Apprentissage de représentations de documents et leur exploitation en recherche d'information

In order to perform the document-query similarity, many information retrieval (IR) models represent documents and queries as sets of weighted key words, called « bag of words », or a bag of concepts derived from a linguistic ontology, or constructed automatically by LSI or LDA techniques, to fill the gap between the query vocabulary and the one used in the document. Recent approaches propose to model the term as an embedded vector, called word2vec approaches, allowing to capture relations beyond the co-occurrence by modelling semantic relations between the terms. In this article, we present the state of the art about this topic, as well as our contribution to integrate these approaches within IR models. MOTS-CLÉS : Recherche d’information, apprentissage profond, word2vec, représentations sémantiques.