<?xml version="1.0" standalone="yes"?>
<Paper uid="N03-2011">
  <Title>Rhetorical Parsing with Underspecification and Forests</Title>
  <Section position="2" start_page="0" end_page="0" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> The task of rhetorical parsing, i.e., automatically determining discourse structure, has been shown to be relevant, inter alia, for automatic summarization (e.g., Marcu, 2000). Not surprisingly, though, the task is very difficult. Previous approaches have thus emphasized the need for heuristic or probabilistic information in the process of finding the best or most likely rhetorical tree.</Paragraph>
    <Paragraph position="1"> As an alternative, we explore the idea of strictly separating &amp;quot;high-confidence&amp;quot; information from hypothetical reasoning and of working with underspecified trees as much as possible. We create a parse forest on the basis of surface cues found in the text. This forest can then be subject to further processing. Depending on the application, such further steps can either calculate the &amp;quot;best&amp;quot; tree out of the forest or continue working with a set of structured hypotheses.</Paragraph>
    <Paragraph position="2"> Section 2 briefly summarizes our proposal on under-specified rhetorical trees; section 3 introduces our grammar approach to text structure; section 4 compares this strategy to earlier work.</Paragraph>
  </Section>
class="xml-element"></Paper>