Comparative Evaluation of RNN-based Bahdanau Attention vs Transformer Model for NMT in South African Low-Resource Languages
Philasande T Mkhwanazi, Skhumbuzo G. Zwane, Matthew Olusegun Adigun, Eyethu K Thwala · 2025
The linguistic diversity of South Africa presents significant challenges for multilingual communication, particularly in resource-constrained environments such as customer call centers. This study uses South African low-resource language pairs to investigate and compare the performance of two prominent neural machine translation (NMT) architectures, namely the Bahdanau Attention Model and the Transformer Model with multi-headed self-attention. The study would illustrate which model performes best in translation between low-resource languages of South Africa. The study uses parallel corpora from JW300 and Autshumato datasets. Three language pairs were evaluated: isiZulu to Sesotho, Swati to Ndebele, and Tsonga to Setswana. The languages were selected to best cover the south African low resource languages families. The study adopts the Design Science Research Methodology (DSRM) for implementing two NMT architectures. Each model was trained under equivalent conditions and assessed using BLEU scores, sentence length distribution, and training time per epoch. Results indicate that the Transformer model consistently outperformed the attention-based model in translation accuracy (BLEU-4), while the attention model showed higher BLEU-1 scores in some instances, suggesting better word-level precision. The results also suggested that similarities in grammatical structure, vocabulary overlap, and sentence distribution patterns are pivotal in translation accuracy and contextual understanding.