<?xml version="1.0" standalone="yes"?>
<Paper uid="P99-1047">
  <Title>A Decision-Based Approach to Rhetorical Parsing</Title>
  <Section position="9" start_page="371" end_page="371" type="concl">
    <SectionTitle>
8 Conclusion
</SectionTitle>
    <Paragraph position="0"> In this paper, we presented a shift-reduce rhetorical parsing algorithm that learns to construct rhetorical structures of texts from tagged data. The parser has two components: a discourse segmenter, which identifies the elementary discourse units in a text; and a shift-reduce action identifier, which determines how these units should be assembled into rhetorical structure trees.</Paragraph>
    <Paragraph position="1"> Our results suggest that a high-performance discourse segmenter would need to rely on more training data and more elaborate features than the ones described in this paper -- the learning curves did not converge to performance limits. If one's goal is, however, to construct discourse trees whose leaves are sentences (or units that can be identified at high levels of performance), then the segmenter described here appears to be adequate. Our results also suggest that the rich set of features that constitute the foundation of the action identifier are sufficient for constructing discourse hierarchies and for assigning to discourse segments a rhetorical status of nucleus or satellite at levels of performance that are close to those of humans. However, more research is needed in order to approach human performance in the task of assigning to segments correct rhetorical relation labels.</Paragraph>
    <Paragraph position="2"> Acknowledgements. I am grateful to Ulf Hermjakob, Kevin Knight, and Eric Breck for comments on previous drafts of this paper.</Paragraph>
  </Section>
class="xml-element"></Paper>