The developpment of NLP tools for dialects faces the severe problem of lack of resources for such dialects. In the case of diglossia, as in arabic, a variant of arabic, Modern Standard Arabic, exists, for which many resources have been developped which can be used to build NLP tools. Taking advantage of the closeness of MSA and dialects, one way to solve the problem consist in performing a surfacic translation of the dialect into MSA in order to use the tools developped for MSA. We describe in this paper an achitecture for such a translation and we evaluate it on arabic verbs.