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

TL;DR

This paper investigates how translation asymmetry in large language models affects data augmentation for low-resource Romansh language varieties, revealing that aligning augmentation direction with resource gradients improves translation quality.

Contribution

It demonstrates that resource-gradient-aligned data augmentation outperforms standard methods, achieving fluent translations in Romansh varieties and surpassing existing models like Gemini 3 Pro.

Findings

Resource-gradient-aligned augmentation improves BLEU scores by 23 in Romansh varieties.

The approach yields the first fluent translations for individual Romansh varieties.

LLMs tend to confuse the 6 Romansh language varieties, affecting translation quality.

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.