Entity-Centric Coreference Resolution with Model Stacking

Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing, pages 1405–1415, Beijing, China, July 26-31, 2015. c⃝2015 Association for Computational Linguistics Entity-Centric Coreference Resolution with Model Stacking Kevin Clark Computer Science Department Stanford University kevclark@cs.stanford.edu Christopher D. Manning Computer Science Department Stanford University manning@cs.stanford.edu Abstract Mention pair models that predict whether or not two mentions are coreferent have historically been very effective for coref- erence resolution, but do not make use of entity-level information. However, we show that the scores produced by such models can be aggregated to define pow- erful entity-level features between clusters of mentions. Using these features, we train an entity-centric coreference system that learns an effective policy for building up coreference chains incrementally. The mention pair scores are also used to prune the search space the system works in, al- lowing for efficient training with an exact loss function. We evaluate our system on the English portion of the 2012 CoNLL Shared Task dataset and show that it im- proves over the current state of the art. 1 Introduction Coreference resolution, the task of identifying mentions in a text that refer to the same real world entity, is an important aspect of text understanding and has numerous applications. Many approaches to coreference resolution learn a scoring function defined over mention pairs to guide the corefer- ence decisions (Soon et al., 2001; Ng and Cardie, 2002; Bengtson and Roth, 2008). However, such systems do not make use of entity-level informa- tion, i.e., features between clusters of mentions in- stead of pairs. Using entity-level information is valuable be- cause it allows early coreference decisions to in- form later ones. For example, finding that Clin- ton and she corefer makes it more likely that Clin- ton corefers with Hillary Clinton than Bill Clin- ton due to gender agreement constraints. Such in- formation has been incorporated successfully into entity-centric coreference systems that build up coreference clusters incrementally, using the in- formation from the partially completed corefer- ence chains produced so far to guide later deci- sions (Raghunathan et al., 2010; Stoyanov and Eisner, 2012; Ma et al., 2014). However, defining useful features between clus- ters of mentions and learning an effective policy for incrementally building up clusters can be chal- lenging, and many recent state-of-the-art systems work entirely or almost entirely over pairs of men- tions (Fernandes et al., 2012; Durrett and Klein, 2013; Chang et al., 2013). In this paper we in- troduce a novel coreference system that combines the advantages of mention pair and entity-centric systems with model stacking. We first propose two mention pair models designed to capture dif- ferent linguistic phenomena in coreference resolu- tion. We then describe how the probabilities pro- duced by these models can be used to generate expressive features between clusters of mentions. Using these features, we train an entity-centric in- cremental coreference system. The entity-centric system builds up coreference chains with agglomerative clustering: each men- tion starts in …

Content truncated to 500 words