Accurate Semantic Class Classifier for Coreference Resolution

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 1232–1240, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP Accurate Semantic Class Classifier for Coreference Resolution Zhiheng Huang1, Guangping Zeng1,2, Weiqun Xu3, and Asli Celikyilmaz1 1EECS Department, University of California at Berkeley, CA 94720, USA {zhiheng,gpzeng,asli}@eecs.berkeley.edu 2Computer Science Department, School of Information Engineering, University of Science and Technology Beijing, China 3ThinkIT, Institute of Acoustics, Chinese Academy of Sciences, Beijing, 100190, China xuweiqun@hccl.ioa.ac.cn Abstract There have been considerable attempts to incorporate semantic knowledge into coreference resolution systems: different knowledge sources such as WordNet and Wikipedia have been used to boost the per- formance. In this paper, we propose new ways to extract WordNet feature. This feature, along with other features such as named entity feature, can be used to build an accurate semantic class (SC) classifier. In addition, we analyze the SC classifica- tion errors and propose to use relaxed SC agreement features. The proposed accu- rate SC classifier and the relaxation of SC agreement features on ACE2 coreference evaluation can boost our baseline system by 10.4% and 9.7% using MUC score and anaphor accuracy respectively. 1 Introduction Coreference resolution is used to determine which noun phrases (including pronouns, proper names, and common nouns) refer to the same entities in documents. Much work on coreference resolution is based on (Soon et al., 2001), which built a de- cision tree classifier to label pairs of mentions as coreferent or not. Recent work aims to improve the performance from two aspects: new models and new features. The former cast the pair wise mention classifications into various forms such as the best path in a Bell tree (Luo et al., 2004), the best graph cut (Nicolae and Nicolae, 2006), in- teger linear programming (Denis and Baldridge, 2007) and graph partition based conditional model (McCallum and Wellner, 2004). The latter de- velop and investigate new linguistic features for the problem. For instance, WordNet (Poesio et al., 2004), Wikipedia (Ponzetto and Strube, 2006), se- mantic neighbor words (Ng, 2007a), and pattern based features (Yang and Su, 2007) have been ex- tensively studied. Deeper linguistic knowledge is required to en- able the coreference resolution to reach a higher level of performance (Kehler et al., 2004). An im- portant type of semantic knowledge that has been employed in coreference resolution system is the semantic class (SC) of an NP, which can be used to filter out the coreference between semantically incompatible NPs. However, the difficulty is to accurately compute the semantic class features. In this paper, we show that the WordNet may not be efficiently employed in the traditional way such as (Soon et al., 2001; Ng, 2007a; Ponzetto and Strube, 2006) to compute the semantic class fea- tures. We introduce new ways to use the WordNet and the experiments show its effectiveness in de- termining the semantic classes for noun phrases. In addition, we analyze the classification errors of the SC classifier and propose to use relaxed SC agreement features. With these proposed features …

Content truncated to 500 words