Perception & Performance Action:
Human Machine Interaction

Virtual Information Spaces

Some background

The ongoing spread of information technology means that increasing amounts of data are being stored on computer based media, computers are being used by a more heterogenous population, work is becoming more 'knowledge oriented' (see Card, Robertson & York, 1996), and connectivity between systems is expanding rapidly. These factors present challenges with respect to the successful representation and manipulation of data. Systems must be developed that facilitate information access, and individual differences in performance must be minimised.

For example, difficulties locating information in a textual database can be seen with respect to the Internet. The development of the Internet as an important information resource brings problems of data navigation on a scale not previously seen. Although a number of Internet ‘search engines’ are available, these tend to rely on content specific (similarity) searching, and require the user to declare one or more specific terms (usually words) as the basis for the search. Unformatunately, the user often has only a vague idea of the area of interest (the very purpose of the search may be to find out more about the topic) so it is difficult to define the search accurately. These problems are compounded by the effects of synonymy (different words can refer to the same concept) and polysemy (one word can refer to several different concepts), and often result in the user being deluged with information the relevance of which is difficult to assess, and yet failing to retrieve all relevant information. This situation can be improved by providing the user with additional contextual information about the contents of a database (see e.g. Fischer & Stevens, 1991; Furnas, 1995; Mukherjea & Foley, 1995).

Figure 1.	Figure 2.

One method of achieving this is through the development of Virtual Information Spaces (VISs), in which items (in this context items could be, e.g. documents or web pages) are represented as objects in a virtual environment (see Figures 1 & 2), whose location is determined by their semantic content. Items that are semantically similar are placed close together, and items that are semantically dissimilar are placed far apart. Information access and retrieval becomes akin to the process of 'real world' navigation (Dillon et al., 1990), with successful performance being dependent on the user possessing an effective mental representation or 'cognitive map' (Tolman, 1948) of the area (i.e. information space) to be navigated. The basis premises on which the user operates is that if they have found an item of interest then other similar items will be located nearby in the VIS, but if a selected item is definitely not of interest then the same probably applies for other nearby items and they might wish to search a different area of the space.

Cognitive processes

The use of spatial reference as an aid to human memory dates back to classical times (Method of Loci). However, the application of spatial metaphors to computerised information retrieval tasks only became recognised as viable in the early 1980s (e.g. Cole, 1982; Herot, 1980). At this time a good deal of research was conducted relating to computer-based data navigation and the format of data structures. However, this was mainly concerned with optimising hierarchical menu structures (see Paap & Roske-Hofstrand, 1988, for review). In the mid to late 1980s relevant research was conducted relating to the use of Hypertext systems (Conklin, 1987; Nielsen, 1989). These systems allow the user to navigate information spaces using complex linking structures based on the semantic properties of the database contents. Problems of ‘getting lost’ in Hypertext were apparent and various graphical browsers were developed in order to convey additional information to the user (Dillon et al., 1990). Initially, these browsers presented restricted views of the dataspace in two dimensions (Furnas, 1986; Valdez et al., 1988). Recent advances in computer processing power and the affordability of such technology now mean that now data can be represented along up to three spatial dimensions.

Although some information retrieval systems comprise stored items that have ‘real world’ spatial relationships (e.g. Geographic Information Systems: GIS), one of the strengths of SDMSs lies in the fact that diverse, non-spatial information can be represented within a spatial context. Some of the mechanics that underlie this mapping have been proposed by Jackendoff (1983; see also Gardenfors, 2000) who argues that the highly developed capacity of the human brain for spatial processing is responsible for the application (during the developmental process) of similar structures to the cognitive organisation of information from other semantic fields. Consequently, the semantic primitives that describe spatial associations (motion and location) are held to form a superset from which associations in any other semantic field can be described. The implications of this position for the development and use of VISs are profound. It would follow that any given semantic dimension of computer-stored information can be represented in a spatial format, and that any computerised information space can be navigated using similar cognitive processes to those that would apply during the process of ‘real world’ navigation. These concepts have been augmented within the context of Spatial Information Systems. For example, Laurini and Thompson (1992) identify the non-metric, topological properties of connectivity, orientation, adjacency, and containment; and Benedikt (1991) distinguishes between extrinsic (x, y, and z location) and intrinsic (e.g. size, shape, colour, spin speed) properties of virtual objects.

Mapping to two or three dimensions?

There are a number of VISs in existence or under development. These vary with respect to the number of spatial dimensions that are used to represent semantic information. Some systems represent semantic information along only two spatial dimensions - usually the x and z dimensions (e.g. Chen, 1999a; Foltz, 1998). Other systems use all three spatial dimensions (e.g. Fox, Frieder, Knepper, & Snowberg, 1999; Mariani & Lougher, 1992; Risch et al., 1997).

Although additional semantic information can be conveyed by 'mapping' to three dimensions, rather than two, there are associated disadvantages. It has been demonstrated (Westerman, 1998) that navigation of three-dimensional spaces is more cognitively demanding than that of two-dimensional spaces. It would follow that, when applying semantic structure, the benefit associated with the spatial representation of additional semantic dimensions must outweigh the additional cognitive demands. In other words, the amount of additional semantic information conveyed by mapping information to an additional spatial dimension must be sufficient to overcome the associated increase in navigational demands. This issue has been the subject of investigation in a recent EPSRC funded research project.

