Dependency-driven Anaphoricity Determination for Coreference Resolution

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 599–607, Beijing, August 2010 Dependency-driven Anaphoricity Determination for Coreference Resolution Fang Kong Guodong Zhou Longhua Qian Qiaoming Zhu* JiangSu Provincial Key Lab for Computer Information Processing Technology School of Computer Science and Technology Soochow University {kongfang, gdzhou, qianlonghua, qmzhu}@suda.edu.cn * Corresponding author Abstract This paper proposes a dependency-driven scheme to dynamically determine the syn- tactic parse tree structure for tree ker- nel-based anaphoricity determination in coreference resolution. Given a full syntactic parse tree, it keeps the nodes and the paths related with current mention based on con- stituent dependencies from both syntactic and semantic perspectives, while removing the noisy information, eventually leading to a dependency-driven dynamic syntactic parse tree (D-DSPT). Evaluation on the ACE 2003 corpus shows that the D-DSPT out- performs all previous parse tree structures on anaphoricity determination, and that apply- ing our anaphoricity determination module in coreference resolution achieves the so far best performance. 1 Introduction Coreference resolution aims to identify which noun phrases (NPs, or mentions) refer to the same real-world entity in a text. According to Webber (1979), coreference resolution can be decomposed into two complementary sub-tasks: (1) anaphoricity determination, determining whether a given NP is anaphoric or not; and (2) anaphor resolution, linking together multiple mentions of a given entity in the world. Al- though machine learning approaches have per- formed reasonably well in coreference resolu- tion without explicit anaphoricity determina- tion (e.g. Soon et al. 2001; Ng and Cardie 2002b; Yang et al. 2003, 2008; Kong et al. 2009), knowledge of NP anaphoricity is ex- pected to much improve the performance of a coreference resolution system, since a non-anaphoric NP does not have an antecedent and therefore does not need to be resolved. Recently, anaphoricity determination has been drawing more and more attention. One common approach involves the design of some heuristic rules to identify specific types of non-anaphoric NPs, such as pleonastic it (e.g. Paice and Husk 1987; Lappin and Leass 1994, Kennedy and Boguraev 1996; Denber 1998) and definite descriptions (e.g. Vieira and Poe- sio 2000). Alternatively, some studies focus on using statistics to tackle this problem (e.g., Bean and Riloff 1999; Bergsma et al. 2008) and others apply machine learning approaches (e.g. Evans 2001;Ng and Cardie 2002a, 2004,2009; Yang et al. 2005; Denis and Bal- bridge 2007; Luo 2007; Finkel and Manning 2008; Zhou and Kong 2009). As a representative, Zhou and Kong (2009) directly employ a tree kernel-based method to automatically mine the non-anaphoric informa- tion embedded in the syntactic parse tree. One main advantage of the kernel-based methods is that they are very effective at reducing the burden of feature engineering for structured objects. Indeed, the kernel-based methods have been successfully applied to mine structured information in various NLP applications like syntactic parsing (Collins and Duffy, 2001; Moschitti, 2004), semantic relation extraction (Zelenko et al., 2003; Zhao and Grishman, 2005; Zhou et al. 2007; Qian et al., 2008), se- mantic role labeling (Moschitti, 2004); corefer- ence …

Content truncated to 500 words