<?xml version="1.0" standalone="yes"?>
<Paper uid="H93-1005">
  <Title>THE HCRC MAP TASK CORPUS: NATURAL DIALOGUE FOR SPEECH RECOGNITION</Title>
  <Section position="10" start_page="28" end_page="29" type="evalu">
    <SectionTitle>
6. PRELIMINARY RESULTS
</SectionTitle>
    <Paragraph position="0"/>
    <Section position="1" start_page="28" end_page="29" type="sub_section">
      <SectionTitle>
6.1. Second Mention
</SectionTitle>
      <Paragraph position="0"> The duration and/or (excerpted) intelligibility of different tokens of a word uttered by the same speaker have been shown to depend on the availability of information outwith the word's acoustic shape which might help listeners to recognize it. In the context of extended discourse, this means word tokens are less intelligible when they refer to Given entities.</Paragraph>
      <Paragraph position="1"> On the face of it, the tendency to produce degraded tokens where they are redundant seems wonderfully cooperative, in the Gricean sense of the term: when there is previous relevant material, intelligibility is reduced. The less intelligible repeated tokens are in fact helpful to listeners, for they make better prompts to earlier discourse material, either because they signal listeners to associate the word's meaning with some entity already established in a discourse model (Fowler and Housum, 1987) or because such stored information must be called into play for successful on-line word recognition (Bard et al., 1991).</Paragraph>
      <Paragraph position="2"> The difficulty is that degraded tokens are not restricted to contexts in which the listener can recover the conditioning informarion. Using the Map Task corpus, we have begun to investigate how far speakers' adjustment of intelligibility is egocentrically rather than cooperatively based, that is, how far the speaker's own relevant knowledge provides his/her model for what the listener knows.</Paragraph>
      <Paragraph position="3"> We have found the expected loss of intelligibility for excerpted second mentions as against both first mentions and citation forms. Interestingly, we found that the co-referential repetition effect found for monologue holds in dialogue: it doesn't matter who utters the word first. When it comes to the second mention of an entity either speaker may reduce intelligibility. This suggests that dialogue participants maintain a common record of textually evoked given entities.</Paragraph>
      <Paragraph position="4"> It would also appear that once an entity is textually evoked there is no further effect of visual information. That is, it doesn't matter whether the listener or speaker can see the object they're referring to.</Paragraph>
      <Paragraph position="5"> Also relevant is a significant intelligibility loss we found in mentions which are only 'second' for the speaker, because the relevant feature was first mentioned not in the current conversation, but in a previous conversation with a different listener.</Paragraph>
      <Paragraph position="6"> Thus on the basis of our investigations to date it would appear, somewhat surprisingly, that speakers reduce articulatory effort on a purely egocentric basis, without regard to listeners' ability to share the contextual conditioning this implies.</Paragraph>
      <Paragraph position="7"> Emea~ of New Item Introduction The Map Task corpus presents an excellent opportunity for examining how speakers introduce new items into a discourse.</Paragraph>
      <Paragraph position="8"> Moreover, because we can measure to overall communicative effectiveness of a conversation by reference to the accuracy of the resulting map, we can go further and attempt to assess the value of particular item introduction strategies.</Paragraph>
      <Paragraph position="9"> Definite versus indefinite article is almost certainly too simplistic a starting point for investigating this issue. This is born out by a tabulation of new item introduction over half  the corpus, as shown in Table 2.</Paragraph>
      <Paragraph position="10"> introductions are those in which the speaker &amp;quot;Right  tion givers (IG) and instruction followers (IF). of introductions per dialogue by form of introductions used by instruc-Overall definites and indefinites appeared with equal frequency.</Paragraph>
      <Paragraph position="11"> If we look at listener's responses to the introduction of items they don't have on their map, we see a significant correlation of informative responses (&amp;quot;I haven't got an extinct volcano&amp;quot;) with question introductions, but not with indefinite article usage as such. Also, using the accuracy of the route drawn as a measure of communicative efficacy, we found a significant correlation between use of question introductions by the IG and route accuracy. There is an independent correlation between informative IF responses and accuracy. See (Anderson &amp; Boyle, in press) for more details.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>