<?xml version="1.0" standalone="yes"?>
<Paper uid="N06-2003">
  <Title>Museli: A Multi-Source Evidence Integration Approach to Topic Segmentation of Spontaneous Dialogue</Title>
  <Section position="4" start_page="9" end_page="9" type="metho">
    <SectionTitle>
3 Overview of Museli Approach
</SectionTitle>
    <Paragraph position="0"> We will demonstrate that lexical cohesion alone does not adequately mark topic boundaries in dialogue. Nevertheless, it can provide one meaningful source of evidence towards segmenting dialogue. In our hybrid Museli approach, we combined lexical cohesion with features that have the potential to capture something about the linguistic style that marks shifts in topic: word-unigrams, word-bigrams, and POS-bigrams for the current and previous contributions; the inclusion of at least one non-stopword term (contribution of content); time difference between contributions; contribution length; and the agent role of the previous and current contribution.</Paragraph>
    <Paragraph position="1"> We cast the segmentation problem as a binary classification problem where each contribution is classified as NEW_TOPIC if the contribution introduces a new topic and SAME_TOPIC otherwise. We found that using a Naive Bayes classifier (John &amp; Langley, 1995) with an attribute selection wrapper using the chi-square test for ranking attributes performed better than other state-of-the-art machine learning algorithms, perhaps because of the evidence integration oriented nature of the problem. We conducted our evaluation using 10fold cross-validation, being careful not to include instances from the same dialogue in both the training and test sets on any fold so that the results we report would not be biased by idiosyncratic communicative patterns associated with individual conversational participants picked up by the trained model.</Paragraph>
    <Paragraph position="2"> Using the complete set of features enumerated above, we perform feature selection on the training data for each fold of the cross-validation separately, training a model with the top 1000 features, and applying that trained model to the test data.</Paragraph>
    <Paragraph position="3"> Examples of high ranking features confirm our intuition that contributions that begin new topic segments are syntactically marked. For example, many typical selected word bigrams were indicative of imperatives, such as lets-do, do-the, ok-lets, ok-try, lets-see, etc. Others included time oriented discourse markers such as now, then, next, etc.</Paragraph>
    <Paragraph position="4"> To capitalize on differences in conversational behavior between participants assigned to different roles in the conversation (i.e., student and tutor in our evaluation corpora), we learn separate models for each role in the conversation  . This decision is based on the observation that participants with different agent-roles introduce topics with a different frequency, introduce different types of topics, and may introduce topics in a different style that displays their status in the conversation. For instance, a tutor may introduce new topics with a contribution that ends with an imperative. A student may introduce new topics with a contribution that ends with a wh-question.</Paragraph>
  </Section>
class="xml-element"></Paper>