Improving Bidirectional English–Assamese Neural Machine Translation with Word Embeddings and Transformer Models

NMT for Assamese language remains challenging because of its morphological rich language and lack of parallel resources. The paper aims to study the effect of word embeddings, different neural architectures and bidirectional pre-training for enhancing bidirectional English–Assamese translation. Three embedding methdos (Word2Vec, GloVe and FastText) were evaluated within RNNs using GRU and LSTM architecture. FastText consistently yielded the highest BLEU scores among recurrent models, achieving 24.99 for Assamese–English and 24.65 for English–Assamese, owing to its ability to capture subword level information and handle rare word forms. To establish stronger baselines, we fine tuned Transformer-based models (MarianMT and mBART) on a 175,000-sentence parallel corpus and further enhanced mBART through a Bidirectional Training (BiT) phase to promote cross-lingual parameter sharing. The BiT-pretrained variant achieved the best performance, with BLEU scores of 28.93 (As–En) and 28.57 (En–As). Finally, we compared these systems against a zero-shot GPT baseline, which underperformed relative to the fine-tuned models, reaffirming the effectiveness of specialized training for low-resource translation. Human evaluation of fluency and adequacy supported these results, confirming that Transformer-based architectures with bidirectional pre-training remain the most effective solution for English–Assamese NMT.