Person retrieval on XML documents by coreference analysis utilizing structural features

Keyword retrieval of the present day exploits frequencies and positions of search keywords in target documents. As for retrieval by two or more keywords, semantic relation between keywords is important. For retrieving information about a person, it is common to search by a pair of keywords consisting of person's name and his/her attribute of the interest. By using dependency analysis and coreference analysis, correct occurrences of pairs of person and his/her attributes can be retrieved. However, existing natural language analysis does not consider the factor that logical structures of the documents strongly influence probabilistic patterns of coreference. In this paper, we propose a new way of person retrieval by computing a maximum entropy model from linguistic features and structural features, where structural features are learned from probabilistic distribution of coreference over XML document structures. Our method can utilize strong correlation between XML document structures and coreference, thus having superior accuracy than existing methods.