<?xml version="1.0" standalone="yes"?>
<Paper uid="P05-2014">
  <Title>Dialogue Act Tagging for Instant Messaging Chat Sessions</Title>
  <Section position="2" start_page="0" end_page="79" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Instant Messaging (IM) dialogue has received relatively little attention in discourse modelling. The novelty and popularity of IM dialogue and the significant differences between written and spoken English warrant specific research on IM dialogue.</Paragraph>
    <Paragraph position="1"> We show that IM dialogue has some unique problems and attributes not found in transcribed spoken dialogue, which has been the focus of most work in discourse modelling. The present study addresses the problems presented by these differences when modelling dialogue acts in IM dialogue.</Paragraph>
    <Paragraph position="2"> Stolcke et al. (2000) point out that the use of dialogue acts is a useful first level of analysis for describing discourse structure. Dialogue acts are based on the illocutionary force of an utterance from speech act theory, and represent acts such as assertions and declarations (Austin, 1962; Searle, 1979). This theory has been extended in dialogue acts to model the conversational functions that utterances can perform. Dialogue acts have been used to benefit tasks such as machine translation (Tanaka and Yokoo, 1999) and the automatic detection of dialogue games (Levin et al., 1999). This deeper level of discourse understanding may help replace or assist a support representative using IM dialogue by suggesting responses that are more sophisticated and realistic to a human dialogue participant.</Paragraph>
    <Paragraph position="3"> The unique problems and attributes exhibited by IM dialogue prohibit existing dialogue act classification methods from being applied directly. We present solutions to some of these problems along with methods to obtain high accuracy in automated dialogue act classification. A statistical discourse model is trained and then used to classify dialogue acts based on the observed words in an utterance.</Paragraph>
    <Paragraph position="4"> The training data are online conversations between two people: a customer and a shopping assistant, which we collected and manually annotated. Table 1 shows a sample of the type of dialogue and discourse structure used in this study.</Paragraph>
    <Paragraph position="5"> We begin by considering the preliminary issues that arise in IM dialogue, why they are problematic when modelling dialogue acts, and present their solutions inSS2. With the preliminary problems solved, we investigate the dialogue act labelling task with a description of our data inSS3. The remainder of the paper describes our experiment involving the training of a naive Bayes model combined with a n-gram discourse model (SS4). The results of this model and evaluation statistics are presented inSS5. SS6 contains a discussion of the approach we used including its strengths, areas of improvement, and issues for future research followed by the conclusion inSS7.</Paragraph>
    <Paragraph position="6">  finished. Here, message (&amp;quot;Msg&amp;quot;) 12 is actually in response to 10, not 11 since turn 6 was sent as 2 messages: 10 and 11. We use the seconds elapsed (&amp;quot;Sec&amp;quot;) since the previous message as part of a method to resynchronise messages. Utterance boundaries and their respective dialogue acts are denoted by Un.</Paragraph>
  </Section>
class="xml-element"></Paper>