A probabilistic approach for extracting opinion-related word chains from texts Extraction probabiliste de cha nes de mots relatives à une opinion

We present a probabilistic method aimed at extracting opinion-related strings from corpora labeled according to customer mind. These strings first allow us to improve text categorization systems according to opinions (positive, negative or neutral). Second, we use them to display easily what are the frequent comments made by customers about products or services. We test the method on two critical corpora written by internet users about video games and movies (respectively in French language and in English language) and on a customer satisfaction phone survey. For each of them, we present some examples of extracted word chains and the observed improvement obtained for opinion-oriented text categorization task.