<?xml version="1.0" standalone="yes"?>
<Paper uid="P02-1047">
  <Title>An Unsupervised Approach to Recognizing Discourse Relations</Title>
  <Section position="7" start_page="0" end_page="0" type="evalu">
    <SectionTitle>
CONTRAST and CAUSE-EXPLANATION-EVIDENCE
</SectionTitle>
    <Paragraph position="0"> relations, as defined in RST, but not so well between ELABORATION and any other relation. This result is consistent with the discourse model proposed by Knott et al. (2001), who suggest that ELABORATION relations are too ill-defined to be part of any discourse theory.</Paragraph>
    <Paragraph position="1"> The analysis above is informative only from a machine learning perspective. From a linguistic perspective though, this analysis is not very useful. If no cue phrases are used to signal the relation between two elementary discourse units, an automatic discourse labeler can at best guess that an ELABORATION relation holds between the units, because ELABORATION relations are the most frequently used relations (Carlson et al., 2001). Fortunately, with the classifiers described here, one can label some of the unmarked discourse relations correctly. null For example, the RST-annotated corpus of Carlson et al. (2001) contains 238 CONTRAST relations that hold between two adjacent elementary discourse units. Of these, only 61 are marked by a cue phrase, which means that a program trained only on Carlson et al.'s corpus could identify at most 61/238 of the CONTRAST relations correctly. Because Carlson et al.'s corpus is small, all unmarked relations will be likely labeled as ELABORATIONs.</Paragraph>
    <Paragraph position="2"> However, when we run our CONTRAST vs. ELABORATION classifier on these examples, we can label correctly 60 of the 61 cue-phrase marked relations and, in addition, we can also label 123 of the 177 relations that are not marked explicitly with cue phrases. This means that our classifier contributes to an increase in accuracy from a91a55a92a14a93a69a94a64a95a64a96a17a97 a94a64a91a63a98 to a23a80a91a69a99a100a65a40a92a52a94a64a95a63a30a7a93a69a94a64a95a64a96a86a97a102a101a64a101a64a98 !!! Similarly, out of the 307 CAUSE-EXPLANATION-EVIDENCE relations that hold between two discourse units in Carlson et al.'s corpus, only 79 are explicitly marked.</Paragraph>
    <Paragraph position="3"> A program trained only on Carlson et al.'s corpus, would, therefore, identify at most 79 of the 307 relations correctly. When we run our CAUSE-EXPLANATION-EVIDENCE vs. ELABORATION classifier on these examples, we labeled correctly 73 of the 79 cue-phrase-marked relations and 102 of the 228 unmarked relations. This corresponds to an increase in accuracy from a101a69a103a63a93a69a95a69a99a43a101a104a97 a94a64a91a63a98 to a23a62a101a69a95a85a65a86a92a10a99a63a94a63a30a7a93a69a95a69a99a43a101a105a97a107a106a63a101a64a98 .</Paragraph>
  </Section>
class="xml-element"></Paper>