The development of natural language processing tools for dialects faces the severe problem of lack of resources. In cases of diglossia, as in Arabic, one variant, Modern Standard Arabic (MSA), has many resources that can be used to build natural language processing tools. Whereas other variants, Arabic dialects, are resource poor. Taking advantage of the closeness of MSA and its dialects, one way to solve the problem of limited resources, consists in performing a translation of the dialect into MSA in order to use the tools developed for MSA. We describe in this paper an architecture for such a translation and we evaluate it on Tunisian Arabic verbs. Our approach relies on modeling the translation process over the deep morphological representations of roots and patterns, commonly used to model Semitic morphology. We compare different techniques for how to perform the cross-lingual mapping. Our evaluation demonstrates that the use of a decent coverage root+pattern lexicon of Tunisian and MSA with a backoff that assumes independence of mapping roots and patterns is optimal in reducing overall ambiguity and increasing recall.