ZurichNLP/romansh-backtranslated

Romansh–German Back-Translation Dataset

Background

This dataset contains Romansh texts paired with German translations generated synthetically using Gemini 2.5 Flash. It was created as part of research on data augmentation for low-resource machine translation of Romansh, a language with 6 distinct written varieties (Rumantsch Grischun, Sursilvan, Sutsilvan, Surmiran, Puter, Vallader).

LLMs tend to confuse Romansh varieties when translating into Romansh, but perform well translating out of Romansh into German. This asymmetry makes LR → HR augmentation (Romansh → German) the ideal strategy. Synthetic German translations produced this way allow models to surpass a Gemini 3 Pro baseline on German → Romansh translation by up to +23 BLEU in the lowest-resource variety.

The language variety of each Romansh text was automatically predicted using `ZurichNLP/romansh-lid`.

Usage

from datasets import load_dataset

dataset = load_dataset("ZurichNLP/romansh-backtranslated")

This dataset is intended for NLP and machine translation research. The license of the individual rows is defined per-row.

Paper

Translation Asymmetry in LLMs as a Data Augmentation Factor: A Case Study for 6 Romansh Language Varieties Jannis Vamvas, Ignacio Pérez Prat, Angela Heldstab, Dominic P. Fischer, Sina Ahmadi, Rico Sennrich arXiv:2603.25489 · https://arxiv.org/abs/2603.25489

Dataset Fields

Source Datasets

This dataset contains backtranslations of the following publicly available Romansh sources:

Citation

If you use this dataset, please cite:

@misc{vamvas2026translationasymmetryllmsdata,
  title={Translation Asymmetry in LLMs as a Data Augmentation Factor: A Case Study for 6 Romansh Language Varieties},
  author={Jannis Vamvas and Ignacio Pérez Prat and Angela Heldstab and Dominic P. Fischer and Sina Ahmadi and Rico Sennrich},
  year={2026},
  eprint={2603.25489},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/abs/2603.25489}
}