Exploiting large pre-trained models for low-resource neural machine translation

Pre-trained models have revolutionized the natural language processing field by leveraging large-scale language representations for various tasks. Some pre-trained models offer general-purpose representations, while others are specialized in particular tasks, like neural machine translation (NMT). Multilingual NMT-targeted systems are often fine-tuned for specific language pairs, but there is a lack of evidence-based best-practice recommendations to guide this process. Additionally, deploying these large pre-trained models in computationally restricted environments, typically found in developing regions where low-resource languages are spoken, has become challenging. We propose a pipeline to tune the mBART50 pre-trained model to 8 diverse low-resource language pairs, and then distill the resulting system to obtain lightweight and more sustainable NMT models. Our pipeline conveniently exploits back-translation, synthetic corpus filtering, and knowledge distillation to deliver efficient bilingual translation models that are 13 times smaller, while maintaining a close BLEU performance.This paper is part of the R+D+i project PID2021-127999NB-I00 funded by the Spanish Ministry of Science and Innovation (MCIN), the Spanish Research Agency (AEI/10.13039/501100011033) and the European Regional Development Fund A way to make Europe. The computational resources used were funded by the European Regional Development Fund through project IDIFEDER/2020/00