatial information. These schema constitute one part of a three

2
component cycle, whereby schema direct action, and the resultant experience of the world
modifies the schema. Downs and Stea [13] define the acquisition of cognitive maps, as:

“. . . a process composed of a series of psychological transformations by which an

individual acquires, codes, stores, recalls and decodes information about the relative
locations and attributes of phenomena in his everyday spatial environment.”

They emphasise that CMs comprise not only spatial knowledge, but all of the other attributes
of ‘place,’ for example accessibility—the effort to reach a place; sensory experience—sounds,
sights, smells; emotional connotations— ‘I nearly crashed here’; evaluative perceptions —
‘The food’s no good.’ These are returned to shortly.

A key point to note is that cognitive maps are functionally equivalent to a cartographic map,
as opposed to structurally equivalent, such as a set of complete representations at different
scales. Detail varies according to subjective importance: ‘Cognitive maps are complex, highly
selective, abstract, generalised representations in various forms’ (Downs and Stea [13] p. 18).

As a simple framework for thinking about the abstraction process involved in mapping,
Downs and Stea define a mapping as the transformation of an object set into an image set via
a mapping function. The object set is the experienced world, whilst the mapping function
determines which attributes of that world to preserve in the image set, by abstracting over
content. How can this characterisation be applied to hypertext? Figure 1 illustrates some
conceptual parallels. It shows the object›image set mapping operating at different levels , e.g.
hypertext›browser, browser›user’s cognitive map, so that the image set from one mapping
becomes the object set of the next.

Object Set (Mapping Function) Image Set

physical cartography map

Physical world
Space
map problem solving cognitive map
activities of physical world

Hypertext hypertext browser algorithm browser

Space
browser user task demands/ cognitive model
network exploration of network

Figure 1. Transformation from an object set to an image set via a mapping function:
parallels between the real world and hypertext. The diagonal arrows reflect the
formation of cognitive maps from direct interaction with the world/hypertext, as well
as via the use of maps/browsers.

Spatial attributes subject to the mapping function include shape, size, relative distance and
direction; in hypertext, properties include colour, node content and size, and links; thus, a
cognitive map might preserve relative position and link direction (e.g. A›B but not B›A) but
filter out other information like size and colour, either intentionally (they are not considered
important), or simply due to memory limitations. The mapping function which hypertext
users employ when using the system will be determined by what they consider significant—a

3
map is only as effective as it is helpful to its owner, which means that a map cannot be
evaluated in terms of its objective completeness, but by its support to the user.

The content of cognitive maps

What knowledge do cognitive maps actually embody? Downs and Stea propose that to
operate effectively in a spatial environment, we need access to two basic classes of
information: the locations, and the attributes of phenomena, and within each of these they
identify two subclasses of information, distance and directional information, and descriptive
and evaluative attributes, respectively. This typology is summarised in Figure 2. It is
interesting to consider the role each of these plays in the context of using hypertext
environments.
Spatial Phenomenon

Locational Information Attributional Information

(where is it?) (what’s it like?)

Distance Direction Descriptive Evaluative

Attributes Attributes

Figure 2. Classes of information used in cognitive maps (summary of Downs and Stea
[13]). Note there is no claim that cognitive maps are structured like this simply that
the content can be classified as such.

Locational information

Figure 3 shows an example of the structure of locational information. Note that distance is
not only quantified in metres and miles, but by the cost in terms of time and money which
exerts an influence over the perceived effort involved in making a journey.

Spatial Phenomenon
Vivaldi’s Four Seasons at the Concert Hall

Locational Information Attributional Information

(where is it?) (what’s it like?)

Distance Direction

- 60 miles - due south

- 1 hour by car - half-way down 7th
- over the river - downtown
- three blocks from John's
- expensive by taxi

Figure 3. An example of locational information for a place. Note that concepts of Distance
may incorporate subjective judgements of cost to reach the destination; Distance
and Direction are expressed in terms of the spatial constructs in which the person
normally thinks

4
Thus, ‘cognitive distance’ to the hypertext user is expressed as concern over system response
time, ease of returning to the previous node, the number of link traversals/mouse clicks
involved. Direction is a more difficult concept to capture in hypertext. Figure 4 shows the
equivalent structure to Figure 3, but in the context of hypertext. It can be seen that the
concepts of distance and direction are relative as opposed to absolute in nature. The reasons
for this are discussed in more detail later, when the number of spatial metaphors and models
are examined.
Spatial Phenomenon
Digitised sequence of cell-division

Locational Information Attributional Information

(where is it?) (what’s it like?)

Distance Direction

- system takes only 2 secs - forwards

- 3 mouse clicks from here - backwards
- two levels from here

Figure 4. An equivalent to Figure 3, in the domain of hypertext – see text for details.

Hypertext designers have been aware for some time of the problems of systems which
effectively increase the ‘distance’ users must travel. For instance, the KMS system [6] makes
speed of link traversal the primary system requirement, aiming for 0.5 sec response on
average, and Wright [14] has emphasised the need for users to know what is waiting for them
at the other end of a link—what is the cost of jumping? A facility which all systems should
provide is a ‘go back to last node, key, the hypertext equivalent of ‘undo.’

