Bridging the Digital Language Divide: A Comprehensive Survey of English-to-Indian Low-Resource Machine Translation

In a nation marked by unparalleled linguistic diversity, translating between English and India’s regional languages poses challenges due to limited parallel corpora, rich morphology, and varied scripts. The paper presents a comprehensive survey of English-to-Indian low-resource machine translation approaches, covering data acquisition from public repositories and in-house sources, corpus preprocessing through tokenization, sub word segmentation, alignment, and synthetic data generation via paraphrasing, back-translation, and noise injection. We analyze three generations of translation paradigms rule-based, statistical, and neural highlighting architectures from RNN-based encoder–decoder models to transformer-based networks and adapter-augmented large language models with sparse Mixture-of-Experts components. We compare evaluation metrics from BLEU and chrF to learned quality estimators like COMET and BERTScore and discuss human assessment protocols for adequacy and fluency. Finally, we identify future directions in transfer learning, unsupervised and semi-supervised adaptation, incontext learning with large language models, and collaborative corpus development, with the goal of bridging India’s digital language divide.