Neural Machine Translation for Low-Resource Tangkhul--English

Researchers have developed a neural machine translation system for Tangkhul, a severely under-resourced Tibeto-Burman language spoken in Manipur, India, achieving a BLEU score of 39.97 using a fine-tuned ByT5-large model trained on 38,336 parallel sentences. This work addresses a significant gap in NLP infrastructure for one of India's marginalized linguistic communities and demonstrates practical approaches to machine translation for languages with minimal computational resources.

The development of machine translation systems for under-resourced languages represents a critical frontier in natural language processing, where technological advancement intersects with linguistic preservation and digital inclusion. Tangkhul, with virtually no prior NLP infrastructure, exemplifies the thousands of languages globally that lack digital representation despite millions of speakers. This research directly addresses that gap by establishing baseline translation capabilities for a language pair that previously had no computational treatment.

The choice of ByT5-large over alternative architectures reflects an understanding of low-resource constraints—byte-level tokenization proves more effective than word-level approaches when training data is limited and orthographic systems are complex. The 39.97 BLEU score, while respectable for a low-resource setting, acknowledges realistic limitations: the corpus comprises primarily biblical texts, stories, and conversational data, introducing domain bias that would restrict real-world applicability beyond religious and narrative contexts.

From a development perspective, this work establishes foundational infrastructure that enables downstream applications—digital literacy tools, educational platforms, and accessibility services for Tangkhul speakers. The explicit discussion of orthographic challenges with Latin-script diacritics provides practical insights for other under-resourced language pairs facing similar technical hurdles. The research pathway toward data diversification and domain adaptation signals achievable improvements without requiring orders-of-magnitude increases in training data.

Future momentum depends on community engagement and data collection initiatives that expand beyond the current corpus composition, while technical advances in transfer learning and multilingual model efficiency may unlock better performance with minimal additional investment.

• →ByT5-large achieved 39.97 BLEU score for Tangkhul-English translation with only 38,336 parallel sentence pairs, demonstrating viable low-resource MT approaches.
• →Domain bias in training corpus (biblical/religious texts) limits applicability and represents a key target for improvement through diversification.
• →Latin-script diacritics in Tangkhul orthography present specific technical challenges requiring specialized handling beyond standard tokenization approaches.
• →Establishment of NLP infrastructure for under-resourced languages enables downstream applications in education, accessibility, and digital inclusion.
• →Comparative analysis with mT5-small validates architectural choices while suggesting multilingual model efficiency as a viable path for future development.