Abstract
In this paper, we address the challenge of lowresource machine translation for the Atayal-Chinese language pair by leveraging Indonesian as a pivot language for cross-lingual transfer. Specifically, we fine-tuned the mBART50 model using Indonesian-Chinese parallel data to adapt the model to a typologically related language. This adapted model was then further fine-tuned on a small set of Atayal-Chinese parallel data. Our experiments demonstrate that this two-stage fine-tuning approach significantly improves translation performance. We achieved a BLEU score of 14.00 in the Atayal-to-Chinese direction and 21.39 in the Chinese-to-Atayal direction, showing the effectiveness of pivot-based transfer for low-resource scenarios.