<?xml version="1.0" standalone="yes"?>
<Paper uid="W03-0412">
  <Title>PhraseNet: Towards Context Sensitive Lexical Semantics/</Title>
  <Section position="6" start_page="0" end_page="0" type="relat">
    <SectionTitle>
5 Related Work
</SectionTitle>
    <Paragraph position="0"> In this section we point to some of the related work on syntax, semantics interaction and lexical semantic resources in computational linguistics and natural language processing. Many current syntactic theories make the common assumption that various aspects of syntactic alternation are predicable via the meaning of the predicate in the sentence (Fillmore, 1968; Jackendoff, 1990; Levin, 1993). With the resurgence of lexical semantics and corpus linguistics during the past two decades, this so-called linking regularity triggers a broad interest of using syntactic representations illustrated in corpora to classify lexical meaning (Baker et al., 1998; Levin, 1993; Dorr and Jones, 1996; Lapata and Brew, 1999; Lin, 1998b; Pantel and Lin, 2002).</Paragraph>
    <Paragraph position="1"> FrameNet (Baker et al., 1998) produces a semantic dictionary that documents combinatorial properties of English lexical items in semantic and syntactic terms based on attestations in a very large corpus. In FrameNet, a frame is an intuitive structure that formalizes the links between semantics and syntax in the results of lexical analysis. (Fillmore et al., 2001) However, instead of derived via attested sentences from corpora automatically, each conceptual frame together with all its frame elements has to be constructed via slow and labor-intensive manual work. FrameNet is not constructed automatically based on observed syntactic alternations. Though deep semantic analysis is built for each frame, lack of automatic derivation of the semantic roles from large corpora3 confines the usage of this network drastically.</Paragraph>
    <Paragraph position="2"> Levin's classes (Levin, 1993) of verbs are based on the assumption that the semantics of a verb and its syntactic behavior are predictably related. She defines 191 verb classes by grouping 4183 verbs which pattern together with respect to their diathesis alternations, namely alternations in the expressions of arguments. In Levin's classification, it is the syntactic skeletons (such as np-v-nppp)to classify verbs directly. Levin's classification is validated via experiments done by (Dorr and Jones, 1996) and some counter-arguments are in (Baker and Ruppenhofer, 2002). Her work provides a a small knowledge source that needs further expansion.</Paragraph>
    <Paragraph position="3"> Lin's work (Lin, 1998b; Pantel and Lin, 2002) makes use of distributional syntactic contextual information to define semantic proximity. Dekang Lin's grouping of similar words is a combination of the abstract syntactic skeleton and concrete word tokens. Lin uses syntactic dependencies such as &amp;quot;Subj-people&amp;quot;, &amp;quot;Modifier-red&amp;quot;, which combine both abstract syntactic notations and their concrete word token representations. He applies this method to classifying not only verbs, but also nouns and adjectives. While no evaluation has ever been done to determine if concrete word tokens are necessary when the syntactic phrase types are already presented, Lin's work indirectly shows that the concrete lexical representation is effective.</Paragraph>
    <Paragraph position="4"> WordNet (Fellbaum, 1998) by far is the most widely used semantic database. However, this database does not 3The attempt to label these semantic roles automatically in (Gildea and Jurafsky, 2002) assumes knowledge of the frame and covers only 20% of them.</Paragraph>
    <Paragraph position="5"> always work as successfully as researchers have expected (Krymolowski and Roth, 1998; Montemagni and Pirelli, 1998). This seems to be due to lack of topical context (Harabagiu et al., 1999; Agirre et al., 2001) as well as local context (Fellbaum, 1998). By adding contextual information, many researchers, (e.g., (Green et al., 2001; Lapata and Brew, 1999; Landes et al., 1998)), have already made some improvements over it.</Paragraph>
    <Paragraph position="6"> The work on the importance of connecting syntax and semantics in developing lexical semantic resources shows the importance of contextual information as a step towards deeper level of processing. With hierarchical sentential local contexts embedded and used to categorize word classes automatically, we believe that PhraseNet provides the right direction for building useful lexical semantic database.</Paragraph>
  </Section>
class="xml-element"></Paper>