Dataset Card for FrBMedQA

FrBMedQA is the first French biomedical question answering dataset. It contains more than 41k automatically generated passage–question instances derived from French biomedical Wikipedia articles. The dataset was constructed using a cloze-style strategy where biomedical entities are masked and must be predicted using the surrounding context.

Dataset Details

Dataset Description

FrBMedQA is a French biomedical question answering (BQA) dataset designed to support research in machine reading comprehension and biomedical NLP for the French language. The dataset contains more than 41,000 passage–question instances, automatically generated from biomedical articles extracted from French Wikipedia.

Each instance consists of:

• a context passage
• a question
• a list of candidate biomedical entities
• the correct answer entity

The dataset follows a cloze-style formulation, where a biomedical entity in the question is replaced by a placeholder token and the task is to identify the correct entity from the candidate list appearing in the passage.

The dataset was created to address the lack of French biomedical QA datasets, which limits the development and evaluation of machine learning and deep learning models for biomedical question answering in French. :contentReference[oaicite:1]{index=1}

• Curated by: Zakaria Kaddari, Toumi Bouchentouf
• Funded by: Moroccan National Center for Scientific and Technical Research (CNRST)
• Language(s) (NLP): French
• License: CC BY 4.0

Dataset Sources

• Repository: (to be added when published on Hugging Face)
• Paper: FrBMedQA: the first French biomedical question answering dataset (IJ-AI 2022) :contentReference[oaicite:2]{index=2}

Uses

Direct Use

The dataset is intended for research in:

• biomedical question answering
• machine reading comprehension
• biomedical NLP for French
• evaluation of language models on biomedical reasoning
• training and benchmarking QA systems on biomedical corpora

FrBMedQA can be used to train and evaluate:

• neural machine reading comprehension models
• transformer-based language models
• statistical baseline models
• retrieval-based QA systems.

Out-of-Scope Use

The dataset should not be used for medical decision making or clinical diagnosis. The dataset contains automatically generated question-answer pairs and may include noise or simplified contexts. It is intended for research purposes only.

The dataset may also not be suitable for:

• conversational QA
• generative QA
• open-domain QA without context passages

since it follows a cloze-style multiple-choice formulation.

Dataset Structure

Each dataset instance is a tuple composed of:

• context: passage containing biomedical information
• question: a sentence where a biomedical entity is replaced by a placeholder
• candidates: list of biomedical entities appearing in the passage
• answer: correct biomedical entity

Example fields: { "context": "...", "question": "... @placeholder ...", "candidates": ["entity1","entity2","entity3"], "answer": "entity2" }

Dataset splits:

Split	Instances
Train	32,888
Validation	4,111
Test	4,110
Total	41,109

Average statistics:

• Average context length: ~111 tokens
• Average question length: ~38 tokens
• Average candidate entities: ~4.8 :contentReference[oaicite:3]{index=3}

Dataset Creation

Curation Rationale

The motivation for creating FrBMedQA was the lack of large-scale French biomedical question answering datasets, which limits research progress in French biomedical NLP.

While several biomedical QA datasets exist in English, none were available for French at the time of publication. FrBMedQA was introduced to fill this gap and enable training and evaluation of QA models for biomedical texts in French. :contentReference[oaicite:4]{index=4}

Source Data

The dataset was constructed from French Wikipedia biomedical articles.

More than 243,000 biomedical-related Wikipedia articles were retrieved using filtering strategies based on biomedical infoboxes and biomedical terminology. :contentReference[oaicite:5]{index=5}

Data Collection and Processing

The dataset construction pipeline included:

1. Retrieval of biomedical articles from French Wikipedia.
2. Cleaning of noisy content (references, HTML artifacts).
3. Splitting articles into paragraphs.
4. Biomedical entity recognition using the SIFR annotator.
5. Filtering entities belonging to selected UMLS semantic groups.
6. Replacement of biomedical entities with pseudo-tokens.
7. Generation of cloze-style question–answer instances.

Biomedical entities were replaced by tokens such as: @entity0 @entity1

A placeholder token (@placeholder) was inserted in the question to create the prediction task. :contentReference[oaicite:6]{index=6}

Who are the source data producers?

The original textual data comes from contributors to French Wikipedia, an open collaborative encyclopedia.

Biomedical entities were automatically annotated using the SIFR biomedical NER system.

Annotations

Annotation process

Biomedical entities were automatically detected using the SIFR annotator, a tool designed for semantic annotation of French biomedical texts.

Only entities belonging to selected UMLS semantic groups were considered, including:

• chemicals & drugs
• anatomy
• physiology
• disorders
• phenomena
• procedures
• genes & molecular sequences
• medical devices. :contentReference[oaicite:7]{index=7}

Who are the annotators?

Annotations were generated automatically using the SIFR biomedical entity recognition system rather than manual human annotation.

Personal and Sensitive Information

The dataset does not contain personal or private information. All texts originate from publicly available Wikipedia articles.

Bias, Risks, and Limitations

Several limitations should be considered:

• The dataset is automatically generated, which may introduce noise.
• Questions follow a cloze-style format, limiting diversity of question types.
• The candidate answer list can sometimes be small, allowing simple heuristics to perform well.
• The dataset is limited to Wikipedia biomedical articles, which may not fully represent clinical or scientific biomedical literature. :contentReference[oaicite:8]{index=8}

Recommendations

Users should treat the dataset as a benchmark dataset for research rather than a source of medical knowledge.

Models trained on FrBMedQA should not be used for clinical applications without further validation.

Quick Start

from datasets import load_dataset

dataset = load_dataset("ZakariaKaddari/FrBMedQA")

print(dataset["train"][0])

Citation

BibTeX

@article{kaddari2022frbmedqa,
  title={FrBMedQA: the first French biomedical question answering dataset},
  author={Kaddari, Zakaria and Bouchentouf, Toumi},
  journal={IAES International Journal of Artificial Intelligence},
  volume={11},
  number={4},
  pages={1588--1595},
  year={2022},
  doi={10.11591/ijai.v11.i4.pp1588-1595}
}