<?xml version="1.0" standalone="yes"?>
<Paper uid="N06-2034">
  <Title>Using Phrasal Patterns to Identify Discourse Relations</Title>
  <Section position="2" start_page="0" end_page="0" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Identifying discourse relations is important for many applications, such as text/conversation understanding, single/multi-document summarization and question answering. (Marcu and Echihabi 2002) proposed a method to identify discourse relations between text segments using Naive Bayes classifiers trained on a huge corpus.</Paragraph>
    <Paragraph position="1"> They showed that lexical pair information extracted from massive amounts of data can have a major impact.</Paragraph>
    <Paragraph position="2"> We developed a system which identifies the discourse relation between two successive sentences in Japanese. On top of the lexical information previously proposed, we added phrasal pattern information. A phrasal pattern includes at least three phrases (bunsetsu segments) from two sentences, where function words are mandatory and content words are optional. For example, if the first sentence is &amp;quot;X should have done Y&amp;quot; and the second sentence is &amp;quot;A did B&amp;quot;, then we found it very likely that the discourse relation is CONTRAST (89% in our Japanese corpus).</Paragraph>
  </Section>
class="xml-element"></Paper>