Résumé
Typed lexical relations between terms are indispensable for the tasks realized in NLP, but collecting lexical information is a difficult process. Indeed, when done manually, it requires the competence of experts and the duration can be prohibitive. When done automatically, the results can be biased by the chosen corpus of texts. The approach we present here consists in having people take part in a collective project by offering them a playful application accessible on the web. From an already existing base of terms, the players themselves thus build the lexical network, by supplying associations which are validated only by an agreeing pair of users. Furthermore, these typed relations are weighted according to the number of pairs of users who provide them. We then approach the question of word usage determination for a term, by searching relations between this term and its neighbours in the network, before introducing the notion of similarity between these different word usages. We are thus able to build the tree of word usages for a term. Finally, we briefly present the realization and the first obtained results.