A Systematic Exploration of the Feature Space for Relation Extraction

Proceedings of NAACL HLT 2007, pages 113–120, Rochester, NY, April 2007. c⃝2007 Association for Computational Linguistics A Systematic Exploration of the Feature Space for Relation Extraction Jing Jiang and ChengXiang Zhai Department of Computer Science University of Illinois at Urbana-Champaign Urbana, IL 61801 {jiang4,czhai}@cs.uiuc.edu Abstract Relation extraction is the task of find- ing semantic relations between entities from text. The state-of-the-art methods for relation extraction are mostly based on statistical learning, and thus all have to deal with feature selection, which can significantly affect the classification per- formance. In this paper, we systemat- ically explore a large space of features for relation extraction and evaluate the ef- fectiveness of different feature subspaces. We present a general definition of fea- ture spaces based on a graphic represen- tation of relation instances, and explore three different representations of relation instances and features of different com- plexities within this framework. Our ex- periments show that using only basic unit features is generally sufficient to achieve state-of-the-art performance, while over- inclusion of complex features may hurt the performance. A combination of fea- tures of different levels of complexity and from different sentence representations, coupled with task-oriented feature prun- ing, gives the best performance. 1 Introduction An important information extraction task is relation extraction, whose goal is to detect and characterize semantic relations between entities in text. For ex- ample, the text fragment “hundreds of Palestinians converged on the square” contains the located rela- tion between the Person entity “hundreds of Pales- tinians” and the Bounded-Area entity “the square”. Relation extraction has applications in many do- mains, including finding affiliation relations from web pages and finding protein-protein interactions from biomedical literature. Recent studies on relation extraction have shown the advantages of discriminative model-based sta- tistical machine learning approach to this problem. There are generally two lines of work following this approach. The first utilizes a set of carefully se- lected features obtained from different levels of text analysis, from part-of-speech (POS) tagging to full parsing and dependency parsing (Kambhatla, 2004; Zhao and Grishman, 2005; Zhou et al., 2005)1. The second line of work designs kernel functions on some structured representation (sequences or trees) of the relation instances to capture the similarity be- tween two relation instances (Zelenko et al., 2003; Culotta and Sorensen, 2004; Bunescu and Mooney, 2005a; Bunescu and Mooney, 2005b; Zhang et al., 2006a; Zhang et al., 2006b). Of particular interest among the various kernels proposed are the convolu- tion kernels (Bunescu and Mooney, 2005b; Zhang et al., 2006a), because they can efficiently compute the similarity between two instances in a huge feature space due to their recursive nature. Apart from their computational efficiency, convolution kernels also implicitly correspond to some feature space. There- fore, both lines of work rely on an appropriately de- 1Although Zhao and Grishman (2005) defined a number of kernels for relation extraction, the method is essentially similar to feature-based methods. 113 fined set of features. As in any learning problem, the choice of features can affect the performance …

Content truncated to 500 words