Comparative Study of Neural Machine Translation Approaches for Hindi–Malayalam: Bi-LSTM Baselines, Word2Vec and Attention Enhancements, and mBART Transfer Learning

This Hindi–Malayalam machine translation faces significant challenges due to structural differences between Indo-Aryan and Dravidian languages. Malayalam is highly agglutinative, whereas Hindi relies more on syntactic structures and postpositions to express grammatical relations. Limited availability of high-quality parallel corpora further complicates the development of robust translation systems. This study presents a comparative evaluation of neural machine translation architectures for Hindi–Malayalam translation. A curated parallel corpus of about 80,000 sentence pairs was created using automated translation followed by manual correction. Five models were implemented and evaluated: a Bi-LSTM sequence-to-sequence baseline, Bi-LSTM with Word2Vec embeddings, Bi-LSTM with Word2Vec and attention, inference using the pretrained multilingual transformer mBART-50, and fine-tuning of mBART-50 on the dataset. All models were trained using a unified preprocessing pipeline including Unicode filtering and Indic tokenization. Results show clear improvements across architectures, with fine-tuned mBART-50 achieving the highest translation quality, highlighting the effectiveness of multilingual transformer models for low-resource Hindi–Malayalam translation.