Baseline Transliteration Corpus for Improved English-Amharic Machine Translation

Abstract

Machine translation (MT) between English and Amharic is one of the least studiedand, performance-wise, least successful topics in the MT field. We therefore proposeto apply corpus transliteration and augmentation techniques in this study to addressthis issue and improve MT performance for the language pairs. This paper presentsthe creation, the augmentation, and the use of an Amharic to English transliterationcorpus for NMT experiments. The created corpus has a total of 450,608 parallelsentences before preprocessing and is used to train three different NMT architecturesafter preprocessing. These models are actually built using Recurrent Neural Networkswith attention mechanism (RNN), Gated Recurrent Units (GRUs), and Transformers.Specifically, for Transformer-based experiments, three different Transformer modelswith different hyperparameters are created. Compared to previous works, the BLEUscore results of all NMT models used in this study are improved. One of the threeTransformer models, in particular, achieves the highest BLEU score ever recorded forthe language pairs.