ABSTRACT
In conjunction with a project for information retrieval from a medical textual data base, we have encoded a fairly large knowledge base of facts (about 1000 axioms) ranging from low-level commonsense knowledge to mildly arcane medical knowledge. A methodology is described for constructing such a knowledge base in a principled way. Facts are selected for the knowledge base by determining what facts are linguistically presupposed by a text in the domain of interest. The facts are then organized into clusters according to their logical dependencies and encoded as predicate calculus axioms. We are currently devising pragmatic processes that use such a knowledge base to interpret medical discourse. Among these is a process that seeks to bring a predicate and its arguments into congruence. At the simplest level, this merely checks selectional constraints; in more complex cases, it seeks to resolve examples of metonymy. It is argued that sublanguage constraints that are usually encoded as selectional constraints on predicate-argument pairs are in fact a surface manifestation of these deeper pragmatic processes operating on such a domain-specific knowledge base.