Elsewhere in the literature, distinctions in navigational knowledge have been made between
route and map knowledge. Thorndyke [15] characterises route knowledge as having a
context-dependent, ego-centred perspective, that is, it can only direct action within a limited
set of situations (they must be on the route) based on recognition of cues en route, like
landmarks, e.g. ‘go through three sets of lights, left at the shops, and up the hill to the green
gates’).

On its own, route knowledge does not tell users where they are in relation to the rest of the
environment; hypertext route knowledge allows one to quote a linear sequence of linked
nodes (A›B›C›D), and perhaps links branching from nodes passed on the way (B›X and Y,
C›Z). It is impossible to say whether a node could be reached faster via another route
(A›?›D), or from somewhere else in the network (?›D). Whilst on its own, fragmented route
knowledge does not inform on the wider picture of spatial relationships, the next step is to
integrate knowledge about different routes to infer nodal relationships if the different routes
share landmarks, a process utilised in the spatial cognition model proposed by Siegel and
White [16].

This eventually leads to map knowledge, which as the name suggests, operates at the level of
global spatial relationships. This is in contrast to route knowledge, taking a world-centred
frame of reference, and operates independent of context (e.g. ‘go west, walking parallel with
the shore, then north for half a mile’). Often deriving from secondary sources such as printed
maps, possession of map knowledge enables us to select from alternative routes, generate
novel ones, and think more flexibly because the knowledge is not limited to a set of highly
specific conditions (there is a similar difference between procedural and declarative device
knowledge [17]).

5
What role do these two types of knowledge play for hypertext users? Map knowledge is
undoubtedly crucial to successful navigation. Navigation and coverage of material is far more
efficient when the structure and extent of the network is known, one effective means being
the provision of a browser [10]. What is the role of route knowledge in hypertext? Do users
navigate by internalising, and then ‘running off’ sequences of link traversals, such as, ‘To get
to the screen on swimming, select leisure pursuits and then sport, and then scroll two screens
to click on aqua-activities’? From observation and personal experience, one would argue that
procedural knowledge of this sort is used, though to what extent in proportion to map
knowledge most likely differs between task domains and systems.

Attributional information

So far, the focus has been on the spatial information encoded within cognitive maps.
However, our representations of environments constitute more than featureless sets of
relationships; we bring a rich set of expectations about how places should look, sound, feel,
and smell, that is to say, the attributes by which they are recognised, and evaluated.

Descriptive attributes, as labelled by Downs and Stea, represent those sensory features which
identify a place; evaluative attributes are the evaluative tags we attach to our conceptions of
places, in order to decide whether or not a visit is likely to meet our needs. It is these
attributional factors which make cognitive maps so personal. Figures 5 and 6 show the
attributional information which might exist for the two hypothetical places appearing in
Figures. 3 and 4.
Spatial Phenomenon
Vivaldi’s Four Seasons at the Concert Hall

Locational Information Attributional Information

(where is it?) (what’s it like?)

Descriptive Attributes Evaluative Attributes

(what's there?) (is it worth visiting?)

- tinted glass building - expensive famous soloist

- people in evening dress - cultured atmosphere
- green and blue Four - terrible food
Seasons posters

Figure 5. Attributional information for a real world place

Spatial Phenomenon
Digitised sequence of cell-division

Locational Information Attributional Information

(where is it?) (what’s it like?)

Descriptive Attributes Evaluative Attributes

(what's there?) (is it worth visiting?)

- first cell is large and in top - interesting for 1st year students
left of screen - shows off system well
- dialogue box says 'click to
start'
- pink background screen

Figure 6. Attributional information for a hypertext node

6
Within this framework, debates about the visual interface (e.g. Hardman [18]) can be seen as
addressing the salience of descriptive attributes—how easily can the user bring to mind an
image of the target information, or recognise it whilst browsing? If used judiciously [19]
other media like sound and video should make this easier by enriching the set of distinctive
retrieval cues.

Turning to evaluative attributes, the fact that locations often take on significance suggests a
potentially useful role for marking and annotating nodes. Given hypertexts of substantial size,
this will most usefully be in a system-readable form, so that the tool can support the user in
pulling out nodes with different classes of mark. Monk [20] describes a personal browser
which uses the number of visits to a node as an index to its subjective importance to the user,
and then displays those nodes visited over a user-defined threshold frequency. It is a difficult,
though worthwhile challenge to designers, to develop effective tool-support for more
elaborate ‘tagging’ schemes (see also Lansdale [21]), so that the strategies we use in the
physical world transfer to hypertext.

Acquiring spatial knowledge

Downs and Stea [13] distinguish between different sources of spatial information, labelling
them direct and vicarious. Direct sources refer to experiences at ‘ground level’ (experiencing
the world at first hand) which are generally a reliable source of information about the
environment. Compared to ‘learning by doing,’ vicarious sources are second-hand
descriptions of places, such as verbal reports, maps, brochures, and television. According to
the purpose of the communication (e.g. tourism; civil engineering; a military operation), each
source abstracts over different kinds of information so that only appropriate kinds are left,
targetting a specific audience.

In hypertext, the acquisition of information via direct sources corresponds to personal

interaction with the hypertext, whether navigating at ‘node level’ (inspecting the contents of
nodes and following links) or at ‘browser level’ work when only the titles of nodes are
displayed (the more common mode of working when manipulating structure).

