<?xml version="1.0" standalone="yes"?>
<Paper uid="W00-0742">
  <Title>Inductive Logic Programming for Corpus-Based Acquisition of Semantic Lexicons</Title>
  <Section position="6" start_page="205" end_page="205" type="concl">
    <SectionTitle>
5 Conclusion
</SectionTitle>
    <Paragraph position="0"> The Inductive Logic Programming learning method that we have proposed in order to define what is a N-V pair whose elements are 17(screw, nut, door, indicator signal, plug, cowl, cap). bound by one of the qualia relations in Pustejovsky's Generative Lexicon formalism leads to very promising results: 83.05% of relevant pairs (after one occurrence) are detected for seven significant nouns; these results have to be compared with the 64% results of Chi-square. It is worth noticing that beyond this simple comparison with one of the possible pure statistics based method is, the interest of using ILP learning is its explanatory characteristic; and it is this characteristic that have motivated our choice: contrary to statistical methods, our ILP method does not just extract statistically correlated pairs but it permits to automatically learn rules that distinguish relevant pairs from others.</Paragraph>
    <Paragraph position="1"> The fact that noise has to be used in Progol to obtain these results however means that something is missing in our E + to fully define the concept &amp;quot;qualia pair&amp;quot; versus &amp;quot;not qualia pair&amp;quot;; some E- have to be covered to define it better.</Paragraph>
    <Paragraph position="2"> A piece of information, maybe syntactic and/or semantic is missing in our E + to fully characterize it. This fact can be easily illustrated by the following example: 'Verbinf det N' structures are generally relevant (ouvrir la porte 19, etc.), except when the N indicates a collection of objects (nettoyer l'ensemble du rdservoir 2deg) or a part of an object (vider le fond du rdservoir21).</Paragraph>
    <Paragraph position="3"> A simple POS-tagging of the sentences offers no difference between them. We are currently working on a semantic tagging of the Matra CCR corpus in order to improve the results.</Paragraph>
    <Paragraph position="4"> Another future work concerns the automatic distinction between the various qualia roles during learning. The last phase of the project will deal with the real use of the N-V pairs obtained by the machine learning method within one information retrieval system and the evaluation of the improvement of its performances.</Paragraph>
  </Section>
class="xml-element"></Paper>