<?xml version="1.0" standalone="yes"?>
<Paper uid="C04-1048">
  <Title>Generating Discourse Structures for Written Texts</Title>
  <Section position="7" start_page="0" end_page="0" type="concl">
    <SectionTitle>
5 Conclusions
</SectionTitle>
    <Paragraph position="0"> We have presented a discourse parser and evaluated it using the RST corpus. The presented discourse parser is divided into two levels: sentence-level and text-level. The experiment showed that syntactic information and cue phrases are quite effective in constructing discourse structures at the sentence-level, especially in discourse segmentation (86.9% F-score). The discourse trees at the text-level were generated by combining the hypothesized discourse relations among nonoverlapped text spans. We concentrated on solving the combinatorial explosion in searching for discourse trees. The constraints of textual adjacency and textual organization, and a beam search were applied to find the best-quality trees in a search space that is much smaller than the one given by Marcu (2000). The experiment on documents from the RST corpus showed that the proposed approach could produce reasonable results compared to human annotator agreements.</Paragraph>
    <Paragraph position="1"> To improve the system performance, future work includes refining the segmentation rules and improving criteria to select optimal paths in the beam search. We would also like to integrate a syntactic parser to this system. We hope this research will aid the development of text processing such as text summarization and text generation.</Paragraph>
  </Section>
class="xml-element"></Paper>