Évaluation de modèles de classification automatique appliqués à la détection d'opinions

This paper describes the problem of classifying opinions expressed into sentences. The system must categorize them as opinionated or factual. To achieve this objective, we have used a Naïve Bayes approach and Support Vector Machines. As a new categorization model, we suggest using a normalized score (Z score) based on a specific vocabulary. Using a NTCIR test collection, our evaluations demonstrate that the suggested model based on the Z score performs significantly better than the others and that a representation based on words tends to show better performance level than surrogates based on lemmas. MOTS-CLÉS : détection d’opinions; classification d'opinions ; évaluation de classifieurs.