The characteristic of a vicarious source is that it is a view of the environment through

another’s eyes—detail has been filtered. In this sense, graphical browsers of the structure may
fall into this category, as users acquire map knowledge without having to actually visit the
places themselves. Map views clearly constitute a vicarious source when supplied with the
hypertext as a read-only display, as the designer effectively imposes a view of the content on
the user, rather than the user controlling the filtering of node/link types (static maps may of
course be quite adequate for the task domain). Other vicarious sources of hypertext
information will depend on the environment of use, and should not be underestimated. In
groupwork situations, for instance, one would expect to find that students working together
with courseware tools like Intermedia [22] or the Hitch-Hiker’s Guide [10] would share
information with each other, increasing individual coverage of the relevant material.

Accessing users’ cognitive maps

Assuming the existence of internal representations of spatial knowledge, researchers have

understandably been interested to try to elicit and analyse them. In his seminal work on
cognitive maps, Lynch [23] obtained inhabitants’ sketched maps of cities, and analysed them
in terms of landmarks (salient visual features), paths (traversable routes), nodes (path
intersections), districts (geographically defined regions) and edges (visibly defined
boundaries to districts). Whilst the first three of these will be strikingly familiar to the
hypertext community, districts with boundaries are less so. It is not always easy to explicitly
mark regions on graphs or hierarchies, even if it makes sense to demarcate different sections.

7
Given that regions clearly play an important role in everyday chunking of spatial information
in memory—for instance, structuring concepts such as downtown, south of the river, second
floor, top shelf, left margin—it should be of some concern to hypertext designers tackling
disorientation problems to capitalise on this human propensity to impose structure.

In some task domains, a graphical metaphor is one means of introducing regions

unobtrusively: a medical hypertext could make natural use of the human body as the top level
map, in which the head, torso and limbs form explicit regions which group related
information; the same might be said of a car, or building. If, however, material is of an
abstract nature, the author does not have resources to create a graphical interface, or different
views of the information are required for which the boundaries are unsuitable, the solution is
less obvious. The layered space model described later uses clearly bounded regions as a
means of aiding spatial memory, but by employing an abstract conceptual model (rather than
a concrete metaphor) exemplifies a domain independent solution.

Methodologically, the use of sketch-maps is now common in spatial cognition research, and
Lynch’s indices have been frequently used. Canter ([24], Chap. 4) provides a fascinating
account of the pragmatics of obtaining and analysing sketch maps, as well as the cognitive
theory underpinning this method. A common finding, for instance, is that places which have
no personal importance, or are felt to be irrelevant are often omitted or underemphasised in
sketch maps, a phenomenon which one might anticipate in user sketches of hypertext
structure.

An alternative means of accessing cognitive representations of spaces is to ask subjects to

arrange cards or models representing buildings and other objects. Within the context of
hypertext, this technique is particularly appropriate when the variable of interest is recall of
structure as opposed to content. Edwards and Hardman [25], asked hypertext users to arrange
cards (representing nodes) on a whiteboard and draw in links, demonstrating how an existing
technique can be successfully recruited and tailored for hypertext research purposes.

CASHING T HE I DEAS : WHAT D OES S PATIAL I MAGERY BUY US ?

A number of conceptual issues in spatial cognition have now been introduced, with the
emphasis on cognitive maps, these being the representations through which we make sense of
the spatial world, and which drive all consequent spatial behaviour. Several striking parallels
have been identified in the different classes of knowledge which exist in physical and
hypertext environments, which is then integrated into our own highly subjective
representations.

This section considers from the cognitive mapping perspective how hypertext might be
enriched spatially, comparing a number of spatial conceptual models which have been
proposed. In doing so, we can characterise them in terms of (i) the number of dimensions
they use (ii) whether they invoke notions of virtual or euclidean space, and in the latter case,
(iii) the extent to which absolute position carries meaning. These terms will be elaborated as
they are met.

Euclidean and virtual-space metaphors

In a paper on information workspaces, Card [26] points to the generalised tendency for
information processing (both machine and human) to focus on and switch between relatively
small clusters of information’ reflecting the tasks to be performed. He then critiques several
interfaces for the support which they offer to this work-pattern. Given the emphasis on
clustering, it is not surprising that the interfaces are also highly spatial in nature, but it may
be this difference in perspective— task basis vs. spatial basis—which is crucial to

8
understanding how spatial location can be used, and abused, as an organising principle . The
following design case study is detailed as it raises a number of issues which when developed
relate to the use of spatial imagery in hypertext.

Let us consider 2D metaphors based on the desktop metaphor. ‘Landscapes,’ providing an

infinitely large area to position objects, were devised to overcome the severe screen-space
limitations of conventional desktops; examples of the basic idea have been implemented in
several systems, including BigScreen [27] and Spatial Data Management System (SDMS)
[28]. Landscapes are two dimensional euclidean spaces, and as such, possess properties of
distance and direction (because objects have consistent positions relative to each other);
depending on the system, the user scrolls, jumps, or flies over the area to find objects. Since
the locational information available comprises both distance and directional components
(unlike basic hypertext—see Figure 4), cognitive maps will evolve in a similar fashion to
mapping an environment in the real world.

