A Listwise Approach to Coreference Resolution in Multiple Languages

A Listwise Approach to Coreference Resolution in Multiple Languages Oanh Thi Tran, Bach Xuan Ngo, Minh Le Nguyen, and Akira Shimazu School of Information Science, JAIST, 1 - 8 Asahidai, Nomi, Ishikawa, Japan 923 - 1211 {oanhtt, bachnx, nguyenml, shimazu}@jaist.ac.jp Abstract. This paper presents a listwise approach as an alternative to commonly used pair- wise approaches to the task of coreference resolution in multiple languages. In this listwise approach, all antecedent candidates are examined simultaneously and assigned corresponding scores expressing the probability that each candidate is coreferent with a given mention. The experimental results on the corpora of SemEval-2010 shared task 1 showed that our proposed system gave the good results in English and Spanish, and comparative results in Catalan when compared to previous participating systems. These results prove that this approach is appropriate and quite efficient for Coreference Resolution in Multiple Languages. Keywords: listwise approach, coreference resolution, multiple languages. 1 Introduction Reference resolution (Jurafsky and Martin, 2009) (chapter 21, section 21.4) is a task of determin- ing to which entities are referred by which linguistic expressions. This task plays an important role in a large number of NLP applications such as Information Retrieval, Question Answering and Machine Translation. Therefore, it has attracted many attentions within the NLP community. Many works on various aspects (linguistic features (Ng, V., 2007), (Haghighi and Klein, 2009); machine learning models (Soon et al., 2001); multiple languages (Recasens et al., 2010a); and so on) of the coreference resolution task have been published. Until the release of the SemEval-2010 task 1 (Recasens et al., 2010a), there has no competi- tion or public corpus that allows evaluating different coreference resolution systems in multiple languages. Most published systems only focus on a specific language and use the same data sets for example ACE or MUC corpora to train and test the systems. This makes the systems easy to unintentionally adapt themselves to the corpus but not to the problem in general. Therefore, this SemEval-2010 task 1 (Recasens et al., 2010a) made it possible to evaluate and compare various automatic coreference resolution systems in the aspects of: (i) the portability of systems across languages, (ii) the relevance of different levels of linguistic information, and (iii) the behavior of scoring metrics. This shared task attracted lots of researchers’ attentions, but finally only six teams submitted their final results. The participating systems differed in terms of architecture, machine learning methods, etc. These systems mostly based on pairwise models, graph partitioning and entity- mention models. Unfortunately, these models suffered from an important weakness (Ng, V., 2010). In these models, each antecedent candidate is resolved independently with the other candidates. So the models could not determine the best candidate in the relation with the other candidates. To address this drawback, ranking models were proved to be a useful solution (Denis and Baldridge, 2007), (Ng, V., 2005), (Yang et al., 2003). Motivated from ranking models, in this paper, we present our proposal approach for learning-based reference resolution task in multiple languages. Copyright 2011 by …

Content truncated to 500 words