Paper Details
Abstract
Recent advances in multilingual machine translation have demonstrated remarkable potential for improving translation quality, particularly for low-resource languages. In this study, we conduct a comparative analysis of two prominent multilingual models: the mBART-50 and the No Language Left Behind (NLLB) models, both fine-tuned on the IWSLT2015 English-Vietnamese dataset. Specifically, we compare facebook/mbart-large-50-many-to-many-mmt with facebook/nllb-200-distilled-600M, implementing identical preprocessing, fine-tuning strategies, and evaluation metrics to ensure a fair comparison. Our experiments reveal that the NLLB model achieves a BLEU score of 35.81, outperforming mBART-50's score of 33.97 on the same test set, despite having a smaller parameter footprint. We analyze the strengths and limitations of each model, particularly examining their ability to handle domain-specific terminology and syntactic structures common in the TED talks domain. This study contributes to the understanding of how different multilingual architectures perform on low-resource language pairs and provides insights into selecting appropriate models for English-Vietnamese translation tasks. The source code, data, and fine-tuned models are publicly available at: GitHub repository ([https://github.com/vuhuyng04/NMT\_mBART-50\_NLLB.git](https://github.com/vuhuyng04/NMT_mBART-50_NLLB.git)), fine-tuned mBART-50 model ([https://huggingface.co/nguyenvuhuy/en-vi-mbart50\_TMG301](https://huggingface.co/nguyenvuhuy/en-vi-mbart50_TMG301)), and fine-tuned NLLB-200 model ([https://huggingface.co/nguyenvuhuy/en-vi-nllb-200\_TMG301](https://huggingface.co/nguyenvuhuy/en-vi-nllb-200_TMG301)).