<?xml version="1.0" standalone="yes"?>
<Paper uid="W04-2328">
  <Title>Multi-level Dialogue Act Tags</Title>
  <Section position="3" start_page="0" end_page="0" type="intro">
    <SectionTitle>
2 Understanding Dialogue Structure:
</SectionTitle>
    <Paragraph position="0"/>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
Dialogue Acts
2.1 The Concepts behind Dialogue Acts
</SectionTitle>
      <Paragraph position="0"> Dialogues are series of speaker turns. Utterances can be de ned as the atomic subparts of a turn that accomplish one or more \functions&amp;quot; with respect to speaker interaction. Utterances are in general signalled by syntactic and/or prosodic means, but the speci city of their \function&amp;quot; belongs to pragmatics (Levinson, 1983, ch. 4). Linguists have identi ed several dimensions for the role of sentences uttered in a dialogue. These dimensions are not mutually exclusive, and there are certainly correlations between some of them (e.g. \question&amp;quot; as a speech act and as a member of an adjacency pair).</Paragraph>
      <Paragraph position="1"> Speech acts (Searle, 1969; Vanderveken, 1990): (1) representatives, such as assertions or conclusions; (2) directives, such as requests, questions, suggestions; (3) commissives, such as promises, threatenings, o ers; (4) expressives such as thanks, apologies, congratulations; (5) declarations, such as excommunications, declarations of war, christening, ring from employment, etc.</Paragraph>
      <Paragraph position="2"> Turn management: backchannel, oor holder, oor grabber, hold; Adjacency pairs: utterances can be the rst part or the second part of exchange pairs such as request / accept (or refuse); o er / accept; assess / (dis)agree; question / answer; etc.</Paragraph>
      <Paragraph position="3"> Overall organization and topics: openings, closings, topic-changers, topic-continuers, etc.</Paragraph>
      <Paragraph position="4"> Politeness management: face-threatening, facesaving, neutral; Rhetorical role: elaboration, purpose, restatement, etc.</Paragraph>
    </Section>
    <Section position="2" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
2.2 Dialogue Acts in Computational
Linguistics
</SectionTitle>
      <Paragraph position="0"> There is not much agreement, within the CL/NLP community, on the de nition of a dialogue act. The term denotes some function of an utterance in a dialogue, not reducible to its syntactic or semantic content. The function is selected, in general, among a set of possible dialogue acts (a DA tagset) that depends on the goals of its creator (Traum, 2000).</Paragraph>
      <Paragraph position="1"> One of the main inspiration sources for DA tagsets are speech acts, but the original repertoire (Searle, 1969; Vanderveken, 1990) has been gradually enriched with other possible functions. From the numerous DA tagsets (Klein and Soria, 1998), the following are particularly relevant to a general-domain meeting recording application.</Paragraph>
      <Paragraph position="2"> The DA tags in DAMSL (Allen and Core, 1997) are nearly all independent: the DAMSL guidelines state that all tags (i.e. all \functions&amp;quot;) that characterize an utterance should be associated with it. The DAMSL tags are grouped in four dimensions: communicative status, information level, forward-looking function and backward-looking function. In fact, several theories are con ated in DAMSL, which was initially designed as a shared resource with a focus primarily on task-oriented dialogs (Core and Allen, 1997). There are about 4 million possible combinations of DAMSL tags, which make a huge search space for automatic annotation.</Paragraph>
      <Paragraph position="3"> The application of DAMSL to the Switchboard data (two-party telephone conversations) lead to SWBD-DAMSL (Jurafsky et al., 1997), a smaller tagset than DAMSL. About 200,000 SWBD utterances were rst annotated with DAMSL tags: it was observed that only 220 combinations of tags occurred (Jurafsky et al., 1998). These 220 labels were then clustered into 42 tags, such as: statement (36%), opinion (13%), agree/accept (5%), yes-no-question (2%). The resulting search space (42 mutually exclusive tags) was well adapted to the initial goals, viz., the automatic annotation of dialogue acts and the use of dialogue act speci c language models in speech recognition (Stolcke et al., 2000).</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>