How can semantic information be mapped to spatial dimensions?

One way to determine the spatial location of items in a VIS is to have participants make a series of judgements relating to the conceptual distance between items, and to apply multidimensional scaling (MDS) to achieve a semantic-spatial mapping. Pallant et al. (1996) suggest that this technique can be used as a means of achieving an acceptable group fit to data. However, it is not well suited to the purposes considered here. It is difficult to apply to large datasets, and not suited to ‘organic’ systems, e.g. the Internet. An alternative approach that can be applied to databases containing textual information is to use Automatic Text Analysis (ATA). There are many forms of ATA. One such method uses the Vector Space Model (VSM) to determine semantic relatedness of items based on the frequency of words or letter strings contained within the database (see Salton, 1983). Initially, a document by term matrix is deveoped in which the frequency with which each unique term (usually words) occurs in each document is recorded (see Table 1). Document similarity is calculated on the basis of the distance or angle between document vectors (see Cribbin & Westerman, 1999, for more detail). From this, the position of databased items (documents) in a VIS can be detemined, as with human similarity ratings, using MDS.

	Document1	Document2	Document3	. . . . . .	Documentn
Term1	1	2	0		3
Term2	1	0	0		1
Term3	2	1	1		0
. . . . .
Termt	2	2	3		2

Table 1. Example of a word by document matrix

Although ATA allows text items to be assessed for semantic similarity without human intervention, it is not clear that the conceptual structures applied by ATA algorithms are strongly associated with the conceptual structures that human users of a system would apply to the represented information. This may be an extremely important factor in determining the utility of such systems, and is an issue that we have been examining (Cribbin & Westerman, 1999; Westerman, Cribbin, & Collins, submitted).

References

Benedikt, M. (1991). Cyberspace: Some proposals. In M. Benedikt (Ed.), Cyberspace: First Steps. MIT Press. Pp. 273-302.

Card, S.K., Robertson, G.C., & York, W. (1996). The WebBook and the Web Forager: An information workspace for the World-Wide Web. In Proceedings of CHI ‘96. NY: ACM. Pp. 111-117.

Cole, I. (1982). Human aspects of office filing: Implications for the electronic office. Proceedings of the Human Factors Society 26th Annual Meeting. Santa Monica, CA: HFS.

Conklin, J. (1987). Hypertext: An introduction and survey. IEEE Computer, 20, 17-41.

Cribbin, T. & Westerman, S.J. (1999). Spatial data management systems: Mapping semantic distance. INTERACT '99. Seventh IFIP Conference on Human-Computer Interaction. Amsterdam: IOS Press. Pp. 171-178.

Dillon, A., McKnight, C., & Richardson, J. (1990). Navigation in hypertext: A critical review of the concept. In Proceedings of INTERACT '90. Amsterdam: North Holland.

Fischer, G. & Stevens, C. (1991). Information access in complex, poorly structured information space. In Proceedings of CHI ‘91. Pp. 63-70.

Furnas, G.W. (1986). Generalized fisheye views. In Proceeding of CHI ‘86. NY: ACM.

Furnas, G.W. (1995). Navigation-by-query, query-by-navigation, and other tricks to try. In Panel (Mackinlay et al.) Browsing vs. Search: Can we find a synergy. Proceedings of CHI ‘95. P. 180.

Herot, C.F. (1980). Spatial management of data. ACM Transactions on Database Systems, 5, 493-514.

Jackendoff, R. (1983). Semantics and Cognition. London: MIT Press.

Laurini, R. & Thompson, D. (1992). Fundamentals of Spatial Information Systems. London: Academic Press.

Mariani, J.A. & Lougher, R. (1992) Triplespace: An experiment in a 3D graphical interface to a binary relational database. Interacting with Computers, 4, 147-162.

Mukherjea, S. & Foley, J.D. (1995). Showing the context of nodes in the World-Wide Web. In Proceedings of CHI ‘95. pp. 326-327.

Nielsen, J. (1989). The matters that really matter for hypertext usability. In Proceedings of Hypertext ‘89. NY: ACM.

Paap, K.R. & Roske-Hofstrand, R.J. (1988). Design of menus. In M. Hellander (Ed.) Handbook of Human-Computer Interaction. Amsterdam: North Holland.

Pallant, A., Timmer, P. , & McRae, S. (1996). Cognitive mapping as a tool for requirements capture. In S.A. Roberson (Ed.), Contemporary Ergonomics 1996. In Proceedings of the Annual Conference of the Ergonomics Society. London: Taylor & Francis. Pp. 495-500.

Salton, G. & McGill, M.J. (1983). Introduction to Modern Information Retrieval. London, McGraw-Hill.

Tolman, E.C. (1948). Cognitive maps in rats and men. Psychological Review, 55, 189-208.

Valdez, F., Chignell, M., & Glenn, B. (1988). Browsing models for hypermedia databases. In Proceedings of the Human Factors Society 32nd Annual Meeting. Santa Monica, CA: HFS.

Westerman, S.J. (1998). A comparison of the cognitive demands of navigating two- versus three-dimensional spatial database layouts. Ergonomics, 41, 207-216.

Westerman, S.J., Cribbin, T., & Collins, J. (submitted). N-gram Automatic Text Analysis Strongly Predicts Average Human Ratings of Document Semantic Similarity.