Quantification et identification des concepts implicites d'une requête

In this paper we introduce an unsupervised method for mining and modeling latent search concepts. We use Latent Dirichlet Allocation (LDA), a generative probabilistic topic model, to exhibit highly-specific query-related topics from pseudo-relevant feedback documents. Our approach automatically estimates the number of latent concepts as well as the needed amount of feedback documents, without any prior training step. Latent concepts are then weighted to reflect their relative adequacy and are further used to automatically reformulate the initial user query. We also explore the use of different types of sources of information for modeling the latent concepts. For this purpose, we use four general sources of information of various nature (web, news, encyclopedic) from which the feedback documents are extracted. We evaluate our approach over two large ad-hoc TREC collections, and results show that it significantly improves document retrieval effectiveness while best results are achieved by combining latent concepts modeled from all available sources. MOTS-CLÉS : Recherche contextuelle, modélisation thématique, retour de pertinence