Parsimonious covering as a method for natural language interfaces to expert systems

摘要

Abductive inference has been characterized in the AI literature as ‘inference to the best explanation’ or as ‘plausible inference involving context-sensitive discrimination among explanatory hypotheses’. Analogously, understanding natural language involves context-sensitive discrimination among word senses, and there has been a growing awareness that it can be viewed as a type of abductive inference. Parsimonious covering theory, first formulated to model the abductive inference underlying medical diagnostic problem solving, is examined here as a method for automating natural language processing for medical expert system interfaces. The nature of ‘parsimony’ in natural language processing and the relationship of parsimonious covering to a notion of focus of attention are discussed.An experimental prototype developed to test these ideas in the context of a medical expert system is briefly described. This prototype is domain-independent in the same sense that a generic expert system shell is domain-independent. Given a knowledge base for a specific medical application, a vocabulary extractor extracts and indexes the linguistic information which it contains. In addition, an indexed domain-independent knowledge base that contains linguistic knowledge common to many domains is used. With a parsimonious covering inference mechanism superimposed on this knowledge, a natural language interface is generated for the specific application defined by the knowledge base.