Same Referent, Different Words: Unsupervised Mining of Opaque Coreferent Mentions

Proceedings of NAACL-HLT 2013, pages 897–906, Atlanta, Georgia, 9–14 June 2013. c⃝2013 Association for Computational Linguistics Same Referent, Different Words: Unsupervised Mining of Opaque Coreferent Mentions Marta Recasens*, Matthew Can†, and Dan Jurafsky* *Linguistics Department, Stanford University, Stanford, CA 94305 †Computer Science Department, Stanford University, Stanford, CA 94305 recasens@google.com, {mattcan,jurafsky}@stanford.edu Abstract Coreference resolution systems rely heav- ily on string overlap (e.g., Google Inc. and Google), performing badly on mentions with very different words (opaque mentions) like Google and the search giant. Yet prior at- tempts to resolve opaque pairs using ontolo- gies or distributional semantics hurt precision more than improved recall. We present a new unsupervised method for mining opaque pairs. Our intuition is to restrict distributional se- mantics to articles about the same event, thus promoting referential match. Using an En- glish comparable corpus of tech news, we built a dictionary of opaque coreferent mentions (only 3% are in WordNet). Our dictionary can be integrated into any coreference system (it increases the performance of a state-of-the-art system by 1% F1 on all measures) and is eas- ily extendable by using news aggregators. 1 Introduction Repetition is one of the most common coreferential devices in written text, making string-match features important to all coreference resolution systems. In fact, the scores achieved by just head match and a rudimentary form of pronominal resolution1 are not far from that of state-of-the-art systems (Recasens and Hovy, 2010). This suggests that opaque men- tions (i.e., lexically different) such as iPad and the Cupertino slate are a serious problem for modern systems: they comprise 65% of the non-pronominal 1Closest NP with the same gender and number. errors made by the Stanford system on the CoNLL- 2011 data. Solving this problem is critical for over- coming the recall gap of state-of-the-art systems (Haghighi and Klein, 2010; Stoyanov et al., 2009). Previous systems have turned either to ontologies (Ponzetto and Strube, 2006; Uryupina et al., 2011; Rahman and Ng, 2011) or distributional semantics (Yang and Su, 2007; Kobdani et al., 2011; Bansal and Klein, 2012) to help solve these errors. But nei- ther semantic similarity nor hypernymy are the same as coreference: Microsoft and Google are distribu- tionally similar but not coreferent; people is a hy- pernym of both voters and scientists, but the peo- ple can corefer with the voters, but is less likely to corefer with the scientists. Thus ontologies lead to precision problems, and to recall problems like missing NE descriptions (e.g., Apple and the iPhone maker) and metonymies (e.g., agreement and word- ing), while distributional systems lead to precision problems like coreferring Microsoft and the Moun- tain View giant because of their similar vector rep- resentation (release, software, update). We increase precision by drawing on the intuition that referents that are both similar and participate in the same event are likely to corefer. We restrict dis- tributional similarity to collections of articles that discuss the same event. In the following two doc- uments on the Nexus One from different sources, we take the …

Content truncated to 500 words