Construction Of Japanese Nominal Semantic Dictionary Using "A NOT B" Phrases In Corpora

This paper describes a method of constructing Japanese nominal semantic dictionary, which is indispensable for text analysis, especially for indirect anaphora resolution. The main idea is to use noun phrases of "A NO(postposition) B" in corpora. Two nouns A and B in "A NO B" can have several semantic relations. By collecting "A NO B" phrases form corpora, analyzing their semantic relations, and arranging them for each "B" and each semantic relation, we can obtain a nominal semantic dictionary. The dictionary we constructed from 130M characters corpora by this method has 22,252 entries, which can be considered as a practically useful coverage. Our method for analyzing "A NO B" phrase is also original which uses a thesaurus as an attribute for decision tree. 1 I n t r o d u c t i o n The role of dictionary is undoubtedly important in Natural Language Processing (NLP). So far, research in NLP has mainly concerned the analysis of individual sentences. The analysis of a sentence is to clarify which element in it has relation with which by what relation. To do such an analysis, a verbal semantic dictionary, in other words, case frame dictionary is necessary. A case frame dictionary describes what kind of cases each verb has and what kinds of noun can fill a case slot. Condition on case slots can be expressed by semantic markers and/or example nouns. For example, a case frame for the verb "YOMU(read)" can be as follows: "Now at Communications Research Laboratory. Email: [email protected] tNow at Sharp Corporation. Y O M U ( r e a d ) agent : human beings, like KARE(he), KEN(ken), SENSEI(teacher) object : something to be read, like HON(book), SHOSETSU(novel) Such dictionaries with a practically useful coverage have been compiled in many institutes, mainly by hand, and used in many NLP systems (EDR, 1993; NTT, 1997). These days, the main target of NLP has been shifting from individual sentences to a series of sentences, that is, a text. Human beings use language to communicate, and the unit of communication is not a sentence, but a text in most cases, especially in the case of written language. The NLP system can only catch enough information when it handles a text as a whole. Similar to sentence analysis, the main part of text analysis is to clarify the relation among its