<?xml version="1.0" standalone="yes"?>
<Paper uid="W00-0103">
  <Title>Reducing Lexical Semantic Complexity with Systematic Polysemous Classes and Underspecification</Title>
  <Section position="1" start_page="0" end_page="14" type="abstr">
    <SectionTitle>
Abstract
</SectionTitle>
    <Paragraph position="0"> This paper presents an algorithm for finding systematic polysemous classes in WordNet and similar semantic databases, based on a definition in (Apresjan 1973). The introduction of systematic polysemous classes can reduce the amount of lexical semantic processing, because the number of disambiguation decisions can be restricted more clearly to those cases that involve real ambiguity (homonymy). In many applications, for instance in document categorization, information retrieval, and information extraction, it may be sufficient to know if a given word belongs to a certain class (underspecified sense) rather than to know which of its (related) senses exactly to pick. The approach for finding systematic polysemous classes is based on that of (Buitelaar 1998a, Buitelaar 1998b), while addressing some previous shortcomings.</Paragraph>
    <Paragraph position="1"> Introduction This paper presents an algorithm for finding systematic polysemous classes in WordNet (Miller et al 1990) and GermaNet (Hamp and Feldweg 1997) -- a semantic database for German similar to WordNet. The introduction of such classes can reduce the amount of lexical semantic processing, because the number of disambiguation decisions can be restricted more clearly to those cases that involve real ambiguity (homonymy). Different than with homonyms, systematically polysemous words need not always be disambiguated, because such words have several related senses that are shared in a systematic way by a group of similar words. In many applications then, for instance in document categorization and other areas of information retrieval, it may be sufficient to know if a given word belongs to this grou p rather than to know which of its (related) senses exactly to pick. In other words, it will suffice to assign a more coarse grained sense that leaves several related senses underspecified, but which can be further specified on demand 1.</Paragraph>
    <Paragraph position="2"> The approach for finding systematic polysemous classes is based on that of (Buitelaar 1998a, Buitelaar 1998b), but takes into account some shortcomings as pointed out in (Krymolowski and Roth 1998) (Peters, Peters and Vossen 1998) (Tomuro 1998). Whereas the original approach identified a small set of top-level synsets for grouping together lexical items, i As pointed out in (Wilks 99), earlier work in AI on 'Polaroid Words' (Hirst 87) and 'Word Experts' (Small 81) advocated a similar, incremental approach to sense representation and interpretation. In line with this, the CoreLex approach discussed here provides a large scale inventory of systematically polysemous lexical items with underspecified representations that can be incrementally refined.</Paragraph>
    <Paragraph position="3">  the new approach compares lexical items according to all of their synsets on all hierarchy levels. In addition, the new approach is both more flexible and precise by using a clustering algorithm for comparing meaning distributions between lexical items. Whereas the original approach took into account only identical distributions (with additional human intervention to further group together sufficiently similar classes), the clustering approach allows for completely automatic comparisons, relative to certain thresholds, that identify partial overlaps in meaning distributions.</Paragraph>
  </Section>
class="xml-element"></Paper>