Unsupervised Ranking Model for Entity Coreference Resolution

Proceedings of NAACL-HLT 2016, pages 1012–1018, San Diego, California, June 12-17, 2016. c⃝2016 Association for Computational Linguistics Unsupervised Ranking Model for Entity Coreference Resolution Xuezhe Ma and Zhengzhong Liu and Eduard Hovy Language Technologies Institute Carnegie Mellon University Pittsburgh, PA 15213, USA {xuezhem, liu}@cs.cmu.edu, ehovy@cmu.edu Abstract Coreference resolution is one of the first stages in deep language understanding and its impor- tance has been well recognized in the natu- ral language processing community. In this paper, we propose a generative, unsupervised ranking model for entity coreference resolu- tion by introducing resolution mode variables. Our unsupervised system achieves 58.44% F1 score of the CoNLL metric on the English data from the CoNLL-2012 shared task (Pradhan et al., 2012), outperforming the Stanford deter- ministic system (Lee et al., 2013) by 3.01%. 1 Introduction Entity coreference resolution has become a critical component for many Natural Language Processing (NLP) tasks. Systems requiring deep language un- derstanding, such as information extraction (Wellner et al., 2004), semantic event learning (Chambers and Jurafsky, 2008; Chambers and Jurafsky, 2009), and named entity linking (Durrett and Klein, 2014; Ji et al., 2014) all benefit from entity coreference infor- mation. Entity coreference resolution is the task of iden- tifying mentions (i.e., noun phrases) in a text or dialogue that refer to the same real-world entities. In recent years, several supervised entity corefer- ence resolution systems have been proposed, which, according to Ng (2010), can be categorized into three classes — mention-pair models (McCarthy and Lehnert, 1995), entity-mention models (Yang et al., 2008a; Haghighi and Klein, 2010; Lee et al., 2011) and ranking models (Yang et al., 2008b; Durrett and Klein, 2013; Fernandes et al., 2014) — among which ranking models recently obtained state-of-the-art performance. However, the manu- ally annotated corpora that these systems rely on are highly expensive to create, in particular when we want to build data for resource-poor languages (Ma and Xia, 2014). That makes unsupervised ap- proaches, which only require unannotated text for training, a desirable solution to this problem. Several unsupervised learning algorithms have been applied to coreference resolution. Haghighi and Klein (2007) presented a mention-pair non- parametric fully-generative Bayesian model for un- supervised coreference resolution. Based on this model, Ng (2008) probabilistically induced corefer- ence partitions via EM clustering. Poon and Domin- gos (2008) proposed an entity-mention model that is able to perform joint inference across mentions by using Markov Logic. Unfortunately, these un- supervised systems’ performance on accuracy sig- nificantly falls behind those of supervised systems, and are even worse than the deterministic rule-based systems. Furthermore, there is no previous work exploring the possibility of developing an unsuper- vised ranking model which achieved state-of-the- art performance under supervised settings for entity coreference resolution. In this paper, we propose an unsupervised genera- tive ranking model for entity coreference resolution. Our experimental results on the English data from the CoNLL-2012 shared task (Pradhan et al., 2012) show that our unsupervised system outperforms the Stanford deterministic system (Lee et al., 2013) by 3.01% absolute on the CoNLL …

Content truncated to 500 words