English-Afaan Oromo Machine Translation Using Deep Attention Neural Network

Abstract

Attention-based neural machine translation (attentional NMT), which jointly aligns and translates, has got much popularity in recent years. Besides, a language model needs an accurate and larger bilingual dataset_ from the source to the target, to boost translation performance. There are many such datasets publicly available for well-developed languages for model training. However, currently, there is no such dataset available for the English-Afaan Oromo pair to build NMT language models. To alleviate this problem, we manually prepared a 25K English-Afaan Oromo new dataset for our model. Language experts evaluate the prepared corpus for translation accuracy. We also used the publicly available English-French, and English-German datasets to see the translation performances among the three pairs. Further, we propose a deep attentional NMT model to train our models. Experimental results over the three language pairs demonstrate that the proposed system and our new dataset yield a significant gain. The result from the English-Afaan Oromo model achieved 1.19 BLEU points over the previous English-Afaan Oromo Machine Translation (MT) models. The result also indicated that the model could perform as closely as the other developed language pairs if supplied with a larger dataset. Our 25K new dataset also set a baseline for future researchers who have curiosity about English-Afaan Oromo machine translation.