The critical lesson to note from the BigScreen work is that clusters of applications windows
grew on the basis of the tasks they supported, and use of an enormous desktop in fact became
more of a hindrance than a help, indeed, “task switching seemed to have a non-spatial
representation in the user’s mind” (Card [26] p. 6). Decisions about where to position new
material so that it could be easily located became an overhead, as did navigating over the
landscape to switch tasks; indeed, rapid jump facilities had to be added in order to switch
location rapidly. Following BigScreen, Henderson and Card built the ‘Rooms’ interface [27],
which allows the user to create and switch between multiple screen-sized desktops (Rooms),
enabling applications windows to be grouped according to the tasks they served. It was the
creation of clusters which was most important, rather than the medium by which this was
achieved.

If we look at the level of the individual Room/desktop, it can be seen that unlike the higher
level relationships between Rooms, organisation of desktop objects is euclidean, as in a
physical area. Malone [29], in an analysis of use of office space for storing information,
concludes that desktop interfaces inherit useful properties from their physical counterparts,
which non-spatial (i.e. non-euclidean-spatial) information systems lose. Malone encountered
the organising strategies that many of us use everyday, namely, positioning objects (piles of
papers; books; letters etc.) in particular places to act as reminders to act—location perhaps
determining priority, and leaving objects out for easy access, or as a useful way to defer
classification of ambiguous items into some more rigid filing system. To expand this last
point, it should also be noted that spatial representation is an ideal way to represent
relationships which are unclear; for instance, if an object does not fit into either of two
discrete categories, simply placing it nearer one of them, or between the two shows its fuzzy
nature, and intuitively means ‘has something to do with . . .,’ without further commitment.

This case study, although not dealing with hypertext systems as such , illustrates two
conceptual issues relevant to the discussion about spatial cognition and hypertext. These can
be summarised as follows:

(i) virtual and euclidean spaces have different properties;

(ii) it is important for a node’s location to carry information about its content, or spatial
organisation is meaningless.

Taking the first point, the evolution of Rooms helps to clarify the difference between virtual
and euclidean spaces. Space in the physical world is ‘euclidean’ being defined by three
dimensions. Location within that space is absolute, and as Figure 3 shows, has attributes of
distance and direction. Euclidean space can also exist as an interface metaphor or model, if
entities possess these attributes. Virtual spaces are only possible online, being simply an
interface presentation of memory areas, without any attempt made to locate them within a
euclidean framework. Rooms employs a virtual-space model—a Room is a virtual storage
area with no meaningful euclidean relationship to the others.

9
The difference between Rooms and BigScreen—virtual vs. euclidean spatial organisation—is
analogous to normal hypertext (occupying a virtual space) vs. hypertext located within a
euclidean space. The spatial metaphor that one is moving through space when navigating
hypertext in fact most commonly refers to movement through virtual space— relative to the
whole structure, but not in relation to euclidean space— simply because normally it has not
been defined. When designers choose to develop this imagery further, they may locate the
structure within a two or three dimensional space, and in doing so make the space euclidean.
Similarly, if Rooms were made ‘more spatial’ by presentation within a uniform space, it
would effectively produce the BigScreen interface.

One can conclude from this example that care needs to be taken in the use of spatial imagery:
at the level of cluster organisation, use of a euclidean spatial interface proved not only
unnecessary, but intrusive. Virtual space was more appropriate as a medium. At the level of
within-cluster work (windows and icons in individual Rooms) local euclidean organisation
was ideal, offering a number of useful properties inherent in the physical world.

User difficulties in assigning position to clusters brings us to the second issue, closely tied to
the first, namely the relationship between space and semantics—what does location tell you?

Another characterisation of the BigScreen/Rooms example would be that an unstructured

space (in this case 2D) seems only to be useful for a small number of objects; as the number
increases, the user can no longer sustain a coherent strategy for allocating position in a
sensible way (nobody organises all of their papers spatially). Other than that clusters be
visually discriminable, a non-task related criterion, spatial layout becomes arbitrary. Jones
and Dumais [30] corroborate this with experimental data indicating that divorced from
meaning or any task context, spatial location is a poor retrieval cue.

Two possible conclusions could be drawn from this. One would be that if the content of
information objects (or nodes) does not lend itself to representation along a set of dimensions
(e.g. general-specific; old-young; novice-expert), spatial cues should not be used (see next
section for detailed treatment of sets of dimensions).

Alternatively, one may argue that spatial representation has a number of benefits to offer
hypertext users in terms of navigation and representation of structure: a physical structure
which reflects conceptual structure helps navigation; we have a well developed ability to
cognitively map distance and direction; we know how to use local space to encode complex
relationships. These cues should be made available, if possible, care being taken not to
impose constraints which interfere with the task. The design goal is to enrich the interface
spatially, without drawbacks of the sort found with BigScreen. The key to creating a
euclidean space which is coherent and therefore navigable, is through giving meaning to
position.

Navigating the space, locating existing information, and allocating meaningful positions to
new information became major cognitive overheads for 2D landscape users. Not surprisingly,
these issues continue to dominate discussion as we move into three dimensional space.

