<?xml version="1.0" standalone="yes"?>
<Paper uid="N03-2012">
  <Title>DETECTION OF AGREEMENT vs. DISAGREEMENT IN MEETINGS: TRAINING WITH UNLABELED DATA</Title>
  <Section position="6" start_page="0" end_page="0" type="concl">
    <SectionTitle>
4 Conclusion
</SectionTitle>
    <Paragraph position="0"> In summary, we have described an approach for automatic recognition of agreement and disagreement in meeting data, using both prosodic and word-based features. The methods can be implemented with a small amount of hand-labeled data by using unsupervised LM clustering to label additional data, which leads to significant gains in both word-based and prosody-based classifiers. The approach is extensible to other types of speech acts, and is especially important for domains in which very little annotated data exists. Even operating on ASR transcripts with high WERs (45%), we obtain a 78% rate of recovery of agreements and disagreements, with a very low rate of confusion between these classes. Prosodic features alone provide results almost as good as the word-based models on ASR transcripts, but no additional benefit when used with word-based features. However, the good performance from prosody alone offers hope for performance gains given a richer set of speech acts with more lexically ambiguous cases (Bhagat et al., 2003).</Paragraph>
  </Section>
class="xml-element"></Paper>