The Impact of Word Segmentation Techniques on Neural and Statistical Machine Translation: English-Arabic Case

This paper deals with Machine Translation between the English and Arabic languages. This task is very tricky given the morphological richness of the Arabic language and the unavailability of large parallel corpora. To overcome those issues, we have examined the impact of word segmentation (sub-word and morphological segmentation) on machine translation performance. We have tested both the statistical approach and the neural approach which is widely employed in recent years owing to its promising results. In our experiments, carried out on English-Arabic direction and based on the United Nations parallel corpus, we show that applying morphological segmentation to the target language proved very beneficial, whereas sub-word segmentation made no significant impact on both neural and statistical models.