Thinking in three dimensions

Let us now consider the use of three dimensional spatial imagery. Perhaps the simplest model
is the unstructured ‘cube’ of infinite size, within which nodes may be positioned anywhere.
This is one of the models supported by SemNet [3l] an experimental browser for
manipulating and ‘soaring through’ the structure of large knowledge bases, and one of the
few systems using the graphics hardware commonly used for displaying chemical structures,
for representing virtual information structures. The key concerns of the project proved to be
translating from semantic to spatial relationships, and providing ways for users to orient and
navigate through large 3D node-link structures.

10
Various heuristics and scaling techniques are available to organise elements in meaningful
ways, defining importance, for instance, on the basis of the number of connections, but there
is a constant tension in trying to compute displays which reveal relationships of interest
through the supposedly ‘natural’ medium of 3D space. It is often difficult to represent
meaningful relationships within the structure (e.g. what relates closely to Node N?) as a
corresponding structure in space (i.e. what is displayed near Node N in space?). One method
attempted to reduce visual complexity by clustering all nodes in a given locality into one
object; however, this only produces meaningful structures if nearby elements are closely
related to start with, which in turn requires algorithms with particular characteristics.
Fairchild et al also discuss a variety of navigational techniques, but acknowledge that they
fail to solve the disorientation problems inherent in such a complex space.

Why is disorientation to be expected in terms of cognitive mapping theory? Perhaps the key
factor is that the distance and direction components required to cognitively map a space
(Figure 3) are lost as soon as the spatial environment becomes dynamic (i.e. new views are
computed). Just as the world would be a confusing place if buildings, roads, cities, and fields
constantly changed location, if nodes no longer have a position which can be relied upon the
concept of place changes entirely, absolute distance and direction are not available as
orientational cues, and consequently cognitive maps become redundant. Thus, in applications
domains where multiple views are to be computed, such as filtering node and link types, or
the use of display functions which operate on nodal distances, real world spatial metaphors
which imply a uniform, static structure will be of limited use in navigation because the
information structure changes. Static, structural metaphors might be ‘the hypertext is like a
multi-storey building’ or ‘like a city which you can explore’—one does not expect entities to
change position over time, or the macro structure to alter. The less structure within the
imagery, the more easily it can accommodate changes, e.g. ‘the hypertext is a 3D cube’ or ‘a
2D plane.’

Defining a space with dimensions

Another way location can be given semantics is by defining space through the use of labelled
dimensions. Here, cartesian coordinates are determined by a node’s value along each
dimension. The user who knows the characteristics of the desired information can then infer
from her current position in the hypertext both the distance and direction of unseen nodes.
This represents a spatial world possessing all the attributes of euclidean space, plus
meaning—information about a node is contained in its absolute position. In spaces with no
semantics, absolute position has no meaning; instead, it is the relationships between nodes
which carry the information (e.g. proximity). Navigation through a space defined by
dimensions is something akin to the difference between finding an address in a city with
numbered streets and avenues, compared to one with only names—with numbered streets, at
any point there are orientational cues available.

Let us consider the use of dimensions a little further. Complications arise if there are more
than three useful dimensions in a hypertext (i.e. nodes have more than three system-readable
attributes of interest), in which case one space is inadequate, and users should be able to
select the dimensions to be plotted. However, an immediate consequence of computing
several different structures is that the spatial world is no longer stable—nodes can no longer
be associated with a unique position. This creates something of a paradox, in that distance
and position are important, having local meaning within a given space and so facilitating the
development of cognitive maps, but users can never develop a uniform cognitive map of the
hypertext because there are many possible spaces, depending on the dimensions selected.

Another approach to using dimensions has been proposed by Haitto [32]. In pursuit of a
uniform framework for the different hypertext models we see proliferating in the literature, he
proposes the use of a 3D euclidean-spatial conceptual model, defined by highly abstract
dimensions intended to cover a wide variety of applications. Possible dimensions suggested

11
are: abstraction (concepts become more specific as one descends the space), backward-
forward (author defined ordering of nodes to reflect train of thought), and previous-next
(degree-of-relatedness of supplementary material not directly relevant to the forward flow of
ideas).

Whilst the problem of finding a hypertext standard is of first order importance, a difficulty
with this particular approach is that the abstractness of the dimensions makes it difficult to
classify material in both authoring, and subsequent spatial problem solving. For instance,
ideas can be classed as principle concepts, but be supplementary to others; generality depends
on the perspective taken; and there may be more than one chain of reasoning through the
material (the whole idea of hypertext being to support multiple pathways through such
loosely structured domains).

This tension between computing multiple views and supporting the development of cognitive
maps creates representational difficulties which may never be completely resolved
satisfactorily.

Having examined a number of ways in which to set up associations between location and
domain semantics, attention now turns to assess a hypertext browser which has been
developed as part of ongoing work.

The layered-space model

This example illustrates how the implications of the preceding analysis might be taken into
account in designing a hypertext environment for certain tasks. No conceptual model is, of
course, suitable for all applications domains, and issues of scope are discussed after
describing its main features.

The emphasis was placed on constraining use of the depth dimension to avoid aggravating the
disorientation problems already common in two dimensions. This led to an interface model of
a 3D space structured into a series of layers (Figure 7). Using multiple 2D layers is similar in
certain respects to the ‘transparent overlays’ metaphor used in CAD and graphic design
software to separate levels of detail. However, the model described here develops simple 2D
layers, by explicitly displaying and using the spaces between them.

