<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-2503">
  <Title>Relating WordNet Senses for Word Sense Disambiguation</Title>
  <Section position="3" start_page="17" end_page="18" type="intro">
    <SectionTitle>
WNs# SEGR gloss
</SectionTitle>
    <Paragraph position="0"> lated. Since only a fixed number of senses are defined for each word, the RLISTs include all senses of the word. A cut-off can then be determined for any particular application.</Paragraph>
    <Paragraph position="1"> Previous research on clustering word senses has focused on comparison to the SEGR goldstandard. We evaluate the RLISTs against a new gold-standard produced by humans for this research since the SEGR does not have documentation with figures for inter-tagger agreement. As well as evaluating against a gold-standard, we also look at the effect of the RLISTs and the gold-standards themselves on WSD. Since the focus of this paper is not the WSD system, but the sense inventory, we use a simple WSD heuristic which uses the first sense of a word in all contexts, where the first sense of every word is specified by a resource. While contextual evidence is required for accurate WSD, it is useful to look at this heuristic since it is so widely used as a back-off model by many systems and is hard to beat on an all-words task (Snyder and Palmer, 2004). We contrast the performance of first sense heuristics i) from SemCor (Miller et al., 1993) and ii) derived automatically from the BNC following (McCarthy etal., 2004) and also iii) anupper-bound firstsense heuristic extracted from the test data.</Paragraph>
    <Paragraph position="2"> The paper is organised as follows. In the next section we describe some related work. In section 3 we describe the two methods we will use to relate senses. Our experiments are described in section 4. In 4.1 we describe the construction of a new gold-standard produced using the same sense inventory used for SEGR, and give inter-annotator agreement figures for the task. In section 4.2 we compare our methods to the new gold-standard and in section 4.3 we investigate how much effect coarser grained sense distinctions have on WSD using naive first sense heuristics. Wefollow this with a discussion and conclusion.</Paragraph>
  </Section>
class="xml-element"></Paper>