<?xml version="1.0" standalone="yes"?>
<Paper uid="W05-0613">
  <Title>Probabilistic Head-Driven Parsing for Discourse Structure</Title>
  <Section position="8" start_page="157" end_page="157" type="concl">
    <SectionTitle>
7 Conclusion
</SectionTitle>
    <Paragraph position="0"> In this paper, we have shown how the complex task of creating structures for SDRT can be adapted to a standard probabilistic parsing task. This is achieved via a headed tree representation from which SDRSs can be recovered. This enables us to directly apply well-known probabilistic parsing algorithms and use features inspired by them. Our results show that using dialogue-based features are a major factor in improving the performance of the models, both in terms of determining segmentation appropriately and choosing the right relations to connect them.</Paragraph>
    <Paragraph position="1"> There is clearly a great deal of room for improvement, even with our best model. Even so, that model performed suf ciently well for use in semi-automated annotation: when correcting the model's output on ten dialogues, one annotator took 30 seconds per utterance, compared to 39 for another annotator working on the same dialogues with no aid.</Paragraph>
    <Paragraph position="2"> In future work, we intend to exploit an existing implementation of SDRT's semantics (Schlangen and Lascarides, 2002), which adopts theorem proving to infer resolutions of temporal anaphora and communicative goals from SDRSs for scheduling dialogues. This additional semantic content can in turn be added (semi-automatically) to a training corpus. This will provide further features for learning discourse structure and opportunities for learning anaphora and goal information directly.</Paragraph>
  </Section>
class="xml-element"></Paper>