BM25-Augmented Many-Shot Translation for Low-Resource North-Eastern Indian Languages

2026-08-13

AI Summary

This study addresses the challenge of bidirectional translation between English and eleven low-resource Northeast Indian languages by proposing a tuning-free, retrieval-augmented multi-example translation framework. The approach leverages BM25 to retrieve similar examples that augment the Gemini 2.5 Flash model, with configuration parameters optimized via grid search to effectively overcome data scarcity in low-resource settings. Experimental results demonstrate that this tuning-free method achieves superior performance on the WMT26 low-resource translation task, significantly improving translation quality for Northeast Indian languages. These findings validate the transferability and effectiveness of retrieval-augmented generation techniques for extremely low-resource language pairs, offering a viable solution without requiring model fine-tuning.

Abstract

This paper describes the University of Florida Gators submission to the WMT26 Low-Resource Indic Language Translation shared task. We adapt the retrieval-augmented many-shot translation pipeline from our AmericasNLP 2026 system to translate between English and eleven North-Eastern Indian languages in both directions. At inference time, BM25 retrieves the most similar parallel examples from a language-specific training bank, and Gemini 2.5 Flash translates the input conditioned on these examples. No model fine-tuning is involved. Training banks combine official WMT26 data with publicly available corpora such as Samanantar and prior WMT shared task releases. A grid search over retrieval count r and development exemplar count d across all 22 language-direction pairs selects the best configuration for each submission.

Problem

• Low-Resource Translation
• North-Eastern Indian Languages
• Machine Translation
• WMT26

Innovation

• Retrieval-Augmented Translation
• Many-Shot Learning
• BM25
• Low-Resource Languages
• In-Context Learning