STYLISTICALLY-AWARE HINDI-ENGLISH POETIC TRANSLATION WITH MBART AND LLM-BASED POST-EDITING

Translating poetry across languages poses significant challenges, particularly for low-resource pairs like Hindi-English, where linguistic, cultural, and stylistic gaps are substantial. This paper introduces a domain-adapted machine translation framework designed to preserve the poetic essence like metaphor, emotion, tone, and rhythm of Hindi poetry in English translation. We fine-tune mBART50, a multilingual sequence-to-sequence model, on a curated parallel corpus of Hindi-English poetic pairs, augmented with stylistic tags. We also explore optional post-editing using large language models (LLMs) to enhance fluency and poetic expressiveness. Our approach outperforms standard translation systems, as demonstrated by both automatic metrics and human evaluations. To the best of our knowledge, this is the first systematic study of Hindi–English poetry translation that combines (i) the construction of a novel annotated parallel corpus, (ii) style-aware fine-tuning of mBART50, and (iii) LLM-based post-editing for poetic refinement. This work represents a step toward building translation systems that capture not just meaning, but the creative and emotional depth of poetry especially in low-resource language settings.