<?xml version="1.0" standalone="yes"?>
<Paper uid="N06-1036">
  <Title>Backoff Model Training using Partially Observed Data: Application to Dialog Act Tagging</Title>
  <Section position="6" start_page="286" end_page="286" type="concl">
    <SectionTitle>
5 Conclusions
</SectionTitle>
    <Paragraph position="0"> In this work, we introduced a training method for hidden backoff models (HBMs) to solve a problem in DA tagging where smoothed backoff models involving training-time hidden variables are useful.</Paragraph>
    <Paragraph position="1"> We tested this procedure in the context of dynamic Bayesian networks. Different hidden states were used to model different positions in a DA. According to empirical evaluations, our embedded EM algorithm effectively increases log likelihood on training data and reduces DA tagging error rate on test data.</Paragraph>
    <Paragraph position="2"> If different numbers of hidden states are used for different DAs, we nd that our prosody-independent HBM reduces the tagging error rate by 6.1% relative to the baseline, a result that improves upon previously reported work that uses prosody, and that is comparable to our own new result that also incorporates prosody. We have not yet been able to combine the bene ts of both an HBM and prosody information. This material is based upon work supported by the National Science Foundation under Grant No.</Paragraph>
    <Paragraph position="3"> IIS-0121396.</Paragraph>
  </Section>
class="xml-element"></Paper>