Building a Tunisian Dialect into Arabic Language Parallel Corpus for a Phrase-based Machine Translation

The purpose of this paper is to build a system capable of translating the dialect of Tunisia's capital into the Modern Standard Arabic. Having such a tool can have an impact in various domains such as translating social network user interactions, subtitles of Tunisian movies, books written with the writers' local dialect, etc. Since the Tunisian dialect is classified as a low-resource dialect as well as the other Arabic dialects. We started by building a parallel corpus that contains 5000 sentences. These latter are extracted from different sources accessible on the web and specifically from Facebook and YouTube. Then, they were manually translated into the Arabic. Afterwards, this resource was used to adopt the statistical approach, which is based on the creation of the Language Model ( LM) and the Translation Model (TM). These two models are then used by the decoder to choose the best translation for an input sentence. The results were promising where we achieved 49.90 as a BLEU score.