<?xml version="1.0" standalone="yes"?>
<Paper uid="C04-1199">
  <Title>Learning to Identify Single-Snippet Answers to Definition Questions</Title>
  <Section position="7" start_page="4" end_page="4" type="concl">
    <SectionTitle>
6 Conclusions and future work
</SectionTitle>
    <Paragraph position="0"> We have presented a new method to identify single-snippet definitions in question answering systems. Our method combines previously proposed techniques as attributes of an SVM learner, to which an automatic pattern acquisition process contributes additional attributes. We have evaluated several configurations of our method on TREC data, with results indicating it outperforms previous techniques.</Paragraph>
    <Paragraph position="1"> The performance of our method may improve if n-grams that start or end within a margin of a few tokens from the term to define are added. This may allow definitions like &amp;quot;X, that Y defined as ...&amp;quot; to be found. Further improvements may be possible by using a sentence splitter instead of windows of fixed length, anaphora resolution, clustering of similar snippets to avoid ranking them separately, and identifying additional n-gram attributes by bootstrapping (Ravichandran et al. 2003).</Paragraph>
    <Paragraph position="2"> We believe that it is possible to address the post-2003 TREC task for definition questions with the same approach, but training the SVM learner to identify snippets that should be included in multi-snippet definitions. With sufficient training, we expect that n-grams indicative of information commonly included in multi-snippet definitions (e.g., dates of birth, important works for persons) will be discovered. Larger amounts of training data, however, will be required. We are currently working on a method to generate training examples in an unsupervised manner from parallel texts.</Paragraph>
  </Section>
class="xml-element"></Paper>