Asymmetric Data Augmentation for Low-Resource Romansh Translation
August 28, 2026
Research on Romansh language varieties shows that translating from low-resource to high-resource languages (e.g., Romansh to German) is a more effective data augmentation strategy than the reverse. This approach outperformed Gemini 3 Pro by 23 BLEU in the lowest-resource variety.
HOW THIS AFFECTS YOU
●
researcherYou should choose the direction of synthetic data generation based on translation asymmetry to improve low-resource model performance.