Combining Dependency and Constituent-based Syntactic Information for Anaphoricity Determination in Coreference Resolution

Combining Dependency and Constituent-based Syntactic Information for Anaphoricity Determination in Coreference Resolution Fang Kong a,b and Guodong Zhou a,b a JiangSu Provincial Key Lab for Computer Information Processing Technology bSchool of Computer Science and Technology Soochow University, Suzhou(215006) ,China {kongfang, gdzhou}@suda.edu.cn Abstract. This paper systematically explores the effectiveness of dependency and constituent-based syntactic information for anaphoricity determination. In particular, this paper proposes two ways to combine dependency and constituent-based syntactic information to explore their complementary advantage. One is a dependency-driven constituent-based structured representation, and the other uses a composite kernel. Evaluation on the Automatic Content Extraction (ACE) 2003 corpus shows that dependency and constituent-based syntactic information are quite complementary and proper combination can much improve the performance of anaphoricity determination, and further improve the performance of coreference resolution. Keywords: Dependency Relation, Constituent Parser, Anaphoricity Determination, Coreference Resolution, Tree Kernel, Composite Kernel 1 Introduction Coreference resolution plays a critical role in many NLP applications. It identifies which noun phrases (NPs) refer to the same entity in the real world and can be divided into two subtasks: anaphoricity determination and antecedent identification. The first subtask, anaphoricity determination, determines whether a given noun phrases (NP) is anaphoric or not. Here we say an NP is anaphoric if it has any antecedent in the context preceding it, and non-anaphoric otherwise. The second subtask, antecedent identification, identifies the antecedent of a given anaphoric NP. Although machine learning approaches have performed reasonably well in coreference resolution without explicit anaphoricity determination (e.g. Soon et al. 2001; Ng and Cardie 2002b; Yang et al. 2003, 2008; Kong et al. 2009), knowledge of NP anaphoricity is expected to much improve the performance of a coreference resolution system, since a non-anaphoric NP does not have an antecedent and thus does not need to be resolved. Recently, anaphoricity determination has been the subject of increased attention in coreference resolution. A variety of techniques have been proposed to address anaphoricity determination as an independent task (e.g. Paice and Husk 1987; Lappin and Leass 1994; Kennedy and Boguraev 1996; Denber 1998; Bean and Riloff 1999; Vieira and Poesio 2000; Evans 2001; Cherry, Cherry and Bergsma 2005; Bergsma et al. 2008). Since then, more studies have been done to incorporate anaphoricity determination into coreference resolution in a pipeline way (e.g. Ng and Cardie 2002a; Yang et al. 2005; Kong et al. 2010) or in a joint way (e.g. Denis and Balbridge 2007; Luo 2007; Finkel and Manning 2008; Zhou and Kong 2009; Ng 2009), and achieved promising results. While it is well known that NP anaphoricity interacts with various kinds of structured factors, most of previous studies only consider constituent-based syntactic information and there are few studies on exploring dependency-based syntactic information for anaphoricity determination. In this paper, we first study the effectiveness of dependency-based syntactic information for anaphoricity determination. Then, we propose different ways to combine dependency and constituent-based syntactic information for anaphoricity determination in better exploring their complementary nature. Our study shows that dependency and constituent-based syntactic information are quite complementary and …

Content truncated to 500 words