<?xml version="1.0" standalone="yes"?>
<Paper uid="P05-1031">
  <Title>Towards Finding and Fixing Fragments: Using ML to Identify Non-Sentential Utterances and their Antecedents in Multi-Party Dialogue</Title>
  <Section position="9" start_page="252" end_page="252" type="relat">
    <SectionTitle>
5 Related Work
</SectionTitle>
    <Paragraph position="0"> To our knowledge, the tasks presented here have so far not been studied with a machine learning approach. The closest to our problem is (Fern'andez et al., 2004b), which discusses classifying certain types of fragments, namely questions of the type &amp;quot;Who?&amp;quot;, &amp;quot;When?&amp;quot;, etc. (sluices). However, that paper does not address the task of identifying those in a corpus (which in any case should be easier than our fragment-task, since those fragments cannot be confused with backchannels).</Paragraph>
    <Paragraph position="1"> Overlapping from another direction is the work presented in (Zechner and Lavie, 2001), where the task of aligning questions and answers is tackled.</Paragraph>
    <Paragraph position="2"> This subsumes the task of identifying questionantecedents for short-answers, but again is presumably somewhat simpler than our general task, because questions are easier to identify. The authors also evaluate the use of the alignment of questions and answers in a summarisation system, and report an increase in summary fluency, without a compromise in informativeness. This is something we hope to be able to show for our tasks as well.</Paragraph>
    <Paragraph position="3"> There are also similarities, especially of the antecedent task, to the pronoun resolution task (see e.g. (Strube and M&amp;quot;uller, 2003; Poesio et al., 2004)). Interestingly, our results for the antecedent task are close to those reported for that task. The problem of identifying the units in need of an antecedent, however, is harder for us, due to the problem of there being a large number of non-sentential utterances that cannot be linked to a single utterance as antecedent. In general, this seems to be the main difference between our task and the ones mentioned here, which concentrate on more easily identified markables (questions, sluices, and pronouns).</Paragraph>
  </Section>
class="xml-element"></Paper>