Incorporating Rule-based and Statistic-based Techniques for Coreference Resolution

Proceedings of the Joint Conference on EMNLP and CoNLL: Shared Task, pages 107–112, Jeju Island, Korea, July 13, 2012. c⃝2012 Association for Computational Linguistics Incorporating Rule-based and Statistic-based Techniques for Coreference Resolution Ruifeng Xu, Jun Xu, Jie Liu, Chengxiang Liu, Chengtian Zou, Lin Gui, Yanzhen Zheng, Peng Qu Human Language Technology Group, Key Laboratory of Network Oriented Intelligent Computation, Shenzhen Graduate School, Harbin Institute of Technology, China {xuruifeng.hitsz;hit.xujun;lyjxcz;matitalk;chsky.zou;monta3pt; zhyz.zheng;viphitqp@gmail.com} Abstract This paper describes a coreference resolution system for CONLL 2012 shared task developed by HLT_HITSZ group, which incorporates rule-based and statistic-based techniques. The system performs coreference resolution through the mention pair classification and linking. For each detected mention pairs in the text, a Decision Tree (DT) based binary classifier is applied to determine whether they form a coreference. This classifier incorporates 51 and 61 selected features for English and Chinese, respectively. Meanwhile, a rule-based classifier is applied to recognize some specific types of coreference, especially the ones with long distances. The outputs of these two classifiers are merged. Next, the recognized coreferences are linked to generate the final coreference chain. This system is evaluated on English and Chinese sides (Closed Track), respectively. It achieves 0.5861 and 0.6003 F1 score on the development data of English and Chinese, respectively. As for the test dataset, the achieved F1 scores are 0.5749 and 0.6508, respectively. This encouraging performance shows the effectiveness of our proposed coreference resolution system. 1 Introduction Coreference resolution aims to find out the different mentions in a document which refer to the same entity in reality (Sundheim and Beth, 1995; Lang et al. 1997; Chinchor and Nancy, 1998;). It is a core component in natural language processing and information extraction. Both rule-based approach (Lee et al. 2011) and statistic-based approach (Soon et al., 2001; Ng and Cardie, 2002; Bengtson and Roth, 2008; Stoyanov et al., 2009; Chen et al. 2011) are proposed in coreference resolution study. Besides the frequently used syntactic and semantic features, the more linguistic features are exploited in recent works (Versley, 2007; Kong et al. 2010). CoNLL-2012 proposes a shared task, “Modeling multilingual unrestricted coreference in the OntoNotes” (Pradhan et al. 2012). This is an extension of the CoNLL-2011 shared task. The task involves automatic anaphoric mention detection and coreference resolution across three languages including English, Chinese and Arabic. HLT_HITSZ group participated in the Closed Track evaluation on English and Chinese side. This paper presents the framework and techniques of HLT_HITSZ system which incorporates both rule- based and statistic-based techniques. In this system, the mentions are firstly identified based on the provided syntactic information. The mention pairs in the document are fed to a Decision Tree based classifier to determine whether they form a coreference or not. The rule-based classifiers are then applied to recognize some specific types of coreference, in particular, the long distance ones. Finally, the recognized coreference are linked to obtain the final coreference resolution results. This system incorporates lexical, syntactical and semantic features. Especially for English, WordNet is used to provide semantic information …

Content truncated to 500 words