Modèles de Document Parcimonieux basés sur les annotations et les word embeddings - Application à la personnalisation

Abstract

In this paper, we define social parcimonious language models that emphasize the most important terms in documents, and lower less important terms. The detection of important terms relies on the document itsef and on the tags that were employed by users to describe the document. Conversely to classical personalization approaches that focus first on user’s profiles or on the matching function, our proposal focuses on the documents representations. Evaluation achieved on the Social Book Search 2016 collection show that our proposal outperforms reference approaches in several cases. MOTS-CLÉS : Profil utilisateur, modèles parcimonieux, plongement de mots