Machine Translation for Low Resource Language using NLP Attention Mechanism

Machine translation (MT) systems play a major role in overcoming language barriers, particularly for underutilized language pairs like Bangla-English. This research compares the relative effectiveness of Neural Machine Translation (NMT) against Statistical Machine Translation (SMT) for translating between Bangla and English. It does this by employing a thorough methodology that involves model training, validation, and inference utilizing publicly available corpora. We developed and trained an NMT system using encoder-decoder architecture and attention methods. The system was trained using a Tatoeba Project dataset that was split into training and validation sets. The models were evaluated and modified using standard metrics, and the trained model checkpoints were saved and recovered for inference. Our experiments demonstrate that when translating between Bangla and English, NMT outperforms SMT by a wide margin in terms of BLEU scores.