A Transformer-Based Approach to Low-Resource African Language Translation

Abstract

We present a transformer architecture fine-tuned for low-resource African languages, achieving competitive BLEU scores against high-resource baselines.

Keywords

Full Text

Machine translation systems for high-resource language pairs have advanced rapidly, but African languages remain severely underrepresented in training corpora. This work fine-tunes a transformer architecture on a curated parallel corpus spanning Yoruba, Hausa and Swahili.

Using back-translation and data augmentation to compensate for limited parallel data, the model achieved BLEU scores within 4 points of comparable high-resource baselines across all three target languages, with the largest gains observed on Swahili due to greater available monolingual data.

References

1. Vaswani A. et al. Attention Is All You Need. NeurIPS, 2017.
2. Masakhane. Participatory Research for Low-Resourced African Languages, 2020.