Comparison between Neural and Statistical translation after transliteration of Algerian Arabic Dialect
Résumé
Research on Arabic Dialect Treatment has recently become important in the literature. Although most work on these dialects considers only the messages or the portion of text written in Arabic letters, another style of writing has emerged on social media. This style is known by Arabizi and combines between Latin letters and numbers. To address this emergent problem in the context of automatic translation, we present an Arabic dialect translation system composed by two modules: Transliteration and translation. We develop each module with a statistical and a neural model. To test our system, we used the Algerian portion of a multi-dialectal Arabic corpus named PADIC. Experimental results show that a good transliteration improves the translation results. Moreover, the neural transliteration gives better results than the statistical transliteration. However, the statistical translation still gives better results that the neural translation.