Résumé
The semantic analysis of texts requires beforehand the building of objects related to lexical semantics. Idea vectors and lexical networks seems to be adequate for such a purpose and are complementary. However, one should still be able to construct them in practice. Vectors can be computed with definition corpora extracted from dictionaries, with thesaurii or with plain texts. They can be derived as conceptual vectors, anonymous vectors or lexical vectors - each of those being a particular balance between precision, coverage and practicality. Concerning lexical networks, they can be efficiently constructed through serious games, which is precisely the goal of the JeuxDeMots project. The semantic analysis can be tackled from the thematic analysis, and can serve as computing means for idea vectors. We can modelise the analysis problem as actviations and propagations. The numerous criteria occuring in the semantic analysis and the difficulties related to the proper definition of a control function, lead us to explore metaheuristics inspired from nature. More precisely, we introduce an analysis moodel based on artificial ant colonies. From a given text, the analysis aims at building a graph holding objects of the text (words, phrases, sentences, etc.), highlighting objects considered as relevant (phrases and concepts) as well as typed and weighted relations between those objects.