Quantitative Common Sense Estimation System and its Application for Membership Function Generation

抄録

Systems capable of autonomous thinking are sometimes required to cope with unanticipated situations. An important issue in this context is knowledge – especially common sense – acquisition. In this paper, we propose novel quantitative common sense estimation methods and apply them to an automatic membership function generation system. Our proposed system estimates threshold values corresponding to large and small for various kinds of object-attribute sets to form membership functions, where it attempts to relate each object to its corresponding impression. Two methods are proposed in this paper. The first, Method-1, obtains data from the top 1,000 snippets through a web search and estimates the global and local tendencies by clustering them. The second, Method-2, uses the number of hits from a web search together with parts of the results obtained through Method-1. In addition, we devise several techniques to eliminate unnecessary information in the retrieved web pages. We also carried out experiments that verified the effectiveness of our proposed methods and the method combining those two.