Distributionnalisme et sémantique : état des lieux en traitement automatique des langues
Résumé
In this article, we would like to show that Harris theories have found a prolific posterity in Natural language Processing (NLP), with a bunch of implementations that prove its language description power. The recent availability of huge corpora and the rise of statistical approaches to language have enabled to assess the validity of the initial intuitions of distributionnalism, such as the main principle that unit distributions are the most appropriate way to explicit the functioning of language, from phonology to semantics. A new field of research in NLP has emerged, called Distributional Semantics, whose goal is to automatically derive semantic structures of languages from contexts similarity. But these works rely uniquely on the notion of similarity, which is too vague for describing the semantics of language. In this paper, we would like to show that Harris has given, essentially in (Harris, 1988), some hints to overcome this dead end, and we try to show, through two experiments, that a comeback to the theoretical assumptions permit to explicit more precisely how we can approch semantics on distributional principles.