10.4 VISAR - Navigation by Inference

As has been noted earlier in the sections about SuperBook and dynamic linking, a reliable automatic link generation system needs some kind of semantic knowledge about the data it intends to link. The VISAR system described by Clitherow et al. [Cli89] employs a sophisticated approach using CYC. VISAR has been created to serve as an intelligent assistant to researchers performing an initial survey in a field of interest new to them. It takes a user request and returns a set of citations that seem to be related to the user's interest. VISAR differs in two aspects from existing document retrieval systems performing similar tasks:

• It allows the user to specify relationships between concepts as search input rather than keywords.
• Instead of returning a list of retrieved titles, VISAR returns a conceptual reduction of the relationship between the concepts found that are related to the query. Such a personalized information representation of the relationship between concepts is called a perinrep.

Perinreps are small extracts of a larger semantic network [Bra79]. In VISAR's case, the semantic network consists of the CYC frames and rules extended by a citations knowledge base. A perinrep is displayed on the screen as a graph structure of linked nodes, where each node represents a concept. Steps 2 and 4 in figure I.47 display generic perinreps.

The perinrep idea allows the user to get an overview of the retrieved information instead of being flooded with a list of all retrieved entries. VISAR works by transforming a textual user request into a request perinrep, based on a default perinrep. It then performs an inference process in the CYC knowledge base to find the matching conceptual relationships. Finally the matching relationships are reduced to manageable quantities and returned to the user (Fig. I.47).

Figure I.47 Functional representation of VISAR with Perinreps

A typical user of the VISAR system has the goal of acquiring overview knowledge in a new research field. To be useful in this task, VISAR has to be initialized first with a corpus of technical citations consisting of journal and conference articles. In the implementation described in [Cli89], the generation of the concept network is restricted to parsing the article titles. Obviously, parsing the full article text would be more accurate, unfortunately this is beyond the technical capabilities of today's systems. To parse the natural language request into a request perinrep, the natural language processing system LUCY [Wit86] is used.

VISAR contains some novel and unique features. Perinreps embody the idea of automatically computing the linking structure of a knowledge base. Furthermore, VISAR offers a flexible solution to the problem of extracting and displaying only information relevant to the answer of a query. It thus seems like an ideal tool for the exploration of large information systems. But for systems like VISAR to become also a practical success, machine readable knowledge bases like CYC, that ideally encompass the whole world knowledge, will have to be made available. Unfortunately, neither today's hardware nor software are capable of handling databases of that scale. But VISAR gives us at least a prototypical experience of how systems having such features will look like.

The next chapter introduces general tools and concepts for the visualization of large information structures.