Zero-shot English–Assamese neural machine translation via pivot-based cross-lingual embedding alignment and transfer learning

Neural Machine Translation (NMT) for low-resource languages such as Assamese faces significant challenges due to the scarcity of parallel corpora. To address this, we propose a novel zero-shot translation framework that leverages linguistic proximity between Assamese and Bengali, enabling English-to-Assamese translation without direct parallel data. Our approach combines a multilingual transfer learning architecture based on mBART with cross-lingual embedding alignment via Procrustes optimization, unifying English, Bengali, and Assamese into a shared semantic space. We further mitigate out-of-vocabulary (OOV) challenges through subword tokenization, embedding projection, and dynamic vocabulary expansion, while advanced language tagging mechanisms reduce off-target errors by enforcing explicit language distinctions during inference. We evaluate our framework on a curated English-Assamese dataset, achieving a BLEU score of 28.65-7.53 points higher than direct translation baselines. Human evaluations highlight significant improvements, with adequacy and fluency scores of 4.5/5 and 4.6/5, respectively, and an 8.4% error rate reduction. Ablation studies confirm the critical roles of embedding alignment and language-aware tagging, while inference speed analysis demonstrates a 26.9% improvement over baseline models. This work demonstrates that (pivot based) transfer learning and cross-lingual alignment is effective in bridging resource gaps for linguistically proximate languages.