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

We have constructed a deductive question answeri ng system which accepts natura l language input in Japanese. The semantic trees of aseert ional input sentences are stored in a semantic network and i n t e r re la t i onsh ips —cond i t i ona l , imp l i ca t i ona l , and so f o r t h — are establ ished among them. A matching rou t ine looks for the semantic trees which have some r e l a t i ons to a query, and returns the mismatch in format ion (d i f fe rence) to a deduction rou t i ne . The deduction rout ine produces sub-goals to diminish t h i s d i f fe rence . This process takes place recurs ive ly u n t i l the d i f f e r ence is completely resolved (success), or there is no other p o s s i b i l i t y of matching in the semantic network ( f a i l u r e ) . Standard problem so lv ing techniques are used in t h i s process. As the resu l t the system is very powerful in handl ing deductive responses. In t h i s paper only the part of the l o g i c a l deduction is explained i n d e t a i l . DESCRIPTIVE TERNS: quest ion answering, deduction, natura l language, semantic network, problem so l v i ng .