<?xml version="1.0" standalone="yes"?>
<Paper uid="I05-6007">
  <Title>Syntactic Identification of Attribution in the RST Treebank</Title>
  <Section position="5" start_page="59" end_page="59" type="relat">
    <SectionTitle>
5 Related Work
</SectionTitle>
    <Paragraph position="0"> Soricut and Marcu (2003) describe a Discourse Parser - a system that uses Penn Treebank syntax to identify intra-sentential discourse relations in the RST Treebank. Since this applies to all intra-sentential relations in the RST Treebank, while our system is limited to Attribution, the systems are not directly comparable. Still, the results and discussion from (Soricut and Marcu, 2003) provide some useful perspective on our results.</Paragraph>
    <Paragraph position="1"> Soricut and Marcu (2003) evaluate their Discourse Parser under a variety of scenarios; the most favorable has human-corrected syntax trees and discourse segmentation. In this scenario, the system achieves an f-score of .703 with the full set of 110 Relation Labels, and 75.5 with the relation label set collapsed to 18 labels. Soricut and Marcu (2003) note that human annotator agreement receives comparable f-scores, of .719 and .77 respectively. In the light of these numbers, our Attribution system f-score of .92 is quite impressive. This provides some measure of support for our hypothesis that the intra-sentential relations in the RST Treebank are in fact properly viewed as alternative notations for syntactic information that is already present in the Penn Treebank.</Paragraph>
    <Paragraph position="2"> Of course, it may well be that some of the other intra-sentential relations present much greater difficulties than Attribution. But these results suggest that it is worth pursuing our project of attempting to automatically derive the intra-sentential RST Treebank relations from specific syntactic features.</Paragraph>
  </Section>
class="xml-element"></Paper>