Mechanism of Deduction in a Question-Answering System with Natural Language Input

Abstract

We have constructed a deductive question answering system which accepts natural language input in Japanese. The semantic trees of assertional input sentences are stored in a semantic network and interrelationships --conditional, implicational, and so forth-- are established among them. A matching routine looks for the semantic trees which have some relations to a query, and returns the mismatch information (difference ) to a deduction routine. The deduction routine pro.duces subgoals to diminish this difference. This process takes place recursively until the difference is completely resolved (success), or there is no other possibility of matching in the semantic network (failure). Standard problem solving techniques are used in this process. As the result the system is very powerful in handling deductive responses. In this paper only the part of the logical deduction is explained in detail. DESCRIPTIVE TERMS: question answering, deduction, natural language, semantic network, problem solving.