The growing popularity of Web 2.0 provides with increasing numbers of documents expressing opinions on different topics. Recently, new research approaches have been defined in order to automatically extract such opinions from the Internet. They usually consider opinions to be expressed through adjectives, and make extensive use of either general dictionaries or experts to provide the relevant adjectives. Unfortunately, these approaches suffer from the following drawback: in a specific domain, a given adjective may either not existor have a different meaning from another domain. In this paper, we propose a new approach focusing on two steps. First, we automatically extract a learning dataset for a specific domain from the Internet. Secondly, from this learning set we extract the set of positive and negative adjectives relevant to the domain. The usefulness of our approach was demonstrated by experiments performed on real data.