<?xml version="1.0" standalone="yes"?>
<Paper uid="P06-1104">
  <Title>Sydney, July 2006. c(c)2006 Association for Computational Linguistics A Composite Kernel to Extract Relations between Entities with both Flat and Structured Features</Title>
  <Section position="3" start_page="0" end_page="825" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> The goal of relation extraction is to find various predefined semantic relations between pairs of entities in text. The research on relation extraction has been promoted by the Message Understanding Conferences (MUCs) (MUC, 19871998) and Automatic Content Extraction (ACE) program (ACE, 2002-2005). According to the ACE Program, an entity is an object or set of objects in the world and a relation is an explicitly or implicitly stated relationship among entities.</Paragraph>
    <Paragraph position="1"> For example, the sentence &amp;quot;Bill Gates is chairman and chief software architect of Microsoft Corporation.&amp;quot; conveys the ACE-style relation &amp;quot;EMPLOYMENT.exec&amp;quot; between the entities &amp;quot;Bill Gates&amp;quot; (PERSON.Name) and &amp;quot;Microsoft Corporation&amp;quot; (ORGANIZATION. Commercial).</Paragraph>
    <Paragraph position="2"> In this paper, we address the problem of relation extraction using kernel methods (Scholkopf and Smola, 2001). Many feature-based learning algorithms involve only the dot-product between feature vectors. Kernel methods can be regarded as a generalization of the feature-based methods by replacing the dot-product with a kernel function between two vectors, or even between two objects. A kernel function is a similarity function satisfying the properties of being symmetric and positive-definite. Recently, kernel methods are attracting more interests in the NLP study due to their ability of implicitly exploring huge amounts of structured features using the original representation of objects. For example, the kernels for structured natural language data, such as parse tree kernel (Collins and Duffy, 2001), string kernel (Lodhi et al., 2002) and graph kernel (Suzuki et al., 2003) are example instances of the well-known convolution kernels  in NLP. In relation extraction, typical work on kernel methods includes: Zelenko et al. (2003), Culotta and Sorensen (2004) and Bunescu and Mooney (2005).</Paragraph>
    <Paragraph position="3"> This paper presents a novel composite kernel to explore diverse knowledge for relation extraction. The composite kernel consists of an entity kernel and a convolution parse tree kernel. Our study demonstrates that the composite kernel is very effective for relation extraction. It also shows without the need for extensive feature engineering the composite kernel can not only capture most of the flat features used in the previous work but also exploit the useful syntactic structure features effectively. An advantage of our method is that the composite kernel can easily cover more knowledge by introducing more kernels. Evaluation on the ACE corpus shows that our method outperforms the previous best-reported methods and significantly outperforms the previous kernel methods due to its effective exploration of various syntactic features.</Paragraph>
    <Paragraph position="4"> The rest of the paper is organized as follows.</Paragraph>
    <Paragraph position="5"> In Section 2, we review the previous work. Section 3 discusses our composite kernel. Section 4 reports the experimental results and our observations. Section 5 compares our method with the  Convolution kernels were proposed for a discrete structure by Haussler (1999) in the machine learning field. This framework defines a kernel between input objects by applying convolution &amp;quot;sub-kernels&amp;quot; that are the kernels for the decompositions (parts) of the objects.</Paragraph>
    <Paragraph position="6">  previous work from the viewpoint of feature exploration. We conclude our work and indicate the future work in Section 6.</Paragraph>
  </Section>
class="xml-element"></Paper>