Abstract. Using semantic knowledge in NLP applications always improves their competence. Broad lexicons have been developed, but there are few resources which contain semantic information available for words and which are non-dedicated to specialized domains. In order to build such a base, we designed a system, SVETLAN', able to learn categories of nouns from texts, whatever their domain. In order to avoid general classes mixing all the meanings of words, they are learned taking into account the contextual use of words.