Nowadays, we have become familiar with software interacting with us using natural language (for example in question-answering systems for after-sale services, human-computer interaction or simple discussion bots). These tools have to either react by keyword extraction or, more ambitiously, try to understand the sentence in its context. Though the simplest of these programs only have a set of pre-programmed sentences to react to recognized keywords (these systems include Eliza but also more modern systems like Siri), more sophisticated systems make an effort to understand the structure and the meaning of sentences (these include systems like Watson), allowing them to generate consistent answers, both with respect to the meaning of the sentence (semantics) and with respect to its form (syntax). In this thesis, we focus on syntax and on how to model syntax using categorial grammars. Our goal is to generate syntactically accurate sentences (without the semantic aspect) and to verify that a given sentence belongs to a language - the French language. We note that AB grammars, with the exception of some phenomena like quantification or extraction, are also a good basis for semantic purposes. We cover both grammar extraction from treebanks and parsing using the extracted grammars. On this purpose, we present two extraction methods and test the resulting grammars using standard parsing algorithms. The first method focuses on creating a generalized tree transducer, which transforms syntactic trees into derivation trees corresponding to an AB grammar. Applied on the various French treebanks, the transducer’s output gives us a wide-coverage lexicon and a grammar suitable for parsing. The transducer, even if it differs only slightly from the usual definition of a top-down transducer, offers several new, compact ways to express transduction rules. We currently transduce 92.5% of all sen- tences in the treebanks into derivation trees.For our second method, we use a unification algorithm, guiding it with a preliminary clustering step, which gathers the words according to their context in the sentence. The comparision between the transduced trees and this method gives the promising result of 91.3% of similarity.Finally, we have tested our grammars on sentence analysis with a probabilistic CYK algorithm and a formula assignment step done with a supertagger. The obtained coverage lies between 84.6% and 92.6%, depending on the input corpus. The probabilities, estimated for the type of words and for the rules, enable us to select only the “best” derivation tree. All our software is available for download under GNU GPL licence.
