Most of natural language descriptions are made of sets of rules modelling the behavior of words. However, whereas many general rules have been established, exceptions to these rules are not often studied. Consequently, these rules are incomplete, and even inaccurate when the number of particular cases is too large. To solve this problem, the LADL team has studied in detail basic sentences of French. This work led to a very fine description of the syntactic properties of these sentences, stored in matrices called lexicon-grammar tables. In 1993, the proof was made by Emmanuel Roche that these data could be used to perform automatic parsing. We have studied a way to extend this work, in order to take into account the whole data contained in l...