README.md

---
license: other
license_name: other
license_link: https://huggingface.co/datasets/microsoft/wiki_qa#licensing-information
task_categories:
- question-answering
language:
- en
- fr
- de
- it
- es
- pt
pretty_name: mWikiQA
size_categories:
- 100K<n<1M
configs:
- config_name: en
  data_files:
  - split: train
    path: "eng-train.jsonl"
  - split: validation
    path: "eng-dev.jsonl"
  - split: test
    path: "eng-test.jsonl"
- config_name: de
  data_files:
  - split: train
    path: "deu-train.jsonl"
  - split: validation
    path: "deu-dev.jsonl"
  - split: test
    path: "deu-test.jsonl"
- config_name: fr
  data_files:
  - split: train
    path: "fra-train.jsonl"
  - split: validation
    path: "fra-dev.jsonl"
  - split: test
    path: "fra-test.jsonl"
- config_name: it
  data_files:
  - split: train
    path: "ita-train.jsonl"
  - split: validation
    path: "ita-dev.jsonl"
  - split: test
    path: "ita-test.jsonl"
- config_name: po
  data_files:
  - split: train
    path: "por-train.jsonl"
  - split: validation
    path: "por-dev.jsonl"
  - split: test
    path: "por-test.jsonl"
- config_name: sp
  data_files:
  - split: train
    path: "spa-train.jsonl"
  - split: validation
    path: "spa-dev.jsonl"
  - split: test
    path: "spa-test.jsonl"
- config_name: en_++
  data_files:
  - split: train
    path: "eng-train.jsonl"
  - split: validation
    path: "eng-dev_no_allneg.jsonl"
  - split: test
    path: "eng-test_no_allneg.jsonl"
- config_name: de_++
  data_files:
  - split: train
    path: "deu-train.jsonl"
  - split: validation
    path: "deu-dev_no_allneg.jsonl"
  - split: test
    path: "deu-test_no_allneg.jsonl"
- config_name: fr_++
  data_files:
  - split: train
    path: "fra-train.jsonl"
  - split: validation
    path: "fra-dev_no_allneg.jsonl"
  - split: test
    path: "fra-test_no_allneg.jsonl"
- config_name: it_++
  data_files:
  - split: train
    path: "ita-train.jsonl"
  - split: validation
    path: "ita-dev_no_allneg.jsonl"
  - split: test
    path: "ita-test_no_allneg.jsonl"
- config_name: po_++
  data_files:
  - split: train
    path: "por-train.jsonl"
  - split: validation
    path: "por-dev_no_allneg.jsonl"
  - split: test
    path: "por-test_no_allneg.jsonl"
- config_name: sp_++
  data_files:
  - split: train
    path: "spa-train.jsonl"
  - split: validation
    path: "spa-dev_no_allneg.jsonl"
  - split: test
    path: "spa-test_no_allneg.jsonl"
- config_name: en_clean
  data_files:
  - split: train
    path: "eng-train.jsonl"
  - split: validation
    path: "eng-dev_clean.jsonl"
  - split: test
    path: "eng-test_clean.jsonl"
- config_name: de_clean
  data_files:
  - split: train
    path: "deu-train.jsonl"
  - split: validation
    path: "deu-dev_clean.jsonl"
  - split: test
    path: "deu-test_clean.jsonl"
- config_name: fr_clean
  data_files:
  - split: train
    path: "fra-train.jsonl"
  - split: validation
    path: "fra-dev_clean.jsonl"
  - split: test
    path: "fra-test_clean.jsonl"
- config_name: it_clean
  data_files:
  - split: train
    path: "ita-train.jsonl"
  - split: validation
    path: "ita-dev_clean.jsonl"
  - split: test
    path: "ita-test_clean.jsonl"
- config_name: po_clean
  data_files:
  - split: train
    path: "por-train.jsonl"
  - split: validation
    path: "por-dev_clean.jsonl"
  - split: test
    path: "por-test_clean.jsonl"
- config_name: sp_clean
  data_files:
  - split: train
    path: "spa-train.jsonl"
  - split: validation
    path: "spa-dev_clean.jsonl"
  - split: test
    path: "spa-test_clean.jsonl"
---
## Dataset Description

**mWikiQA** is a translated version of WikiQA. It contains 3,047 questions sampled from Bing query logs. The candidate answer sentences are extracted from Wikipedia and then manually labeled to assess whether they are correct answers.

The dataset has been translated into five European languages: French, German, Italian, Portuguese, and Spanish, as described in this paper: "Datasets for Multilingual Answer Sentence Selection."

## Splits:

For each language (English, French, German, Italian, Portuguese, and Spanish), we provide:

- **train** split
- **validation** split
- **test** split

In addition, the validation and the test splits are available also in the following preprocessed versions:

- **++**: without questions with only negative answer candidates
- **clean**: without questions with only negative and only positive answer candidates

### How to load them:
To use these splits, you can use the following snippet of code replacing ``[LANG]`` with a language identifier (en, fr, de, it, po, sp), and ``[VERSION]`` with the version identifier (++, clean)

```
from datasets import load_dataset

"""
if you want the default splits, replace [LANG] with an identifier in: en, fr, de, it, po, sp
dataset = load_dataset("matteogabburo/mWikiQA", "[LANG]")
"""
# example:
italian_dataset = load_dataset("matteogabburo/mWikiQA", "it")


"""
if you want the processed splits ("clean" and "no all negatives" sets), replace [LANG] with a language identifier and [VERSION] with "++" or "clean"
dataset = load_dataset("matteogabburo/mWikiQA", "[LANG]_[VERSION]")
"""
# example:
italian_clean_dataset = load_dataset("matteogabburo/mWikiQA", "it_clean")

```


## Format:
Each example has the following format:

```
{
  'eid': 1214,
  'qid': 141,
  'cid': 0,
  'label': 1,
  'question': 'Was bedeutet Karma im Buddhismus?',
  'candidate': 'Karma (Sanskrit, auch karman, Pali: Kamma) bedeutet "Handlung" oder "Tun"; was auch immer man tut, sagt oder denkt, ist ein Karma.'
}
```

Where:

- **eid**: is the unique id of the example (question, candidate)
- **qid**: is the unique id of the question
- **cid**: is the unique id of the answer candidate
- **label**: identifies whether the answer candidate ``candidate`` is correct for the ``question`` (1 if correct, 0 otherwise)
- **question**: the question
- **candidate**: the answer candidate



## Citation

If you find this dataset useful, please cite the following paper:

**BibTeX:**
```
@misc{gabburo2024datasetsmultilingualanswersentence,
      title={Datasets for Multilingual Answer Sentence Selection}, 
      author={Matteo Gabburo and Stefano Campese and Federico Agostini and Alessandro Moschitti},
      year={2024},
      eprint={2406.10172},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2406.10172}, 
}
```