Word Sense Disambiguation (WSD) is often omitted in Statistical Machine Translation (SMT) systems, as it is considered unnecessary for lexical selection. The discussion on the need ofWSD is currently very active. In this article we present the main positions on the subject.We analyze the advantages and weaknesses of the current conception of WSD in SMT, according to which the senses of ambiguous words correspond to their translations in a parallel corpus. Then we present some arguments towards a more thorough analysis of the semantic information induced from parallel corpora and we explain how the results of this analysis could be exploited for a more flexible and conclusive evaluation of the impact of WSD on SMT.