<?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="4" start_page="825" end_page="825" type="relat">
    <SectionTitle>
2 Related Work
</SectionTitle>
    <Paragraph position="0"> Many techniques on relation extraction, such as rule-based (MUC, 1987-1998; Miller et al., 2000), feature-based (Kambhatla 2004; Zhou et al., 2005) and kernel-based (Zelenko et al., 2003; Culotta and Sorensen, 2004; Bunescu and Mooney, 2005), have been proposed in the literature. null Rule-based methods for this task employ a number of linguistic rules to capture various relation patterns. Miller et al. (2000) addressed the task from the syntactic parsing viewpoint and integrated various tasks such as POS tagging, NE tagging, syntactic parsing, template extraction and relation extraction using a generative model. Feature-based methods (Kambhatla, 2004; Zhou et al., 2005; Zhao and Grishman, 2005  ) for this task employ a large amount of diverse linguistic features, such as lexical, syntactic and semantic features. These methods are very effective for relation extraction and show the best-reported performance on the ACE corpus. However, the problems are that these diverse features have to be manually calibrated and the hierarchical structured information in a parse tree is not well preserved in their parse tree-related features, which only represent simple flat path information connecting two entities in the parse tree through a path of non-terminals and a list of base phrase chunks.</Paragraph>
    <Paragraph position="1"> Prior kernel-based methods for this task focus on using individual tree kernels to exploit tree structure-related features. Zelenko et al. (2003) developed a kernel over parse trees for relation extraction. The kernel matches nodes from roots to leaf nodes recursively layer by layer in a top-down manner. Culotta and Sorensen (2004) generalized it to estimate similarity between dependency trees. Their tree kernels require the matchable nodes to be at the same layer counting from the root and to have an identical path of ascending nodes from the roots to the current nodes. The two constraints make their kernel high precision but very low recall on the ACE 2003 corpus. Bunescu and Mooney (2005) proposed another dependency tree kernel for relation extraction.</Paragraph>
    <Paragraph position="2">  We classify the feature-based kernel defined in (Zhao and Grishman, 2005) into the feature-based methods since their kernels can be easily represented by the dot-products between explicit feature vectors.</Paragraph>
    <Paragraph position="3"> Their kernel simply counts the number of common word classes at each position in the shortest paths between two entities in dependency trees.</Paragraph>
    <Paragraph position="4"> The kernel requires the two paths to have the same length; otherwise the kernel value is zero.</Paragraph>
    <Paragraph position="5"> Therefore, although this kernel shows performance improvement over the previous one (Culotta and Sorensen, 2004), the constraint makes the two dependency kernels share the similar behavior: good precision but much lower recall on the ACE corpus.</Paragraph>
    <Paragraph position="6"> The above discussion shows that, although kernel methods can explore the huge amounts of implicit (structured) features, until now the feature-based methods enjoy more success. One may ask: how can we make full use of the nice properties of kernel methods and define an effective kernel for relation extraction? In this paper, we study how relation extraction can benefit from the elegant properties of kernel methods: 1) implicitly exploring (structured) features in a high dimensional space; and 2) the nice mathematical properties, for example, the sum, product, normalization and polynomial expansion of existing kernels is a valid kernel (Scholkopf and Smola, 2001). We also demonstrate how our composite kernel effectively captures the diverse knowledge for relation extraction. null</Paragraph>
  </Section>
class="xml-element"></Paper>