10.2 SEMNET - A Testbed for Three-Dimensional Knowledge Base Representation

SemNet is an experimental environment for exploring a variety of navigation mechanisms for structured knowledge bases [Fai88]. Its goal is to present large knowledge bases to the users in ways that can easily be understood. The knowledge bases are represented as three-dimensional directed graphs. Fig. I.41 shows an optimized representation of a knowledge base consisting of Prolog rules.

Figure I.41 SemNet representation of a Prolog knowledge base
(ExFig. 3-2 of [Fai88])

The three-dimensional graphs on the screen can be manipulated with a mouse in real-time. Fairchild et al. used a two-layered architecture with the graphical representation part running on an Silicon Graphics IRIS workstation and the knowledge base running on another computer. To position the elements of the knowledge base, i.e., the nodes of the graph, as understandably as possible, Fairchild et al. experimented with a variety of different methods:

Random positioning:

this simple method resulted, not surprisingly, in the cognitively worst representation. Although this method may work for small knowledge bases, it results in a meaningless tangle for larger bases.

Mapping functions:

Frequently, the elements in the knowledge base have properties that allow one to map the elements into the three-dimensional space. In a knowledge base about animals, the animals could, for example, be mapped using the three dimensions size, predacity, and domesticity.

Connectivity between elements:

In a knowledge base, there are many relationships between elements. Applying some measure for the proximity or similarity of elements results in effective representations of the whole knowledge base. Among the methods used by Fairchild et al. were multidimensional scaling,, where the distance between elements is monotonically related to the number of interconnections between elements; and clustering-based heuristics where the position of an element is based on the weighted mean (called centroid[8]) of the positions of all related knowledge elements. One of the most effective positioning methods was based on simulated annealing [Kir83], where a new position for an element is computed by adding a random vector to the position of the element. The element is moved to the new position if it is closer to the centroid, but not too close to other elements. This method resulted in graph layouts that by the users were subjectively judged to be the best. Figure I.38 shows a graph layout based on simulated annealing.

Personalized positioning:

Sometimes users wish to position nodes manually by dragging and grouping them with the mouse.

To reduce the cognitive overhead of navigating in large knowledge bases, not only do the nodes have to be placed and grouped meaningfully, but they also have to be filtered such that only the nodes of interest are actually displayed. The main technique employed for information reduction was "generalized fish eye views". As the main DOI[9]-function Fairchild et al. used the spatial location of the elements, i.e., the tree-distance among the elements. Objects more distant from the user's focus were represented by object-clusters. Objects-clusters and user specified elements as, e.g., nodes high in a ISA-subtree, got higher API (a priori importance). A second fish eye representation was achieved by showing important nodes larger and in greater detail by, e.g., showing not only the node icon, but also its subtitle and a brief textual description. Less important nodes were only marked by their icon.

To assist in orientation, users could place landmarks on the view to easily identify clusters of interesting nodes. But the most important navigation aid was the immediate animated three-dimensional navigation feedback. In particular, Fairchild et al. found that the users favored absolute movement over relative movement. This means, that users preferred to move to their desired viewpoint by using three two-dimensional maps showing the x-y, y-z, and x-z planes instead of using a "helicopter flight option" that allowed them to directly fly over the knowledge base.

In a newer application of the ideas first described in the SemNet project, Fairchild, Meredith and Wexelblatt extended their ideas to a tourist artificial reality [Fai89]. In this project, the authors tried to organize all information and documentation related to a large software project. They applied the ideas prototyped in SemNet, where the entities are represented in graphs and filtered using fish eye views. The novel idea in their new system is the tourist artificial reality: All workstations are linked to offer a shared workspace. There is also a speakerphone link between the workstations. This means, that two persons at two different workstations can walk through text or source code and discuss the same piece of information. In the tourist artificial reality one of the two participants is the tour guide, who gets full control over all aspects of the tour and the interface on the workstation of the tourist. The tourist can record the tour and replay it later to take the same trip at a more leisurely pace.

The techniques of Fairchild et al. demand intimate knowledge of the internal structure of the knowledge base. They either get this knowledge by using well structured information (like a collection of Prolog rules) or by manually preprocessing the information. In addition they extend their ideas to the integration of guided tours and guides. Contrary to the guides and agents described in chapter 12, Fairchild et al.'s guides are human beings that have the necessary domain knowledge.

The next system is based on the largest machine-readable database of structured knowledge that is currently available.