te -based Figurative Descriptions for Japanese Word

In this paper, we propose a method for describing a Japanese word, not with explaining or defining sentences, but with figurative descriptions. Utilizing a simile pattern, our method gathers a large number of noun-noun relations from the World Wide Web. On the basis of those relations and their statistical information, associative pieces of knowledge called descriptors are estimated. The descriptors, which describe a query word figuratively, are sorted by ranking in order of descriptive ability level with generality and locality. Moreover, combining property of figurative relation and some fixed patterns, the descriptors are classified into concept words, attribute words, and the others. As output, a set of sorted descriptors is shown with several types of output forms. Some experiments using a prototype system 'Murasaki' have been conducted. The experimental results show that the fundamental performance of our method is significantly better than the bag-of-words approach. Additionally, the responsiveness for hot keywords on information retrieval web sites shows that the outcome of the evaluation had 60% precision, which exceeds that of a common dictionary. The method also functioned effectively in ranking performance (74% on MRR) and classification performance (63% accuracy). Furthermore, it is possible that the proposed method could be comparable to Wikipedia if steady coverage of the figurative descriptions for a query word could be ensured.