Hindi Poetry Translation Using Neural Machine Translation

Abstract

Neural Machine Translation (NMT) has become an essential tool in Natural Language Processing (NLP), enabling the automatic translation of text across languages. However, translating poetry remains a relatively unexplored and complex task. Unlike general text, poetry carries multiple layers of meaning, emotion, rhythm, and cultural nuance, making it difficult to capture in another language through standard translation techniques. This study focuses on the translation of Hindi poetry into English using modern NMT approaches. The goal is to make the works of celebrated Indian poets more accessible to non-Hindi-speaking audiences while preserving their poetic qualities. We review existing poetry translation systems, discuss the challenges they face, and explore how translation quality is typically measured. We implemented two NMT models, an attention-based recurrent neural network and a transformer architecture, and fine-tuned them using a curated Hindi poetry dataset. Evaluation results show that our models significantly improve translation quality compared to commonly used online tools, with up to a 15% increase in BLEU scores. In addition, we consider how newer AI techniques, such as Retrieval-Augmented Generation (RAG), could further enhance poetry translation by providing contextual and cultural information during generation. Our work highlights the need for translation systems that go beyond literal meaning and better capture the expressive nature of poetic language.