•Title : “Synchronization of syntactical and semantic knowledge for natural language analysis in the Lexicalized Tree Adjoining Grammars framework” (Synchronisation des connaissances syntaxiques et sémantiques pour l'analyse d'énoncés en langage naturel à partir du formalisme des Grammaires d'Arbres Adjoints)
PhD in Computer Sciences at the University Henry Poincaré (UHP, Nancy, France) Defended the 5th of November 2004
Supervised by Jean Marie Pierrel (Pr. University Henry Poincaré), Bertrand Gaiffe (Researcher, CNRS-Loria)
Phillipe Blache, Research Director, CNRS & University of Province
Christophe Fouqueré, Pr. University PARIS Nord (13) & CNRS-LIPN
Jean Yves Marion, Pr. Institut National Polytechnique de Lorraine & INRIA-Lorraine
Eric Villemonte de la Clergerie, Researcher, INRIA-Rocquencourt
Laurence Danlos, Pr. University Paris Denis Diderot (7) & CNRS
An interface between syntax and semantic aims to propose a logical formalization of the relations between parts of a sentence. This thesis is a proposal based upon the analysis of problematic linguistic phenomena in the Lexicalized Tree Adjunct Grammars (LTAG) framework. LTAG is a linguistic formalism whose minimal units are lexicalized trees. LTAG provides two structures of representation, derived tree and derivation tree. The last one is an almost perfect structure to be used as a canvas for semantic analysis.
However, the derivation tree cannot represent coindexations in a autonomous way. We based our proposition upon the study of linguistic phenomena induced by control verbs. In order to allow their treatment and their complete formalization, we modify the initial LTAG formalism by introducing a new lexical information : the control canvas and a new operation of derivation : the fusion. Its purpose is to integrate inference of missing argumental links into a synchronous course of both the derived tree and the derivation tree through the use of a shared forest which contains all the needed informations. We propose a dynamic reconstruction algorithm based on inference rules dynamically built. These rules are executed during the derivation tree extraction process from the shared forest. As we use tabular methods we can extract, into a dependency graph à la Meaning Text Theory, all the argumental relations described by one shared forest.
LTAG, tabular parsing, parsing as intersection, parsing as deduction, syntax, semantic, dependency graph, predicate argument structure, co indexation, shared forest, derivation forest, dynamic programming, tabulation, formal logic.