Résumé
We present the model of language implemented in the words prediction engine of the "Plateforme de Communication Aternative" (PCA), an assistive communication software designed for impaired persons. The model relies on a large coverage lexicon for french langage which provides for each entry its word frequency and its set of associated morphosyntactic categories. The engine includes a user model (personal lexicon for unknown words typed by the user, computation of personal word frequencies, storage of the generated sentences) and a morphosyntactic prediction system which weight the word frequencies of predicted words according to the syntactical context of the sentence. The evaluation of the model of language is quite encouraging : a keystrokes saving rate of 55% for 9 propositions. The major contribution comes from the raw prediction obtained by using the words frequencies of the general lexicon. MOTS-CLES: Communication assistee pour personnes handicapees, systeme de prediction de mots, modele utilisateur, prediction morphosyntaxique.