<?xml version="1.0" standalone="yes"?>
<Paper uid="N06-2014">
  <Title>Agreement/Disagreement Classi cation: Exploiting Unlabeled Data using Contrast Classi ers</Title>
  <Section position="5" start_page="55" end_page="55" type="concl">
    <SectionTitle>
4 Conclusion
</SectionTitle>
    <Paragraph position="0"> In summary, our experiments on agreement/disagreement detection show that semi-supervised learning using contrast classi ers is an effective method for taking advantage of a large unlabeled data set for a problem with imbalanced classes. The contrast classi er approach outperforms co-training and self-training in detecting the infrequent classes. We also obtain good performance relative to other methods using simple lexical features and performance comparable to the best result reported.</Paragraph>
    <Paragraph position="1"> The experiments here kept the feature set xed, but results of (Galley et al., 2004) suggest that further gains can be achieved by augmenting the feature set. In addition, it is important to assess the impact of semi-supervised training with recognizer output, where gains from using unlabeled data may be greater than with reference transcripts as in (Hillard et al., 2003).</Paragraph>
  </Section>
class="xml-element"></Paper>