Relation Extraction Using Convolution Tree Kernel Expanded with Entity Features

Relation Extraction Using Convolution Tree Kernel Expanded with Entity Features* Longhua Qian, Guodong Zhou, Qiaomin Zhu, Peide Qian School of Computer Science and Technology, Soochow University. 215006, Suzhou, China {qianlonghua, gdzhou, qmzhu, pdqian}@suda.edu.cn Abstract. This paper proposes a convolution tree kernel-based approach for relation extraction where the parse tree is expanded with entity features such as entity type, subtype, and mention level etc. Our study indicates that not only can our method effectively capture both syntactic structure and entity information of relation instances, but also can avoid the difficulty with tuning the parameters in composite kernels. We also demonstrate that predicate verb information can be used to further improve the performance, though its enhancement is limited. Evaluation on the ACE2004 benchmark corpus shows that our system slightly outperforms both the previous best-reported feature-based and kernel-based systems. Keywords: Information Extraction; Kernel-based Relation Extraction; Support Vector Machines. 1. Introduction Information extraction is an important research sub-field in natural language processing (NLP) which aims to identify relevant information from large amount of text documents in digital archives and the WWW. Information extraction subsumes three main tasks, including Entity Detection and Tracking (EDT), Relation Detection and Characterization (RDC), and Event Detection and Characterization (EDC). This paper will focus on the ACE RDC task1 and employ kernel method to extract semantic relationships between named entity pairs. Many feature-based approaches transform relation instances into feature vectors of high dimension, and compute the inner dot product between these feature vectors. Current research (Kambhatla 2004, Zhao et al 2005, Zhou et al. 2005, Wang et al. 2006) shows that it is very difficult to extract new effective features from relation examples. Kernel methods are non-parametric estimation techniques that computer a kernel function between data instances. By avoiding transforming data examples into feature vectors, kernel methods can implicitly explore much larger feature space than could be searched by a * This research is supported by Project 60673041 under the National Natural Science Foundation of China and Project 2006AA01Z147 under the “863” National High-Tech Research and Development of China. We would also like to thank Dr. Alessando Moschitti for his great help in using his Tree Kernel Toolkits, including binary package and source codes. Copyright 2007 by Longhua Qian, Guodong Zhou, Qiaomin Zhu, Peide Qian 1 http://www.ldc.upenn.edu/Projects/ACE/. 415 feature-based approach. Thereafter, kernel methods especially on discrete structures (Haussler 1999) attract more and more attentions in relation extraction as well as other fields in NLP. Prior work on kernel methods for relation extraction includes Zelenko et al. (2003), Culotta and Sorensen (2004), Bunescu and Mooney (2005). Due to strong constraints that matching nodes be at the same layer and in the identical path starting from the roots to the current nodes, their kernels achieve good precision but much lower recall on the ACE2003 corpus. Zhang et al. (2006) proposed a composite kernel that consists of two individual kernels: an entity kernel that allows for entity-related features and a convolution parse tree kernel that models syntactic information of relation examples. However, their …

Content truncated to 500 words