<?xml version="1.0" standalone="yes"?>
<Paper uid="W03-1007">
  <Title>Maximum Entropy Models for FrameNet Classification</Title>
  <Section position="5" start_page="5" end_page="5" type="concl">
    <SectionTitle>
5 Conclusion
</SectionTitle>
    <Paragraph position="0"> The results reported here show that ME models provide higher performance on frame element classification tasks, given both human and automatically identified frame element boundaries, than the linear interpolation models examined in previous work. We attribute this increase to the benefits of the ME framework itself, the incorporation of sentence-level syntactic patterns into our feature set, and the use of previous tag information to find the most probable sequence of roles for a sentence.</Paragraph>
    <Paragraph position="1"> But perhaps most striking in our results are the effects of varying training set size on the performance of the classification and identification models. While for classification, the learning curve appears to be still increasing with training set size, the learning curve for identification appears to have already begun to plateau. This suggests that while classification will continue to improve as the FrameNet database gets larger, increased performance on identification will rely on the development of more sophisticated models.</Paragraph>
    <Paragraph position="2"> In future work, we intend to apply the lessons learned here to the problem of frame element identification. Gildea and Jurafsky have shown that improvements in identification can be had by more closely integrating the task with classification (they report an F-Score of .719 using an integrated model). We are currently exploring a ME approach which integrates these two tasks under a tagging framework. Initial results show that significant improvements can be had using techniques similar to those described above.</Paragraph>
  </Section>
class="xml-element"></Paper>