Using Semantic Relations to Refine Coreference Decisions

Proceedings of Human Language Technology Conference and Conference on Empirical Methods in Natural Language Processing (HLT/EMNLP), pages 17–24, Vancouver, October 2005. c⃝2005 Association for Computational Linguistics Using Semantic Relations to Refine Coreference Decisions Heng Ji David Westbrook Ralph Grishman Department of Computer Science New York University New York, NY, 10003, USA hengji@cs.nyu.edu westbroo@cs.nyu.edu grishman@cs.nyu.edu Abstract We present a novel mechanism for im- proving reference resolution by using the output of a relation tagger to rescore coreference hypotheses. Experiments show that this new framework can im- prove performance on two quite different languages -- English and Chinese. 1 Introduction Reference resolution has proven to be a major obstacle in building robust systems for information extraction, question answering, text summarization and a number of other natural language processing tasks. Most reference resolution systems use represen- tations built out of the lexical and syntactic attrib- utes of the noun phrases (or “mentions”) for which reference is to be established. These attributes may involve string matching, agreement, syntactic dis- tance, and positional information, and they tend to rely primarily on the immediate context of the noun phrases (with the possible exception of sen- tence-spanning distance measures such as Hobbs distance). Though gains have been made with such methods (Tetreault 2001; Mitkov 2000; Soon et al. 2001; Ng and Cardie 2002), there are clearly cases where this sort of local information will not be suf- ficient to resolve coreference correctly. Coreference is by definition a semantic relationship: two noun phrases corefer if they both refer to the same real-world entity. We should therefore expect a successful coreference system to exploit world knowledge, inference, and other forms of semantic information in order to resolve hard cases. If, for example, two nouns refer to people who work for two different organizations, we want our system to infer that these noun phrases cannot corefer. Further progress will likely be aided by flexible frameworks for representing and using the information provided by this kind of semantic relation between noun phrases. This paper tries to make a small step in that di- rection. It describes a robust reference resolver that incorporates a broad range of semantic information in a general news domain. Using an ontology that describes relations between entities (the Auto- mated Content Extraction program1 relation ontol- ogy) along with a training corpus annotated for relations under this ontology, we first train a classi- fier for identifying relations. We then apply the output of this relation tagger to the task of refer- ence resolution. The rest of this paper is structured as follows. Section 2 briefly describes the efforts made by previous researchers to use semantic information in reference resolution. Section 3 describes our own method for incorporating document-level semantic context into coreference decisions. We propose a representation of semantic context that isolates a particularly informative structure of interaction between semantic relations and coreference. Sec- tion 4 explains in detail our strategies for using relation information to modify coreference deci- sions, and the linguistic intuitions behind these strategies. Section …

Content truncated to 500 words