Context-Sensitive Convolution Tree Kernel for Pronoun Resolution

Context-Sensitive Convolution Tree Kernel for Pronoun Resolution ZHOU GuoDong KONG Fang ZHU Qiaoming JiangSu Provincial Key Lab for Computer Information Processing Technology School of Computer Science and Technology Soochow Univ. Suzhou, China 215006 Email: {gdzhou, kongfang, qmzhu}@suda.edu.cn Abstract This paper proposes a context-sensitive convo- lution tree kernel for pronoun resolution. It re- solves two critical problems in previous researches in two ways. First, given a parse tree and a pair of an anaphor and an antecedent candidate, it implements a dynamic-expansion scheme to automatically determine a proper tree span for pronoun resolution by taking predicate- and antecedent competitor-related information into consideration. Second, it ap- plies a context-sensitive convolution tree ker- nel, which enumerates both context-free and context-sensitive sub-trees by considering their ancestor node paths as their contexts. Evalua- tion on the ACE 2003 corpus shows that our dynamic-expansion tree span scheme can well cover necessary structured information in the parse tree for pronoun resolution and the con- text-sensitive tree kernel much outperforms previous tree kernels. 1 Introduction It is well known that syntactic structured informa- tion plays a critical role in many critical NLP ap- plications, such as parsing, semantic role labeling, semantic relation extraction and co-reference reso- lution. However, it is still an open question on what kinds of syntactic structured information are effective and how to well incorporate such struc- tured information in these applications. Much research work has been done in this direc- tion. Prior researches apply feature-based methods to select and define a set of flat features, which can be mined from the parse trees, to represent particu- lar structured information in the parse tree, such as the grammatical role (e.g. subject or object), ac- cording to the particular application. Indeed, such feature-based methods have been widely applied in parsing (Collins 1999; Charniak 2001), semantic role labeling (Pradhan et al 2005), semantic rela- tion extraction (Zhou et al 2005) and co-reference resolution (Lapin and Leass 1994; Aone and Ben- nett 1995; Mitkov 1998; Yang et al 2004; Luo and Zitouni 2005; Bergsma and Lin 2006). The major problem with feature-based methods on exploring structured information is that they may fail to well capture complex structured information, which is critical for further performance improvement. The current trend is to explore kernel-based methods (Haussler, 1999) which can implicitly explore features in a high dimensional space by employing a kernel to calculate the similarity be- tween two objects directly. In particular, the ker- nel-based methods could be very effective at reducing the burden of feature engineering for structured objects in NLP, e.g. the parse tree struc- ture in coreference resolution. During recent years, various tree kernels, such as the convolution tree kernel (Collins and Duffy 2001), the shallow parse tree kernel (Zelenko et al 2003) and the depend- ency tree kernel (Culota and Sorensen 2004), have been proposed in the literature. Among previous tree kernels, the convolution tree kernel represents the state-of-the-art and have been successfully ap- plied by Collins and Duffy (2002) on parsing, Mo- schitti (2004) on …

Content truncated to 500 words