Cognition-based arguments in support of using multimedia aids for the learning of vocabulary have so far offered only an imprecise, general framework. CALL experimentalists have also tried to establish the effectiveness of multimedia for vocabulary learning, but their attempts reveal that the underlying representations have not been clearly defined. After reviewing these points, we propose criteria for evaluating the quality of a visual representation in a lexical environment. These criteria are then used to discuss visual representations in paper and electronic dictionaries and in CALL environments. A kind of confusion has been made between multimedia and nonverbal knowledge. Hence visual representations are scarce and limited to concrete words. One way to extend multimedia in lexical learning is to rely on linguistic knowledge and build lexical networks. We present the ALEXIA system, a lexical learning environment for French as a second/foreign language. We detail its network module which can automatically build graphs of some lexical semantic relations. It is a first step for offering learners representations they can easily interpret. Visual representations which can cover a significant part of the lexicon are computable, extendable and interactive.