<?xml version="1.0" standalone="yes"?>
<Paper uid="J00-3003">
  <Title>Dialogue Act Modeling for Automatic Tagging and Recognition of Conversational Speech</Title>
  <Section position="14" start_page="367" end_page="368" type="concl">
    <SectionTitle>
9. Conclusions
</SectionTitle>
    <Paragraph position="0"> We have developed an integrated probabilistic approach to dialogue act modeling for conversational speech, and tested it on a large speech corpus. The approach combines models for lexical and prosodic realizations of DAs, as well as a statistical discourse 10 The inadequacy of n-gram models for nested discourse structures is pointed out by Chu-Carroll (1998), although the suggested solution is a modified n-gram approach.</Paragraph>
    <Paragraph position="1">  Stolcke et al. Dialogue Act Modeling grammar. All components of the model are automatically trained, and are thus applicable to other domains for which labeled data is available. Classification accuracies achieved so far are highly encouraging, relative to the inherent difficulty of the task as measured by human labeler performance. We investigated several modeling alternatives for the components of the model (backoff n-grams and maximum entropy models for discourse grammars, decision trees and neural networks for prosodic classification) and found performance largely independent of these choices. Finally, we developed a principled way of incorporating DA modeling into the probability model of a continuous speech recognizer, by constraining word hypotheses using the discourse context. However, the approach gives only a small reduction in word error on our corpus, which can be attributed to a preponderance of a single dialogue act type (statements). Note The research described here is based on a project at the 1997 Workshop on Innovative Techniques in LVCSR at the Center for Speech and Language Processing at Johns Hopkins University (Jurafsky et al. 1997; Jurafsky et al. 1998). The DA-labeled Switchboard transcripts as well as other project-related publications are available at http://www.colorado.</Paragraph>
    <Paragraph position="2"> edu/ling/jurafsky/ws97/.</Paragraph>
  </Section>
class="xml-element"></Paper>