<?xml version="1.0" standalone="yes"?>
<Paper uid="P98-1022">
  <Title>A Probabilistic Corpus-Driven Model for Lexical-Functional Analysis</Title>
  <Section position="5" start_page="150" end_page="150" type="concl">
    <SectionTitle>
4. Conclusion and computational issues
</SectionTitle>
    <Paragraph position="0"> Previous DOP models were based on context-free tree representations that cannot adequately represent all linguistic phenomena. In this paper, we gave a DOP model based on the more articulated representations provided by LFG theory. LFG-DOP combines the advantages of two approaches: the linguistic adequacy of LFG together with the robustness of DOP. LFG-DOP triggers a new, corpus-based notion of grammaticality, and its probability models exhibit a preference for the most specific analysis containing the fewest number of feature generalizations.</Paragraph>
    <Paragraph position="1"> The main goal of this paper was to provide the theoretical background of LFG-DOP. As to the computational aspects of LFG-DOP, the problem of finding the most probable representation of a sentence is NP-hard even for Tree-DOP. This problem may be tackled by Monte Carlo sampling techniques (as in Tree-DOP, cf. Bod 1995) or by computing the Viterbi n best derivations of a sentence. Other optimization heuristics may consist of restricting the fragment space, for example by putting an upper bound on the fragment depth, or by constraining the decomposition operations. To date, a couple of LFG-DOP implementations are either operational (Cormons, forthcoming) or under development, and corpora with LFG representations have recently been developed (at XRCE France and Xerox PARC). Experiments with these corpora will be presented in due time.</Paragraph>
  </Section>
class="xml-element"></Paper>