Apprentissage de modèles de langue neuronaux pour la recherche d'information

Information Retrieval (IR) faces different difficulties, notably those related to vocabulary mismatch issues and term dependencies. In the last few years, language models based on neural networks have been proposed to deal with both term dependencies and vocabulary mismatch issues in complex natural language processing tasks. However, to be efficient, these models require huge amounts of training data. They have thus never been employed for IR adhoc tasks directly, where the estimation of one language model per document is required. We propose an approach based on the specialization of a generic language model, learned on the whole document collection, by a set of document-specific parameters, to define neural language models fitted for ad-hoc IR tasks. MOTS-CLÉS : Apprentissage de Représentation, Recherche d’Information, Réseau de Neurones.