Résumé
Recently, machine translation has made significant progress by using lists of compound expressions such as " armed robbery " (" vol à main armée " in French). However, current systems are not yet capable of translating correctly many unknown expressions, such as "caisse centrale", mistranslated by Systran as "central case" instead of "central fund". This expression could have been automatically translated using the Web as a source of examples. The aim of this study is to acquire automatically a database of complex lexical units translations from the Web. We especially tackle difficult translation problems such as polysemy and idioms, for which better translation mechanisms are proposed. Beyond the practical aspects of machine translation improvement, we discuss the possible uses of the Web for general studies in linguistics.