Leveraging retrieval-augmented generation for machine translation in low-resource Indian languages

Low-resource machine translation (MT) remains a challenging task due to the scarcity of parallel corpora. This work addresses the English–Bhojpuri translation task by integrating back-translation and retrieval-augmented generation (RAG) into a transformer-based pipeline. We begin with a parallel corpus of 30,000 English–Bhojpuri sentence pairs to fine-tune the IndicBART model and employ iterative back-translation to synthesize additional training data from a monolingual Bhojpuri corpus. To further enhance performance, we adopt a RAG framework using BM25-based retrieval to provide contextual support during generation. Evaluation using SacreBLEU, ChrF, and F1-score metrics demonstrates significant gains over traditional baselines, confirming the effectiveness of retrieval-augmented strategies for low-resource Indian languages.