<?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="9" start_page="831" end_page="831" type="concl">
    <SectionTitle>
6 Conclusion and Future Work
</SectionTitle>
    <Paragraph position="0"> Kernel functions have nice properties. In this paper, we have designed a composite kernel for relation extraction. Benefiting from the nice properties of the kernel methods, the composite kernel could well explore and combine the flat entity features and the structured syntactic features, and therefore outperforms previous best-reported feature-based methods on the ACE corpus. To our knowledge, this is the first research to demonstrate that, without the need for extensive feature engineering, an individual tree kernel achieves comparable performance with the feature-based methods. This shows that the syntactic features embedded in a parse tree are particularly useful for relation extraction and which can be well captured by the parse tree kernel. In addition, we find that the relation instance representation (selecting effective portions of parse trees for kernel calculations) is very important for relation extraction.</Paragraph>
    <Paragraph position="1"> The most immediate extension of our work is to improve the accuracy of relation detection.</Paragraph>
    <Paragraph position="2"> This can be done by capturing more features by including more individual kernels, such as the WordNet-based semantic kernel (Basili et al., 2005) and other feature-based kernels. We can also benefit from machine learning algorithms to study how to solve the data imbalance and sparseness issues from the learning algorithm viewpoint. In the future work, we will design a more flexible tree kernel for more accurate similarity measure.</Paragraph>
    <Paragraph position="3"> Acknowledgements: We would like to thank Dr. Alessandro Moschitti for his great help in using his Tree Kernel Toolkits and fine-tuning the system. We also would like to thank the three anonymous reviewers for their invaluable suggestions. null</Paragraph>
  </Section>
class="xml-element"></Paper>