<?xml version="1.0" standalone="yes"?>
<Paper uid="W00-0742">
  <Title>Inductive Logic Programming for Corpus-Based Acquisition of Semantic Lexicons</Title>
  <Section position="3" start_page="199" end_page="200" type="intro">
    <SectionTitle>
2 Motivation
</SectionTitle>
    <Paragraph position="0"> As stated in the introduction, our work makes two strong claims: firstly N-V associations defined in GL are relevant for IR and secondly this information can be acquired from a corpus on the basis of surrounding POS context. These presuppositions have to be motivated before explaining the learning method: 1. The aim of GL is to define underspecified lexical representations that will acquire their specifications in context. For example, the qualia structure of book indicates that its default function is read and that it is created by the act of writing. But this information has to be enriched in context in order to characterize how words are used in specific domains. For example, the qualia structure of book will also have to indicate that the book can be shelved or indexed if this information is necessary to interpret texts from information science domain. GL  is therefore a theory of words in context. It can also be seen as a way to structure information in corpora and, in that sense, the relations it defines are therefore privileged information for IR. In this perspective, GL has been preferred to existing lexical resources such as WordNet (Fellbaum, 1998) for two main reasons: lexical relations that we want to exhibit - namely N-V links - are unavailable in WordNet, which focuses on paradigmatic lexical relations; Word-Net is a domain-independent, static resource, which can not be used as such to describe lexical associations in specific texts, considering the great variability of semantic associations from one domain to another.</Paragraph>
    <Paragraph position="1"> 2. In GL, the qualia structures are not arbitrary repository of information. They contain the information necessary to explain the syntactic behaviour of the item. We would therefore expect that there are strong connections between some specific syntactic phenomena and some specific qualia relations. For example, the middle construction seems to be only possible if a telic relation holds between the N and V (Bassac and Bouillon, 2000) (for example: ??this book writes well vs this book reads well). Similarly, imperative constructions (e.g. open the door, follow the links) or adjectival sentences (a book difficult to write/read) may also indicate a qualia relation. These are some of the constructions that we want to identify primilarly in corpora by the learning method.</Paragraph>
  </Section>
class="xml-element"></Paper>