Abstract
With the growing popularity of the Web 2.0, we are more and more provided with documents expressing opinions on different topics. Recently, new research approaches were defined in order to automatically extract such opinions on the Internet. Usually they consider that opinions are expressed through adjectives and they extensively use either general dictionaries or experts in order to provide the relevant adjectives. Unfortunately these approach suffer the following drawback: for a specific domain either the adjective does not exist or its meaning could be different from another domain. In this paper, we propose a new approach focusing on two steps. First we automatically extract from the Internet a learning dataset for a specific domain. Second we extract from this learning set, the set of positive and negative adjectives relevant for the domain. Conducted experiments performed on real data show the usefulness of our approach.