A Transformer-Based Model for English to Yoruba Machine Translation

In Nigeria, Yoruba is a predominant language among over 500 spoken languages. Despite its importance, Yoruba is classified as a low-resourced language, lacking the extensive datasets and linguistic resources needed for advanced natural language processing (NLP) applications. While machine translation (MT) has advanced significantly over the past decade, these improvements mainly benefit high-resourced languages due to the large corpora required for training neural networks. Consequently, Yoruba remains under-resourced in computational linguistics, and NLP research for Yoruba is still in its early stages. This research focuses on developing a machine translation model for the agricultural domain, translating from English to the resource-scarce language Yoruba using transfer learning. The research involved creating approximately 11,000 agriculture-related texts and generating human translations for these texts. This translation pair was then used to train a model based on the pre-trained Google T5-small model. The transformer's ability to handle complex linguistic patterns and variable-length sequences improves translation quality and reduces training time and development complexity, even with a small training dataset. The model achieved a BLEU score of 0.5354 on the evaluation set against the BLEU score of 0.5369 obtained during the model training. Additionally, the translation quality was evaluated by humans in terms of representation accuracy, contextual correctness, and syntactic correctness. Using the mean opinion score of ten (10) local speakers, the translation quality of the model generated in this research ranks above translations from a state-of-the-art model in terms of syntactic and semantic correctness. The results demonstrated that the model developed in this research used domain-specific terms from the training corpus, providing more informative and beneficial translations for local farmers than Google Translate, which often relies on transliteration.