Extraction d'un vocabulaire de surprise par mélange de filtrage collaboratif et d'analyse de sentiments

Computer science is undergoing a profound transformation: gains in computing power and the availability of large datasets provide a fertile ground for machine learning research. One challenge in this context is to take account of user preferences to provide personalized access to information.Recommender systems create user and item profiles using past user reviews, profiles that reflect user preferences and item characteristics. We propose here an analysis of a combination of recommendation systems and polarity classifiers which highlights a vocabulary of surprise: words indicating a gap between the expectations of users and their actual experience with an item. Indeed, recommender systems analyze past ratings to predict user preferences while sentiment classifiers use existing reviews as input: we show that the difference between expectations and observations enables the construction of a surprise lexicon. MOTS-CLÉS : Filtrage collaboratif, Analyse de sentiments, Surprise.