Towards Robust Unsupervised Personal Name Disambiguation

Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pp. 190–198, Prague, June 2007. c⃝2007 Association for Computational Linguistics Towards Robust Unsupervised Personal Name Disambiguation Ying Chen Center for Spoken Language Research University of Colorado at Boulder yc@colorado.edu James Martin Department of Computer Science University of Colorado at Boulder James.Martin@colorado.edu Abstract The increasing use of large open-domain document sources is exacerbating the problem of ambiguity in named entities. This paper explores the use of a range of syntactic and semantic features in unsu- pervised clustering of documents that re- sult from ad hoc queries containing names. From these experiments, we find that the use of robust syntactic and semantic fea- tures can significantly improve the state of the art for disambiguation performance for personal names for both Chinese and Eng- lish. 1 Introduction An ever-increasing number of question answering, summarization and information extraction systems are coming to rely on heterogeneous sets of documents returned by open-domain search en- gines from collections over which application developers have no control. A frequent special case of these applications involves queries containing named entities of various sorts and receives as a result a large set of possibly relevant documents upon which further deeper processing is focused. Not surprisingly, many, if not most, of the returned documents will be irrelevant to the goals of the application because of the massive ambiguity associated with the query names of people, places and organizations in large open collections. Without some means of separating documents that contain mentions of distinct entities, most of these applications will produce incorrect results. The work presented here, there- fore, addresses the problem of automatically problem of automatically separating sets of news documents generated by queries containing per- sonal names into coherent partitions. The approach we present here combines unsu- pervised clustering methods with robust syntactic and semantic processing to automatically cluster returned news documents (and thereby entities) into homogeneous sets. This work follows on the work of Bagga & Baldwin (1998), Mann & Yarowsky (2003), Niu et al. (2004), Li et al. (2004), Pedersen et al. (2005), and Malin (2005). The results described here advance this work through the use of syntactic and semantic features that can be robustly extracted from the kind of arbitrary news texts typically returned from open- domain sources. The specific contributions reported here fall into two general areas related to robustness. In the first, we explore the use of features extracted from syntactic and semantic processing at a level that is robust to changes in genre and language. In par- ticular, we seek to go beyond the kind of bag of local words features employed in earlier systems (Bagga & Baldwin, 1998; Gooi & Allan, 2004; Pedersen et al., 2005) that did not attempt to ex- ploit deep semantic features that are difficult to extract, and to go beyond the kind of biographical information (Mann & Yarowsky, 2003) that is unlikely to occur with great frequency (such as place …

Content truncated to 500 words