romansh-backtranslated
收藏资源简介：
罗曼什语-德语反向翻译数据集是一个专为低资源机器翻译研究设计的数据增强资源。该数据集包含罗曼什语文本及其对应的德语合成译文，这些译文是使用Gemini 2.5 Flash大型语言模型通过反向翻译策略生成的。研究背景表明，大型语言模型在将文本翻译成罗曼什语时容易混淆其六种书面变体（Rumantsch Grischun, Sursilvan, Sutsilvan, Surmiran, Puter, Vallader），但在从罗曼什语翻译成德语时表现良好。因此，本数据集采用了罗曼什语到德语的增强策略，旨在利用高质量的合成德语译文来提升德语到罗曼什语的翻译模型性能，据称可使模型在资源最稀缺的变体上超越Gemini 3 Pro基线高达+23 BLEU分数。数据集中的每个罗曼什语文本都通过`ZurichNLP/romansh-lid`工具自动预测了其所属的语言变体。数据来源于多个公开的罗曼什语语料库，包括新闻（La Quotidiana）、网页文本（FineWeb2, FineWiki）、PDF文档（FinePDFs）、戏剧剧本、市政文件、历史词典以及未对齐的平行语料（Mediomatix），涵盖了全部六种罗曼什语变体。数据集包含以下字段：`rm`（源罗曼什语文本）、`de_baseline`（基线提示生成的德语翻译）、`de_with_dictionary`（词典增强提示生成的德语翻译）、`predicted_variety`（自动预测的罗曼什语变体）、`dataset`（源数据集名称）、`metadata`（源数据集的额外元数据）以及`license`（源数据集的许可证）。本数据集主要用于自然语言处理和机器翻译研究，各行的许可证遵循其源数据集的规定。
The Romansh-German back-translation dataset is a data augmentation resource specifically designed for low-resource machine translation research. The core content of this dataset consists of Romansh texts and their corresponding synthetic German translations, which are generated using the Gemini 2.5 Flash large language model through a back-translation strategy. Research background indicates that large language models tend to confuse the six written varieties of Romansh (Rumantsch Grischun, Sursilvan, Sutsilvan, Surmiran, Puter, Vallader) when translating text into Romansh, but perform well when translating from Romansh to German. Therefore, this dataset employs a Romansh-to-German enhancement strategy, aiming to leverage high-quality synthetic German translations to improve the performance of German-to-Romansh translation models, reportedly enabling models to surpass the Gemini 3 Pro baseline by up to +23 BLEU score on the most resource-scarce varieties. Each Romansh text in the dataset is automatically assigned a predicted language variety using the `ZurichNLP/romansh-lid` tool. The data is sourced from multiple public Romansh corpora, including news (La Quotidiana), web texts (FineWeb2, FineWiki), PDF documents (FinePDFs), play scripts, municipal documents, historical dictionaries, and unaligned parallel corpora (Mediomatix), covering all six Romansh varieties. The dataset includes the following fields: `rm` (source Romansh text), `de_baseline` (German translation generated with baseline prompts), `de_with_dictionary` (German translation generated with dictionary-enhanced prompts), `predicted_variety` (automatically predicted Romansh variety), `dataset` (source dataset name), `metadata` (additional metadata from the source dataset), and `license` (license of the source dataset). This dataset is primarily used for natural language processing and machine translation research, with each rows license adhering to the regulations of its source dataset.
数据集概述
数据集名称：Romansh–German Back-Translation Dataset
许可证：ODC-By 1.0（逐行许可证，详见每行数据）
目标语言：罗曼什语（rm）、德语（de）
覆盖罗曼什语变体：Rumantsch Grischun, Sursilvan, Sutsilvan, Surmiran, Puter, Vallader（共6种）
用途：自然语言处理与机器翻译研究，特别适用于低资源语言的数据增强
数据集构成
| 字段 | 说明 |
|---|---|
rm |
源语言罗曼什语文本 |
de_baseline |
基线提示生成的德语翻译 |
de_with_dictionary |
字典增强提示生成的德语翻译 |
predicted_variety |
自动预测的罗曼什语变体 |
dataset |
来源数据集名称 |
metadata |
来源数据集的附加元数据 |
license |
来源数据集的许可证 |
数据来源
该数据集包含以下公开罗曼什语数据源的回译结果：
| 来源名称 | 包含的变体 | 许可证 |
|---|---|---|
| FineWeb2 | RG | ODC-By 1.0 |
| La Quotidiana (1997–2008, 2021–2025) | RG, Surs., Suts., Surm., Puter, Vall. | CC BY 4.0 |
| FinePDFs | RG | ODC-By 1.0 |
| Mediomatix (unaligned) | Surs., Suts., Surm., Puter, Vall. | CC BY-NC-SA 4.0 |
| FineWiki | RG | CC BY-SA 4.0 |
| Theater plays | RG, Surs., Suts., Surm., Puter, Vall. | CC0 1.0 |
| Municipal documents | Surs., Suts., Surm., Vall. | CC0 1.0 |
| Historical Dictionary of Switzerland | RG | CC BY-SA 4.0 |
核心特点
- 使用 Gemini 2.5 Flash 合成生成德语翻译，利用大语言模型从罗曼什语翻译为德语的优势（即低资源→高资源方向增强）。
- 研究表明，这种增强方法在德国→罗曼什语翻译任务上，可使最低资源变体提升最多 +23 BLEU。
- 每个罗曼什文本的变体标签由
ZurichNLP/romansh-lid自动预测。
使用方式
python from datasets import load_dataset
dataset = load_dataset("ZurichNLP/romansh-backtranslated")
相关论文
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 · 论文地址
引用格式
bibtex @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} }
