Attention-based neural machine translation with quality analysis on low-resource Digaru-English Pairs

Attention-based neural machine translation (NMT) has significantly improved translation quality for well-resourced language pairs. However, enhancing translation accuracy for low-resource languages remains a challenge due to limited datasets and technical constraints. Despite these challenges, the latest NMT method has notably improved translation precision for language combinations with limited data by employing techniques aimed at boosting the training models. This study focuses on improving translations for the low-resource North-East Indian language, Digaru. We utilize attention-based NMT and the Transformer model with hyperparameter tuning, experimenting with datasets of 5k, 10k, and 50k sentences. Our optimized model shows a notable improvement in BLEU scores over the base models, resulting in significantly better translation accuracy. Additionally, human evaluations and multidimensional quality metrics (MQM) confirm the enhanced translation quality.