Représentations et régularisations pour la classification de sentiments

As web 2.0 is spreading, users get used to give their opinion on forums, blogs and e-commerce websites. This is a valuable piece of information for many applications such as consumer modeling, sales prediction or opinion survey. According to the literature, the efficiency of opinion mining tools will mainly relies on the ability of discriminating texts that express positive sentiments from texts that express negative ones. Previous experiments show that this task is difficult. We compare various classical descriptors and point out the interest of large representations of texts (N-grams, sub-sequences) for this task. The dimensionality of the data causes problems during the learning step: we demonstrate the inefficiency of the classical regularization framework as well as the interest of penalizing frequent terms. We demonstrate the efficiency of our approach on classical Movie Reviews and Amazon data-sets. MOTS-CLÉS : Classification de sentiments, régularisation, pénalisation des termes fréquents