<?xml version="1.0" standalone="yes"?>
<Paper uid="N03-2008">
  <Title>A Maximum Entropy Approach to FrameNet Tagging</Title>
  <Section position="5" start_page="3" end_page="3" type="concl">
    <SectionTitle>
4 Conclusion
</SectionTitle>
    <Paragraph position="0"> It is clear that using a tagging framework and syntactic patterns improves performance of the semantic classifier when features are extracted from either automatically generated parse trees or human annotations. The most striking result of these experiments, however, is the dramatic decrease in performance associated with using features extracted from a parse tree.</Paragraph>
    <Paragraph position="1"> This decrease in performance can be traced to at least two aspects of the automatic extraction process: noisy parser output and limited grammatical information.</Paragraph>
    <Paragraph position="2"> To compensate for noisy parser output, our current work is focusing on two strategies. First, we are looking at using shallower but more reliable methods for syntactic feature generation, such as part of speech tagging and text chunking, to either replace or augment the syntactic parser. Second, we are using ontological information, such as word classes and synonyms, in the hopes that semantic information may supplement the noisy syntactic information.</Paragraph>
    <Paragraph position="3"> The models trained on features extracted from parse trees do not have access to rich grammatical information. Following Gildea and Jurafsky (2000), automatic extraction of grammatical information here is limited to the governing category of a Noun Phrase.</Paragraph>
    <Paragraph position="4"> The FrameNet annotations, however, are much richer and include information about complements, modifiers, etc. We are looking at ways to include such information either by using alternative parsers (Hermjakob, 1997) or as a post processing task (Blaheta and Charniak, 2000). In future work, we will extend the strategies outlined here to incorporate Frame Element identification into our model. By treating semantic classification as a single tagging problem, we hope to create a unified, practical, and high performance system for Frame Element tagging.</Paragraph>
  </Section>
class="xml-element"></Paper>