<?xml version="1.0" standalone="yes"?>
<Paper uid="A00-2034">
  <Title>Using Semantic Preferences to Identify Verbal Participation in Role Switching Alternations.</Title>
  <Section position="3" start_page="0" end_page="256" type="intro">
    <SectionTitle>
RSAS.
2 Motivation
</SectionTitle>
    <Paragraph position="0"> Diathesis alternations have been proposed for a number of NLP tasks. Several researchers have suggested using them for improving lexical acquisition.</Paragraph>
    <Paragraph position="1"> Korhonen (1997) uses them in subcategorization frame (SCF) acquisition to improve the performance of a statistical filter which determines whether a SCF observed for a particular verb is genuine or not.</Paragraph>
    <Paragraph position="2"> They have also been suggested for the recovery of predicate argument structure, necessary for SCF acquisition (Briscoe and Carroll, 1997; Boguraev and Briscoe, 1987). And Ribas (1995) showed that selectional preferences acquired using alternations performed better on a word sense disambiguation task compared to preferences acquired without alternations. He used alternations to indicate where the argument head data from different slots can be combined since it occupies the same semantic relationship with the predicate.</Paragraph>
    <Paragraph position="3"> Different diathesis alternations give different emphasis and nuances of meaning to the same basic content. These subtle changes of meaning are important in natural language generation (Stede, 1998).</Paragraph>
    <Paragraph position="4"> Alternations provide a means of reducing redundancy in the lexicon since the alternating scFs need not be enumerated for each individual verb if a marker is used to specify which verbs the alternation applies to. Alternations also provide a means of generalizing patterns of behaviour over groups of verbs, typically the group members are semantically related. Levin (1993) provides a classification of over 3000 verbs according to their participation in alternations involving NP and PP constituents. Levin's classification is not intended to be exhaustive. Automatic identification of alternations would be a useful tool for extending the classification with new participants. Levin's taxonomy might also be used alongside observed behaviour, to predict unseen behaviour. null Levin's classification has been extended by other NLP researchers (Doff and Jones, 1996; Dang et al.,  1998). Dang et al. (1998) modify it by adding new classes which remove the overlap between classes from the original scheme. Dorr and Jones (1996) extend the classification by using grammatical information in LDOCE alongside semantic information in WordNet. What is missing is a way of classifying verbs when the relevant information is not available in a manmade resource. Using corpora by-passes reliance on the availability and adequacy of MRDs.</Paragraph>
    <Paragraph position="5"> Additionally, the frequency information in corpora is helpful for estimating alternation productivity (Lapata, 1999). Estimations of productivity have been suggested for controlling the application of alternations (Briscoe and Copestake, 1996). We propose a method to acquire knowledge of alternation participation directly from corpora, with frequency information available as a by-product.</Paragraph>
  </Section>
class="xml-element"></Paper>