Section: New Results

Statistical models of word order in French

Participants : Juliette Thuilier, Benoît Crabbé.

We study the problem of choice in the ordering of French words using statistical models along the lines of [66] and [67] . This work aims at describing and model preferences in syntax, bringing additional elements to Bresnan's thesis, according to which the syntactic competence of human beings can be largely simulated by probabilistic models. We previously investigated the relative position of attributive adjectives with respect to the noun.

This year, we mainly studied the problem of the relative ordering of postverbal complements. The focus of this investigation is the relative order of direct object and indirect object of French ditransitive verbs. The first part of this work is based on corpora data that we extracted from two journalistic corpora (French Tree Bank and Est-Républicain) and a radio corpus (ESTER). These data were manually annotated and validated for semantic categories (animacy and semantic class of the ditransitive verb). Based on these data, we built statistical models showing that the relative length of complements and verbal lemmas are the most important factors, and that, differently from English or German, categories as animacy or definiteness seem to play no role in the relative ordering.

In collaboration with Anne Abeillé (Laboratoire de Linguistique Formelle, Université Paris 7), we extended our corpora study with psycholinguistic questionnaires, in order to show that statistical models are reflecting some linguistic knowledge of French speakers. The preliminary results confirm that animacy is not a relevant factor in ordering French complements.

As regards to corpus work, we are extending the database with spontaneous speech corpora (CORAL-ROM and CORPAIX) and a wider variety of verbal lemmas, in order to enhance sample representativeness and statistical modelling. In a crosslinguistic perspective, we plan to strengthen the comparison with the constraints observed in other languages such as English or German.

As can be seen from the outline above, this line of research brings us closer to cognitive sciences. We hope in the very long run that these investigations will bring new insights on the design of probabilistic parsers or generators. In NLP the framework that is closest to implementing construction grammar is Data Oriented Parsing.