Unsupervised Word Segmentation Improves Dialectal Arabic to English Machine Translation

Abstract

We demonstrate the feasibility of using unsupervised morphological segmentation for dialects of Arabic, which are poor in linguistics resources. Our experiments using a Qatari Arabic to English machine translation system show that unsupervised segmentation helps to improve the translation quality as compared to using no segmentation or to using ATB segmentation, which was especially designed for Modern Standard Arabic (MSA). We use MSA and other dialects to improve Qatari Arabic to English machine translation, and we show that a uniform segmentation scheme across them yields an improvement of 1.5 BLEU points over using no segmentation.