Figure 7. The basic layered-space model, in which 3D euclidean space is structured into
layers and subspaces between layers. Each layer is dedicated to a different class of
information; new layers can be defined as necessary

Each layer (and associated space) is devoted to a particular subtopic of the complete
information space. A further constraint, therefore, is that clusters of nodes must be rooted on
one of these discrete levels along the Z-axis.

Figure 8 shows how layers can be subdivided, in this instance into quadrants, but this clearly
depends on the context. The root node of a subnetwork is located within a quadrant, with an
associated subspace containing the subnetwork.

12
 y

Figure 8. Layers can themselves be structured, to create hierarchical organisation, for

instance, division into quadrants and associated subspaces

This design has been simulated as a NoteCards browser [33] over a guide to tennis
equipment. Figure 9 shows the top level browser, and the use of subspace nodes, which
‘contain’ a subnetwork rather like composite IPA nodes in the gIBIS system [34].

Four layers are shown, each with one or more appropriately named subspace nodes. Clicking
on a subspace node displays its subnetwork of notecards in a new window. Figure 10
describes the information available from a subspace display, with particular effort made to
orient users viewing different parts of the space.

Figure 11 shows use of all three levels of the information structure for Accessories , that is,
the top level browser, the subspace, and the contents of the nodes themselves; Figure 12
illustrates part of the Rackets layer.

13
Figure 9. The structure of a hypertext represented as four semantically organised layers.
Each layer has one or more subnetworks, each represented by a subspace-node with
an appropriate title. The contents of the subspaces remain hidden until the user
selects a subspace node.

14
Figure 10. Two subspace windows displaying the structure of information on support and
cushioning ; these appear when their respective composite nodes at the top level
browsers are selected. If a subnetwork has links to other parts of the hypertext, the
location of the remote nodes is indicated by a schematic icon (lower right)
representing the quadrants of the layer on which they are situated. This conveys
locational and contextual information to the user, showing how the small portion
currently being viewed relates to the rest of the hypertext. The window showing
cushioning indicates the node cross-trainer support situated within the top left

15
 quadrant of the same layer: by checking the top level browser (shown behind), the
user can see that this quadrant is devoted to information on support. When the
support subspace is inspected (left-hand window), the link from cross-trainer
support is confirmed. Information across multiple views can thus be integrated
through the use of shared nodes.

Figure 11. The structure of Accessories information. All three levels of display are
represented here: layered representations, subnetworks, and actual nodes.

16
 Figure 12. Using the browser to access information on Rackets .

As with other spatial imagery, the layered-space design can be characterised according to
how many dimensions it uses, the sort of space invoked, and the relationship, if any, between
location and semantics. The browser display is 2. 5D, that is, represents 3D space in flat 2D.
True 3D displays allow manipulation of the structure in three dimensions, as with SemNet
(c.f. Poltrock et al [7] who consider the display of any depth, e.g. overlapping windows, to be
3D). The space represented is 3D euclidean space: layers have spatial relationships with each

17
other, as do notecards within subspaces. In the examples, the global space acquires meaning
by setting up an explicit relationship between layers and content. However, one can imagine
labelling one or more of the dimensions to further enrich the space, for instance in a tutorial,
the further to the right layers are, the more advanced the material becomes. Other strategies
might be to locate general material in the top half of each layer, and detailed information in
the lower half, or to define tiers, so that information in the top, middle and bottom thirds of
each layer differed in the perspective offered, for instance, information for managers,
electronic engineers, and usability engineers for each phase of a project.

A number of hypotheses about the layered-space browser are proposed:

- Depth cues help at the perceptual level, by displaying link arcs more clearly: The simple
depth cues enable visualisation of a space rather than a plane, and as a result intersecting
links are now delineated, appearing to lie along different axes (XY and YZ). This effect is
often the motivation behind other systems employing 3D structural displays [35, 36].

- The combination of composite nodes and layers makes it easier to internalise structure: The
top level browser presents a clean, high level overview of the primary classes of
information present, and the subgroupings within (subspace nodes on the layers). Fine
structure is hidden until requested.

- Providing strong spatial cues at the browser level aids orientation when navigating at node
level—The critical factor for a static 2.5D browser is the pervasiveness of the spatial
imagery evoked, that is, the extent to which the layered space model exerts an influence as
users navigate at node level— how strongly will they feel that they moving ‘up’ a
quadrant, or ‘left’ to another layer as they click on a word to traverse a link, or that a
particular node is ‘in front’ or ‘above’ them as they read a notecard? As has already been
discussed, the tighter the mapping between location and semantics (the ideal example
being the space of labelled, orthogonal dimensions), the more easily users can work out
their relative position to other nodes, and so the imagery evoked is more powerful.

Scope of the layered-space model

The layers browser employs a structural model chosen initially for the express purpose of
tackling problems in locating nodes, preserving orientation, and conveying structure, and
would seem best suited to presenting well structured knowledge domains with well defined
classes of information. The layers model is clearly well suited to hierarchical decomposition,
a feature of well structured domains, and one anticipates problems in fitting ill-structured
networks into this structured framework.

