<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-3402">
  <Title>Off-Topic Detection in Conversational Telephone Speech</Title>
  <Section position="4" start_page="8" end_page="8" type="relat">
    <SectionTitle>
3 Related Work
</SectionTitle>
    <Paragraph position="0"> Three areas of related work in natural language processing have been particularly informative for our research.</Paragraph>
    <Paragraph position="1"> First, speech act theory states that with each utterance, a conversant is committing an action, such as questioning, critiquing, or stating a fact. This is quite similar to the notion of transactional and interactional goals. However, speech acts are generally focused on the lower level of breaking apart utterances and understanding their purpose, whereas we are concerned here with a coarser-grained notion of relevance. Work closer to ours is that of Bates et al. (Bates et al., 2005), who define meeting acts for recorded meetings. Of their tags, commentary is most similar to our notion of relevance.</Paragraph>
    <Paragraph position="2"> Second, there has been research on generating small talk in order to establish rapport between an automatic system and human user (Bickmore and Cassell, 2000). Our work complements this by potentially detecting off-topic speech from the human user as an indication that the system should also respond with interactional language.</Paragraph>
    <Section position="1" start_page="8" end_page="8" type="sub_section">
      <SectionTitle>
Label Utterance
</SectionTitle>
      <Paragraph position="0"> S 2: [LAUGH] Hi.</Paragraph>
      <Paragraph position="1"> S 2: How nice to meet you. S 1: It is nice to meet you too. M 2: We have a wonderful topic. M 1: Yeah.</Paragraph>
      <Paragraph position="2"> M 1: It's not too bad. [LAUGH] T 2: Oh, I - I am one hundred percent in favor of, uh, computers in the classroom. T 2: I think they're a marvelous tool, educational tool.</Paragraph>
      <Paragraph position="3"> Table 1: A conversation fragment with annotations: (S)mall Talk, (M)etaconversation, and On-(T)opic.  The two speakers are identified as &amp;quot;1&amp;quot; and &amp;quot;2&amp;quot;. Third, off-topic detection can be viewed as a segmentation of conversation into relevant and irrelevant parts. Thus our work has many similarities to topic segmentation systems, which incorporate cue words that indicate an abrupt change in topic (e.g. &amp;quot;so anyway...&amp;quot;), as well as long term variations in word occurrence statistics (Hearst, 1997; Reynar, 1999; Beeferman et al., 1999, e.g.). Our approach uses previous and subsequent sentences to approximate these ideas, but might benefit from a more explicitly segmentation-based strategy.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>