---
license: cc-by-sa-4.0
dataset_info:
- config_name: de-enGB
  features:
  - name: source
    dtype: large_string
  - name: translation_A
    dtype: large_string
  - name: translation_B
    dtype: large_string
  - name: A
    dtype: bool
  - name: equal
    dtype: bool
  - name: B
    dtype: bool
  - name: label_A
    dtype: large_string
  - name: label_B
    dtype: large_string
  - name: text
    dtype: large_string
  - name: text_type
    dtype: large_string
  splits:
  - name: train
    num_bytes: 99615
    num_examples: 390
  download_size: 51640
  dataset_size: 99615
- config_name: de-frCH
  features:
  - name: source
    dtype: large_string
  - name: translation_A
    dtype: large_string
  - name: translation_B
    dtype: large_string
  - name: A
    dtype: bool
  - name: equal
    dtype: bool
  - name: B
    dtype: bool
  - name: label_A
    dtype: large_string
  - name: label_B
    dtype: large_string
  - name: text
    dtype: large_string
  - name: text_type
    dtype: large_string
  splits:
  - name: train
    num_bytes: 106345
    num_examples: 385
  download_size: 55015
  dataset_size: 106345
- config_name: de-itCH
  features:
  - name: source
    dtype: large_string
  - name: translation_A
    dtype: large_string
  - name: translation_B
    dtype: large_string
  - name: A
    dtype: bool
  - name: equal
    dtype: bool
  - name: B
    dtype: bool
  - name: label_A
    dtype: large_string
  - name: label_B
    dtype: large_string
  - name: text
    dtype: large_string
  - name: text_type
    dtype: large_string
  splits:
  - name: train
    num_bytes: 102833
    num_examples: 378
  download_size: 54128
  dataset_size: 102833
- config_name: en-deCH
  features:
  - name: source
    dtype: large_string
  - name: translation_A
    dtype: large_string
  - name: translation_B
    dtype: large_string
  - name: A
    dtype: bool
  - name: equal
    dtype: bool
  - name: B
    dtype: bool
  - name: label_A
    dtype: large_string
  - name: label_B
    dtype: large_string
  - name: text
    dtype: large_string
  - name: text_type
    dtype: large_string
  splits:
  - name: train
    num_bytes: 99510
    num_examples: 330
  download_size: 49779
  dataset_size: 99510
configs:
- config_name: de-enGB
  data_files:
  - split: train
    path: de-enGB/train-*
- config_name: de-frCH
  data_files:
  - split: train
    path: de-frCH/train-*
- config_name: de-itCH
  data_files:
  - split: train
    path: de-itCH/train-*
- config_name: en-deCH
  data_files:
  - split: train
    path: en-deCH/train-*
task_categories:
- translation
language:
- en
- de
- it
- fr
tags:
- Supertext
- DeepL
- Translation
- A/B-test
pretty_name: A/B Test Supertext vs DeepL
size_categories:
- 1K<n<10K
---
# A/B Test Supertext vs DeepL

We release all evaluation data and scripts for further analysis and reproduction of the accompanying paper: [A comparison of translation performance between DeepL and Supertext](https://arxiv.org/abs/2502.02577).
The data consists of document-level translations by Supertext and DeepL as well as accompanying ratings by professional translators. Please find more details in the paper.

Please note that the empty lines correspond to paragraph boundaries (i.e., double line breaks) in the original documents.

``` python
# for each language pair, there is a separate subset
data = load_dataset("Supertext/mt-doclevel-ab-test", "en-deCH")
```

## Dataset Details

As strong machine translation (MT) systems are increasingly based on large language models (LLMs), reliable quality benchmarking requires methods that capture their ability to leverage extended context. This study compares two commercial MT systems -- DeepL and Supertext -- by assessing their performance on unsegmented texts. We evaluate translation quality across four language directions with professional translators assessing segments with full document-level context. While segment-level assessments indicate no strong preference between the systems in most cases, document-level analysis reveals a preference for Supertext in three out of four language directions, suggesting superior consistency across longer texts. We advocate for more context-sensitive evaluation methodologies to ensure that MT quality assessments reflect real-world usability.

### Citation
If you use any of the data released in our work, please cite the following paper:

```
@misc{flückiger2025comparisontranslationperformancedeepl,
      title={A comparison of translation performance between DeepL and Supertext}, 
      author={Alex Flückiger and Chantal Amrhein and Tim Graf and Frédéric Odermatt and Martin Pömsl and Philippe Schläpfer and Florian Schottmann and Samuel Läubli},
      year={2025},
      eprint={2502.02577},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2502.02577}, 
}
```
### Dataset Description


- **Curated by:** Supertext
- **Language(s) (NLP):** English, French, German, Italian
- **License:** CC BY-SA 4.0

### Dataset Sources

- **Repository:** https://github.com/Supertext/evaluation_deepl_supertext
- **Paper:** https://arxiv.org/abs/2502.02577