This would appear to limit application to domains requiring a single view of a static hypertext
(of which there are many). However, whilst recomputing new layered structures would be
complex, one can foresee users defining new layers and subspaces, rather like plotting a new
set of dimensions; nodes would be placed on the layer with which they had the strongest
relationship (definable in terms of number or type of links), and links to other layers indicated
as shown in Figure 10. (To date, implementational work with the layers browser has not
progressed further than the visual simulation shown in the examples, which offers standard
NoteCards linking and navigational functionality).

Virtual realities

Before concluding, a brief note regarding virtual realities is needed, if only to acknowledge
their entrance into the arena of information-space technologies. Also known as artificial
worlds or realities, these systems offer compelling 3D graphics to take the user or users into a

18
space with objects from the real world. However, users may also be able to fly, or change
their form, and the physical laws modified, whilst the addition of datagloves, full bodysuits
and head-mounted displays brings ‘total human-computer interaction’ a step closer. Most
work to date has been experimental and purely aesthetic, but clearly there is potential for
useful application.

In relation to information structuring and navigation, there has been discussion [37] about the
potential of ‘virtual offices,’ with 3D desks, chairs, bookshelves, papers, and so forth. In light
of the preceding analysis, it is far from clear whether this cloning of the real world, especially
the visuospatial aspects, is a productive direction to take in terms of information
management. Underpinning it seems to be an assumption that the more we can get our hands
on the information and physically manipulate it—in some cases literally—the better. There is
certainly gain to be had in preserving the best aspects of our physical world in electronic form
(euclidean properties of desktops being one example); however, computers also free us from
the constraints which the physical world places on our interaction with information, and care
needs to be taken that we do not build those into our information worlds in our zeal to make
them familiar.

CONCLUSIONS
In an early article, Miller [38] makes an eloquent plea for the richness of spatial cues to
become part of information system interfaces. He writes:

The fact that space is unimportant to modern information processing systems should
mean that it could be placed wholly at the disposition of the user, not that all spatial
features . . . be systematically homogenised. (p.288)

In the systems and spatial images examined in this paper, we see that technology is capable
of realising Miller’s hopes of 1968, but also begin to appreciate the conceptual hurdles which
are erected when spatial imagery is applied to complex information structures.

It is maintained that the conceptual parallels in mapping real and hypertext spatial
environments are potentially extremely useful, and worth pursuing through the use of spatial,
in particular euclidean spatial models and metaphors. However, an understanding of the
user’s tasks must be the moderating variable which determines whether spatial coding will
add or detract from the interface.

It is hoped that this discussion will generate further research into the relationship between
spatial cognition, and the representation and navigation of information structures.

ACKNOWLEDGEMENTS

Many thanks to Nick Hammond, Allan MacLean, Lynda Hardman, Terry Mayes, and
Jonathan Maissel for discussion and comments on earlier drafts. I am also grateful to Rank
Xerox EuroPARC for providing the facilities and environment in which to develop these
ideas.

This work was funded by Rank Xerox EuroPARC, and Science and Engineering Research
Council, CASE Award No. 88504176.

An early version of this paper appears in Open University Computing Technical Report 89/7.

19
REFERENCES
1. BOLT, R. Spatial data management. Cambridge, Massachussetts: Architecture Machine
Group, MIT, 1979.
2. CANTER, D., RIVERS, R. and STORRS, G. Characterising user navigation through
complex data structures. Behaviour and Information Technology, 4 (2), 1985, 93-102.
3. ELM, W.C. and WOODS, D.D. Getting lost: a case study in interface design. Proceedings
of the Human Factors Society 29th Annual Meeting, 1985, 927-931
4. VIGIL, P.J. Spatial ability, imagery, and human-computer search dynamics. Online
Information 88: Proceedings of 12th International Online Information Meeting, London
6-8 December, Oxford and New Jersey, 1988, 51-55
5. WOODS, D.D. Visual momentum: a concept to improve the cognitive coupling of person
and computer. International Journal of Man-Machine Studies, 21, 1984, 229-244.
6. AKSCYN, R., McCRACKEN, D. and YODER, E. KMS: A distributed hypermedia system
for managing knowledge in organisations. Communicationsof the ACM, 31(7), 1988, 820-
835.
7. POLTROCK, S.E., SHOOK R.E., FAIRCHILD K., TARLTON N., LOVGREN J.,
TARLTON M. and HAUSER M. Three-dimensional interfaces: The promise and the
problems. MCC Technical Report No. HI-291-86, 1986.
8. DURHAM, T. Think what you could do with all this space. Computing, 13April, 1989,36-
37.
9. JONES, S. Graphical interfaces for knowledge engineering: An overview of relevant
literature. The Knowledge Engineering Review, 3 (3), 1988, 221-247.
10. HAMMOND, N.V. and ALLINSON, L.J. Taking the hype out of hypertext: Why it’s
inadequate for learning. Proceedings of 4th European Conference on Cognitive
Ergonomics, Cambridge, 1988.
11. KAHN, P. and MYEROWITZ, N. Guide, HyperCard, and Intermedia: A comparison of
hypertext/hypermedia systems. IRIS Technical Report 88- 7, 1988.
12. NElSSER, U. Cognitive maps. In: Neisser, U. Cognition and Reality. San Francisco:
Freeman and Co., 1976.
13. DOWNS, R.M. and STEA, D. Cognitive maps and spatial behaviour: Process and
product. In: R.M. Downs and D. Stea, eds. Image and Environment. London: Edward
Arnold, 1973.
14. WRIGHT, P. Cognitive ergonomics cf hypertexts. HyperHyper: Developments across the
field of Hypermedia. British Computer Society BCI-SIG Workshop, London, 1989.
15. THORNDYKE, P.W. Performance models for spatial and locational cognition. Technical
Report R-2676-0NR. Washington DC: Rand Corporation, 1980.
16. SIEGEL, A.W. and WHITE, S.H. The development of spatial representations of large
scale environments. In H.W. Reese (ed) Advances in Child Development and Behaviour
(vol. 10), New York: Academic Press, 1975.
17. HALASZ, F. and MORAN, T.P. Mental models and problem solving in using a pocket
calculator. In: Proceedings CHI’83: Human Factors in Computing Systems, ACM, 1983.
18. HARDMAN, L. Evaluating the usability of the Glasgow Online hypertext. Hypermedia,
1(1), 1989, 34-63.
19. WADLOW, M.G. The ANDREW system: The role of human interface guidelines in the
design of multimedia applications. Current Psychology: Research and Reviews, Summer,
1990.
20. MONK, A. The Personal Browser: a tool for directed navigation in hypertext systems.
Interacting with Computers, 1(2), 1989, 190-196.
21. LANSDALE, M.W., YOUNG, D.R. and BASS, C.A. MEMOIRS: A personal multimedia
informaion system. In A. Sutcliffe and L. Macaulay (eds), Proceedings of HCI’89, People
and Computers V, Cambridge: Cambridge University Press, 1989, 315-327.

