Translation Asymmetry in LLMs as a Data Augmentation Factor: A Case Study for 6 Romansh Language Varieties

2026-03-26

AI Summary

This work addresses the challenge of mutual interference among the six low-resource Romansh dialects during synthetic data generation with large language models, which degrades machine translation performance. The authors propose a novel data augmentation strategy that aligns the direction of synthetic data generation with the resource gradient between source and target languages, thereby guiding the large language model to produce high-quality, dialect-specific synthetic data. This targeted approach effectively mitigates cross-dialect confusion and significantly improves translation quality across all variants. Notably, on the lowest-resource dialect, the method outperforms Gemini 3 Pro by 23 BLEU points. Human evaluations further confirm substantial gains in both fluency and accuracy of the generated translations.

Technology Category

• Natural Language Processing: Machine Translation, Multilinguality, Cross-Lingual NLP
• Machine Learning: Large Multimodal Models (LMMs)
• Planning, Routing, and Scheduling: Planning with Language Models

Application Category

• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• Web Mining and Content Analysis: Web data generation and simulation
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Recent strategies for low-resource machine translation rely on LLMs to generate synthetic data from higher-resource languages. We find that this method fails for Romansh, because LLMs tend to confuse its 6 distinct language varieties. Our experiments show that instead, the direction of data augmentation should be aligned with the resource gradient between source and target language. This approach surpasses Gemini 3 Pro in the lowest-resource variety of Romansh by 23 BLEU. A human evaluation confirms that our experiments yield the first model that generates fluent translations in the individual Romansh varieties.

Problem

• low-resource machine translation
• Romansh language varieties
• translation asymmetry
• synthetic data generation
• language confusion

Innovation

• translation asymmetry
• data augmentation
• low-resource machine translation
• Romansh language varieties
• LLM-based synthetic data