A Pragmatic Comparison of NMT Strategies for Assamese: Transfer Learning vs. From-Scratch

Developing high-quality Neural Machine Translation (NMT) for low-resource languages (LRLs) like Assamese is challenging due to data scarcity and computational constraints. This paper presents a controlled experiment comparing finetuning a large multilingual model (NLLB-600M) against training a Transformer from scratch. Using a 50,000 parallel Assamese-to-English training dataset on a Google Colab GPU, we compared the NLLB model, which uses a pre-trained 256k tokenizer, against an optimized 3-layer Transformer built with a bespoke 32k tokenizer. We found the finetuned NLLB model achieved substantially higher translation quality, outperforming the fromscratch model by 28+ BLEU. Furthermore, the NLLB model converged dramatically faster, highlighting the overwhelming advantage of its transferred linguistic knowledge, while the fromscratch model struggled to learn basic translation. We conclude that for researchers working under both data and computational constraints, finetuning a large pre-trained model is the more effective and practical strategy.