Combining the Best of Two Worlds: A Hybrid Approach to Multilingual Coreference Resolution

Proceedings of the Joint Conference on EMNLP and CoNLL: Shared Task, pages 56–63, Jeju Island, Korea, July 13, 2012. c⃝2012 Association for Computational Linguistics Combining the Best of Two Worlds: A Hybrid Approach to Multilingual Coreference Resolution Chen Chen and Vincent Ng Human Language Technology Research Institute University of Texas at Dallas Richardson, TX 75083-0688 {yzcchen,vince}@hlt.utdallas.edu Abstract We describe our system for the CoNLL-2012 shared task, which seeks to model corefer- ence in OntoNotes for English, Chinese, and Arabic. We adopt a hybrid approach to coreference resolution, which combines the strengths of rule-based methods and learning- based methods. Our official combined score over all three languages is 56.35. In particu- lar, our score on the Chinese test set is the best among the participating teams. 1 Introduction The CoNLL-2012 shared task extends last year's task on coreference resolution from a monolingual to a multilingual setting (Pradhan et al., 2012). Unlike the SemEval-2010 shared task on Coreference Reso- lution in Multiple Languages (Recasens et al., 2010), which focuses on coreference resolution in European languages, the CoNLL shared task is arguably more challenging: it focuses on three languages that come from very different language families, namely En- glish, Chinese, and Arabic. We designed a system for resolving references in all three languages. Specifically, we participated in four tracks: the closed track for all three lan- guages, and the open track for Chinese. In compari- son to last year's participating systems, our resolver has two distinguishing characteristics. First, unlike last year's resolvers, which adopted either a rule- based method or a learning-based method, we adopt a hybrid approach to coreference resolution, attempt- ing to combine the strengths of both methods. Sec- ond, while last year's resolvers did not exploit genre- specific information, we optimize our system's pa- rameters with respect to each genre. Our decision to adopt a hybrid approach is mo- tivated by the observation that rule-based meth- ods and learning-based methods each have their unique strengths. As shown by the Stanford coref- erence resolver (Lee et al., 2011), the winner of last year's shared task, many coreference relations in OntoNotes can be identified using a fairly small set of simple hand-crafted rules. On the other hand, our prior work on machine learning for coreference res- olution suggests that coreference-annotated data can be profitably exploited to (1) induce lexical features (Rahman and Ng, 2011a, 2011b) and (2) optimize system parameters with respect to the desired coref- erence evaluation measure (Ng, 2004, 2009). Our system employs a fairly standard architecture, performing mention detection prior to coreference resolution. As we will see, however, the parameters of these two components are optimized jointly with respect to the desired evaluation measure. In the rest of this paper, we describe the men- tion detection component (Section 2) and the coref- erence resolution component (Section 3), show how their parameters are jointly optimized (Section 4), and present evaluation results on the development set and the official test set (Section 5). 2 Mention Detection To build a mention detector …

Content truncated to 500 words