20
 22. YANKELOVICH, N., SMITH, K.E., GARRET, N. and MYEROWITZ, N. Issues in
designing a hypermedia document system: The Intermedia case study. In: K. Hooper and
S. Ambron, eds. Interactive multimedia: visions of multimedia for developers, educators,
and information providers. Microsoft Press, 1988, 33-85.
23. LYNCH, K. The Image of the city. Cambridge, Massachussetts: MIT Press, 1960.
24. CANTER, D. The psychology of place. London: The Architectural Press, 1977.
25. EDWARDS, D.M. and HARDMAN, L. ‘Lost in hyperspace’: Cognitive mapping and
navigation in a hypertext environment. In: R. McAleese, ed. Hypertext: Theory into
Practice. Oxford: Intellect Ltd, 1989.
26. CARD, S. Information workspaces. International Symposium on Next Generation Human
Interface Technologies, Sept. S-6, Tokyo, 1989.
27. HENDERSON, D.A. and CARD, S.K. Rooms: The use of multiple virtual workspaces to
reduce space contention in a window-based graphical user interface. ACM Transactions
on Graphics, 5(3), 1986, 211-243.
28. HEROT, C.F., CARLING, R., FRIEDELL, M. and KRAMLICH, D. A prototype spatial
data management system. ACM Conference on Computer Graphics, 1980, 63-70.
29. MALONE, T.W. How do people organise their desks? Implications for the design of
office information systems. ACM Transactions on Office Information Systems, 1, 1985,
99-112.
30. JONES, W.P. and DUMAIS, S.T. The spatial metaphor for user interfaces: Experimental
tests of reference by location versus name. ACM Transactions of Office Information
Systems, 4(1), 1986, 42-63.
31. FAIRCHILD, K.M., POLTROCK, S.E. and FURNAS, G.W. SemNet: Three-dimensional
graphic representations of large knowledge bases. In: R. Guindon, ed. Cognitive Science
and its Applications for Human-Computer Interaction. Hillsdale, N.J: LEA, 1988.
32. HAITTO, H. Towards portable hypertext: A plan of attack. Technical Report TRITA-NA-
P9008, IPLab-23. Stockho1m: Dept. Numerical Analysis and Computing Science, Royal
Institute of Technology, Sweden, 1990.
33. HALASZ, F.G., MORAN, T.P. and TRIGG, R.H. NoteCards in a Nutshell. Proceedings
CHI and GI’87, Human Factors in Computing Systems and Graphic Interface, ACM,
1987, 45-52.
34. CONKLIN, J. and BEGEMAN, M.L. glBlS: A tool for all reasons. Journal of the
American Society for Information Science, March, 1988.
35. KOMMERS, P.A.M. TextVision: Elicitation and acquisition of conceptual knowledge by
graph representation and multi-windowing. In: G. van der Veer and G. Mulder,eds.
Human Computer Interaction: Psychonomic Aspects. Berlin: Springer Verlag, 1988.
36. DODSON, D. Interaction with knowledge systerns through connection diagrams: Please
adjust your diagrams. In: Proceedings of Expert Systems 88. Cambridge: Cambridge
University Press, 1988.
37. MARCUS, A. Beyond the desktop metaphor. Special Interest Group Panel Session,
Human Factors in Computing Systems, CHI’90, Seattle, 1990.
38. MILLER, G.A. Psychology and information. American Documentation, July, 1968, 286-
289.

21

View publication stats