Automatic processing of the Arabic language is complicated because of its many forms of writing, spelling, structure, etc., on the one hand, and the lack of preprocessed and normalized data, on the other. Implementing solutions that can help remedy these problems is a real need and a big challenge for the standardization process that this language must know, especially in the new world of publishing which is the Web; characterized by many forms of writing styles where everyone writes in his own way without any constraints. It is in this context that we propose an unsupervised approach based on deep learning implementing a system to help a normalization of a writing, according to a context characterized mainly by Arabic texts written in “Arabic” script.