Fine-Tuning Self-Supervised Multilingual Sequence-To-Sequence Models for Extremely Low-Resource NMT

Neural Machine Translation (NMT) tends to perform poorly in low-resource language settings due to the scarcity of parallel data. Instead of relying on inadequate parallel corpora, we can take advantage of monolingual data available in abundance. Training a denoising self-supervised multilingual sequence-to-sequence model by noising the available large scale monolingual corpora is one way to utilize monolingual data. For a pair of languages for which monolingual data is available in such a pre-trained multilingual denoising model, the model can be fine-tuned with a smaller amount of parallel data from this language pair. This paper presents fine-tuning self-supervised multilingual sequence-to-sequence pre-trained models for extremely low-resource domain-specific NMT settings. We choose one such pre-trained model: mBART. We are the first to implement and demonstrate the viability of non-English centric complete fine-tuning on multilingual sequence-to-sequence pretrained models. We select Sinhala, Tamil and English languages to demonstrate fine-tuning on extremely low-resource settings in the domain of official government documents. Experiments show that our fine-tuned mBART model significantly outperforms state-of-the-art Transformer based NMT models in all pairs in all six bilingual directions, where we report a 4.41 BLEU score increase on Tamil -> Sinhala and a 2.85 BLUE increase on Sinhala -> Tamil translation.