<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-3401">
  <Title>Prosodic Correlates of Rhetorical Relations</Title>
  <Section position="2" start_page="0" end_page="0" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Rhetorical Structure Theory (RST) (Mann and Thompson, 1988) attempts to describe a given text in terms of its coherence, i.e. how it is that the parts of the text are related to one another and how each part plays a role. Two adjacent text spans will often exhibit a nucleus-satellite relationship, where the satellite plays a role that is relative to the nucleus.</Paragraph>
    <Paragraph position="1"> For example, one sentence might make a claim and the following sentence give evidence for the claim, with the second sentence being a satellite and the evidence relation existing between the two spans.</Paragraph>
    <Paragraph position="2"> In a text containing many sentences, these nucleus-satellite pairs can be built up to produce a document-wide rhetorical tree. Figure 1 gives an example of a rhetorical tree for a three-sentence text1.</Paragraph>
    <Paragraph position="3"> Theories such as RST have been popular for some time as a way of describing the multi-levelled rhetorical relations that exist in text, with relevant applications such as automatic summarization (Marcu, 1997) and natural language generation (Knott and Dale, 1996). However, implementing automatic rhetorical parsers has been a problematic area of research. Techniques that rely heavily on explicit signals, such as discourse markers, are of limited use both because only a small percentage of rhetorical relations are signalled explicitly and because explicit markers can be ambiguous. RST trees are binary branching trees distinguishing between nuclei and satellites, and automatically determining nuclearity is also far from trivial. Furthermore, there are some documents which are simply not amenable to being described by a document-wide rhetorical tree (Mann and Thompson, 1988). Finally, sometimes more than one relation can hold between two given units (Moore and Pollack, 1992). Given the problems of automatically parsing text for rhetorical relations, it seems prohibitively difficult to attempt rhetorical parsing of speech documents - data which are marked by disfluencies, low information density, and sometimes little cohesion. For that reason, this pilot study sets out a comparatively modest task: to determine whether one of five relations holds between two adjacent dialogue acts in meeting speech. All relations are of the form nucleussatellite, and the five relation types are contrast,</Paragraph>
    <Paragraph position="5"> work solely investigates the usefulness of prosodic features in classifying these five relations, rather than relying on discourse or lexical cues. A central motivation for this study is the hope that rhetorical parsing using prosodic features might aid an automatic summarization system.</Paragraph>
  </Section>
class="xml-element"></Paper>