A library of generic concepts for composing knowledge bases

Citation: K. Barker, B. Porter, and P. Clark. A Library of Generic Concepts for Composing Knowledge Bases. First International Conference on Knowledge Capture, October 21-23, 2001.
Link: ACM Portal

Summary

Building a knowledge base traditionally involves a domain expert and a knowledge engineer. A goal of this research is to eliminate the knowledge engineer from this process, that is, to enable domain experts to build their own knowledge bases. A claim of authors' research is that users without knowledge engineering background will be able to represent knowledge from their domain of expertise by using generic components from a small library.

The authors have chosen to build a small library of components, and aim to use composition of these components as the means to achieve coverage, rather than through enumeration of a large number of components.

The research questions are: is such a system (1) easy to master for users not trained in knowledge engineering?, (2) and sufficient to represent sophisticated domain knowledge?

There are three requirements for library components:

• Coverage: The library should contain sufficient components in the library to allow the user to encode a variety of knowledge from any domain. This means defining a restricted set of components which are generic enough to allow a user to compose them consistently, but also be specific enough to allow use in individual domains.
  
• Access: The interface to the library should assist the user in finding appropriate components from the library.
  
• Semantics: The components in the library should have well defined axioms that encode their meanings as well as information about how the component can consistently interact with other components.
  

The library consists of entities and events (states and actions, where states are relatively static situations brought about or changed by actions).

Composition in this system is the ability to connect components in a way that allows inferences beyond the union of individual axioms of the components involved. By using composition, inferences like conditional rules, definitions (reclassification of instances) and simulations can be drawn. However, to achieve these inferences, composition language (relations and properties) must have predictable semantics. Relations connect entities and events (event-entity, entity-entity, event-event, entity-role). Properties link entities to values (cardinal, scalar, categorical etc.)

Evaluation of this library was done through user sessions and user feedback. The criteria for the feedback were (1) ease of finding relevant components, (2) understanding components, (3) use of components to represent knowledge, (4) ease of relations language, and (5) cast biological knowledge in terms of components and relations in the library. The library showed promising results.