<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-1101">
  <Title>Linguistic Distances</Title>
  <Section position="6" start_page="2" end_page="3" type="concl">
    <SectionTitle>
4 Semantics
</SectionTitle>
    <Paragraph position="0"> While similarity as such has not been a prominent term in theoretical and computational research on natural language semantics, the study of LEXICAL SEMANTICS, which attempts to identify regularities of and systematic relations among word meanings, is more often than not predicated on an implicit notion of 'semantic similarity'. Research on the lexical semantics of verbs tries to identify verb classes whose members exhibit similar syntactic and semantic behavior. In logic-based theories of word meaning (e.g., Vendler (1967) and Dowty (1979)), verb classes are identified by similarity patterns of inference, while Levin's (1993) study of English verb classes demonstrates that similarities of word meanings for verbs can be gleaned from their syntactic behavior, in particular from their ability or inability to participate in diatheses, i.e. patterns of argument alternations.</Paragraph>
    <Paragraph position="1"> With the increasing availability of large electronic corpora, recent computational research on word meaning has focused on capturing the notion of 'context similarity' of words. Such studies follow the empiricist approach to word meaning summarized best in the famous dictum of the British  linguist J.R. Firth: &amp;quot;You shall know a word by the company it keeps.&amp;quot; (Firth, 1957, p. 11) Context similarity has been used as a means of extracting collocations from corpora, e.g. by Church &amp; Hanks (1990) and by Dunning (1993), of identifying word senses, e.g. by Yarowski (1995) and by Sch&amp;quot;utze (1998), of clustering verb classes, e.g. by Schulte im Walde (2003), and of inducing selectional restrictions of verbs, e.g. by Resnik (1993), by Abe &amp; Li (1996), by Rooth et al. (1999) and by Wagner (2004).</Paragraph>
    <Paragraph position="2"> A third approach to lexical semantics, developed by linguists and by cognitive psychologists, primarily relies on the intuition of lexicographers for capturing word meanings, but is also informed by corpus evidence for determining word usage and word senses. This type of approach has led to two highly valued semantic resources: the Princeton WordNet (Fellbaum, 1998) and the Berkeley Framenet (Baker et al., 1998). While originally developed for English, both approaches have been successfully generalized to other languages.</Paragraph>
    <Paragraph position="3"> The three approaches to word meaning discussed above try to capture different aspects of the notion of semantic similarity, all of which are highly relevant for current and future research in computational linguistics. In fact, the five papers that discuss issues of semantic similarity in the present volume build on insights from these three frameworks or address open research questions posed by these frameworks. Zesch and Gurevych (this volume) discuss how measures of semantic similarity--and more generally: semantic relatedness--can be obtained by similarity judgments of informants who are presented with word pairs and who, for each pair, are asked to rate the degree of semantic relatedness on a pre-defined scale. Such similarity judgments can provide important empirical evidence for taxonomic models of word meanings such as wordnets, which thus far rely mostly on expert knowledge of lexicographers. To this end, Zesch and Gurevych propose a corpus-based system that supports fast development of relevant data sets for large subject domains.</Paragraph>
    <Paragraph position="4"> St-Jacques and Barri`ere (this volume) review and contrast different philosophical and psychological models for capturing the notion of semantic similarity and different mathematical models for measuring semantic distance. They draw attention to the fact that, depending on which underlying models are in use, different notions of semantic similarity emerge and conjecture that different similarity metrics may be needed for different NLP tasks. Dagan (this volume) also explores the idea that different notions of semantic similarity are needed when dealing with semantic disambiguation and language modeling tasks on the one hand and with applications such as information extraction, summarization, and information retrieval on the other hand.</Paragraph>
    <Paragraph position="5"> Dridan and Bond (this volume) and Hachey (this volume) both consider semantic similarity from an application-oriented perspective. Dridan and Bond employ the framework of robust minimal recursion semantics in order to obtain a more adequate measure of sentence similarity than can be obtained by word-overlap metrics for bag-of-words representations of sentences. They show that such a more fine-grained measure, which is based on compact representations of predicate-logic, yields better performance for paraphrase detection as well as for sentence selection in question-answering tasks than simple word-overlap metrics. Hachey considers an automatic content extraction (ACE) task, a particular subtask of information extraction. He demonstrates that representations based on term co-occurrence outperform representations based on term-by-document matrices for the task of identifying relationships between named objects in texts.</Paragraph>
  </Section>
class="xml-element"></Paper>