A Multi-Pass Sieve for Coreference Resolution

Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing, pages 492–501, MIT, Massachusetts, USA, 9-11 October 2010. c⃝2010 Association for Computational Linguistics A Multi-Pass Sieve for Coreference Resolution Karthik Raghunathan, Heeyoung Lee, Sudarshan Rangarajan, Nathanael Chambers, Mihai Surdeanu, Dan Jurafsky, Christopher Manning Computer Science Department Stanford University, Stanford, CA 94305 {kr,heeyoung,sudarshn,natec,mihais,jurafsky,manning}@stanford.edu Abstract Most coreference resolution models determine if two mentions are coreferent using a single function over a set of constraints or features. This approach can lead to incorrect decisions as lower precision features often overwhelm the smaller number of high precision ones. To overcome this problem, we propose a simple coreference architecture based on a sieve that applies tiers of deterministic coreference mod- els one at a time from highest to lowest preci- sion. Each tier builds on the previous tier’s entity cluster output. Further, our model prop- agates global information by sharing attributes (e.g., gender and number) across mentions in the same cluster. This cautious sieve guar- antees that stronger features are given prece- dence over weaker ones and that each deci- sion is made using all of the information avail- able at the time. The framework is highly modular: new coreference modules can be plugged in without any change to the other modules. In spite of its simplicity, our ap- proach outperforms many state-of-the-art su- pervised and unsupervised models on several standard corpora. This suggests that sieve- based approaches could be applied to other NLP tasks. 1 Introduction Recent work on coreference resolution has shown that a rich feature space that models lexical, syn- tactic, semantic, and discourse phenomena is cru- cial to successfully address the task (Bengston and Roth, 2008; Haghighi and Klein, 2009; Haghighi and Klein, 2010). When such a rich representation is available, even a simple deterministic model can achieve state-of-the-art performance (Haghighi and Klein, 2009). By and large most approaches decide if two men- tions are coreferent using a single function over all these features and information local to the two men- tions.1 This is problematic for two reasons: (1) lower precision features may overwhelm the smaller number of high precision ones, and (2) local infor- mation is often insufficient to make an informed de- cision. Consider this example: The second attack occurred after some rocket firings aimed, apparently, toward [the israelis], apparently in retaliation. [we]’re checking our facts on that one. ... the president, quoted by ari fleischer, his spokesman, is saying he’s concerned the strike will undermine efforts by palestinian authorities to bring an end to terrorist at- tacks and does not contribute to the security of [israel]. Most state-of-the-art models will incorrectly link we to the israelis because of their proximity and compatibility of attributes (both we and the israelis are plural). In contrast, a more cautious approach is to first cluster the israelis with israel because the de- monymy relation is highly precise. This initial clus- tering step will assign the correct animacy attribute (inanimate) to the corresponding geo-political entity, which will prevent the incorrect merging …

Content truncated to 500 words