Exploring Syntactic Features for Relation Extraction using a Convolution Tree Kernel

Proceedings of the Human Language Technology Conference of the North American Chapter of the ACL, pages 288–295, New York, June 2006. c⃝2006 Association for Computational Linguistics Exploring Syntactic Features for Relation Extraction using a Convolution Tree Kernel Min ZHANG Jie ZHANG Jian SU Institute for Infocomm Research 21 Heng Mui Keng Terrace, Singapore 119613 {mzhang, zhangjie, sujian}@i2r.a-star.edu.sg Abstract This paper proposes to use a convolution kernel over parse trees to model syntactic structure information for relation extrac- tion. Our study reveals that the syntactic structure features embedded in a parse tree are very effective for relation extrac- tion and these features can be well cap- tured by the convolution tree kernel. Evaluation on the ACE 2003 corpus shows that the convolution kernel over parse trees can achieve comparable per- formance with the previous best-reported feature-based methods on the 24 ACE re- lation subtypes. It also shows that our method significantly outperforms the pre- vious two dependency tree kernels on the 5 ACE relation major types. 1 Introduction Relation extraction is a subtask of information ex- traction that finds various predefined semantic re- lations, such as location, affiliation, rival, etc., between pairs of entities in text. For example, the sentence “George Bush is the president of the United States.” conveys the semantic relation “President” between the entities “George Bush” (PER) and “the United States” (GPE: a Geo-Political Entity --- an entity with land and a government (ACE, 2004)). Prior feature-based methods for this task (Kambhatla 2004; Zhou et al., 2005) employed a large amount of diverse linguistic features, varying from lexical knowledge, entity mention informa- tion to syntactic parse trees, dependency trees and semantic features. Since a parse tree contains rich syntactic structure information, in principle, the features extracted from a parse tree should contrib- ute much more to performance improvement for relation extraction. However it is reported (Zhou et al., 2005; Kambhatla, 2004) that hierarchical struc- tured syntactic features contributes less to per- formance improvement. This may be mainly due to the fact that the syntactic structure information in a parse tree is hard to explicitly describe by a vector of linear features. As an alternative, kernel meth- ods (Collins and Duffy, 2001) provide an elegant solution to implicitly explore tree structure features by directly computing the similarity between two trees. But to our surprise, the sole two-reported dependency tree kernels for relation extraction on the ACE corpus (Bunescu and Mooney, 2005; Cu- lotta and Sorensen, 2004) showed much lower per- formance than the feature-based methods. One may ask: are the syntactic tree features very useful for relation extraction? Can tree kernel methods effectively capture the syntactic tree features and other various features that have been proven useful in the feature-based methods? In this paper, we demonstrate the effectiveness of the syntactic tree features for relation extraction and study how to capture such features via a con- volution tree kernel. We also study how to select the optimal feature space (e.g. the set of sub-trees to represent relation instances) to …

Content truncated to 500 words