Principles for Organizing Semantic Relations in Large Knowledge Bases

This paper defines principles for organizing semantic relations represented by slots in frame-structured knowledge bases. We consider not only the ways in which slots are used for reasoning about a given domain but also the features of the representation language of the knowledge-based system in which the slots reside. We find that the organization of slots may be based on the knowledge-level semantics of relations and the symbol-level function of slots that implement the representation language. However, the organization of slots is more understandable if these two fundamental distinctions are made explicit. The symbol-level organization of slots depends on the inferencing and expressive capabilities of the knowledge representation system. At the knowledge level, two entirely different classification schemes are identified: one based on linguistic similarities and differences, and another based on the types of concepts being related. As a case study, we show how to use the three organizational principles to revise the hierarchy of slot classes in an existing large-scale knowledge-based system. We show that each of these hierarchies is useful in retrieving existing slots and storing new ones. The results are applicable to knowledge reuse in general.