A Comparative Study of Fine-Tuned mBART and IndicTrans2 for English-Hindi Legal Machine Translation

Natural Language Processing has a hard time translating legal documents because they have complicated syntax and specialised language. Because India has many languages in its government and courts, accurate English to Hindi legal machine translation is very important. This research evaluates two prominent multilingual sequence-to-sequence methodologies for English-Hindi legal translation, namely mBART and Indic-Trans2. A high-quality dataset that is specific to the domain improves both models. The semantically informed COMET metric and the standard BLEU meter measure how well something works. Fine-tuning for a specific domain makes both models better. The fine-tuned mBART model gets a BLEU score of 51.78, while IndicTrans2 gets a score of 47.77. A thorough look at COMET scores shows that mBART makes better translations, which shows how its architecture is good for adapting to new domains in important, specialised translation projects.