Accepted 18th September, 2020

Michael Adjeisah¹, Guohua Liua¹, Richard Nuetey Nortey² and Jinling Song3
¹School of Computer Science and Technology, Donghua University, Shanghai, China.
²School of Information Science and Technology, Donghua University, Shanghai, China.
³School of Mathematics and Information Technology, Hebei Normal University of Science & Technology, Qinhuangdao, Hebei, China

This research presents the first work on the Deep natural language processing DNLP approach for translation purposes from English-Twi (a popular Ghanaian language) language. The study reports machine translation (MT) experiments on English-Twi parallel Bible corpus with the state-of-the-art neural Transformer, a self-attention encoder-decoder. The first attempt at engaging MT on English-Twi corpora was with about 124K parallel-aligned sentences. We investigated the dynamics of layers and the number of models to fit this language pair's demands through the experiment. Furthermore, we leveraged the maximum posterior-based decoding for better performance and performed the n-best beam search on the test data. However, with no baseline translation system, we engaged the Moses toolkit, a Statistical MT (SMT), and the well-known sequence to sequence recurrent neural network (RNN) to serve as a baseline for comparison against the neural Transformer. The results were recorded both in BLEU and TER. This study's findings presented the possibility of conveniently adducing one of the most important directions for MT's future in these language pairs.

Key words: Akan, English-Twi corpora, Ghanaian language, transformer, beam search.