Datasets:
license: mit
dataset_info:
features:
- name: tokens
sequence: string
- name: annotated_labels
sequence: int64
- name: annotated_labels_max
sequence: int64
- name: file
dtype: string
splits:
- name: train
num_bytes: 4977276
num_examples: 12423
download_size: 952960
dataset_size: 4977276
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
task_categories:
- token-classification
language:
- fr
tags:
- medical
- NER
pretty_name: Dataset for NER in medical trials eligibility criteria in french.
Dataset Card for French Clinical Trials NER Dataset
Dataset Details
Dataset Description
This dataset provides French translations and annotations for Named Entity Recognition (NER) in clinical trials' eligibility criteria. It was generated as part of the final project of the Hands-On NLP course at Université Paris-Saclay M1-AI. The dataset builds on the English CHIA dataset by using a cross-lingual approach, including neural machine translation and fine-tuning a multilingual model for entity annotation.
Entities are annotated using the BIO format, focusing on commonly represented entities like Condition, Value, Drug, Procedure, Measurement, Temporal, Observation, and Person.
Dataset Sources
- Github Repository: GitHub Repository
- Original dataset link: CHIA Dataset
- Original dataset presentation paper: Transformer-Based Named Entity Recognition for Parsing Clinical Trial Eligibility Criteria
Uses
Direct Use
The dataset is designed for training and evaluating NER models in French for clinical trials. Potential applications include text parsing, clinical trial management systems, and medical information extraction.
Dataset Structure
Features
- tokens: List of tokens in a sentence.
- annotated_labels: Entity labels following BIO format aligned with tokens.
- annotated_labels_max: Entity labels aggregated by subword annotations.
- file: Filename corresponding to the clinical trial source.
Labels:
sel_ent = {
"O": 0,
"B-Condition": 1,
"I-Condition": 2,
"B-Value": 3,
"I-Value": 4,
"B-Drug": 5,
"I-Drug": 6,
"B-Procedure": 7,
"I-Procedure": 8,
"B-Measurement": 9,
"I-Measurement": 10,
"B-Temporal": 11,
"I-Temporal": 12,
"B-Observation": 13,
"I-Observation": 14,
"B-Person": 15,
"I-Person": 16
}
Splits
The dataset contains a single split for training, with 12,423 examples.
Dataset Creation
Curation Rationale
The dataset was created to enable NER in French for clinical trials, leveraging the cross-lingual capabilities of XLM-RoBERTa and state-of-the-art neural machine translation models.
Source Data
- Original Dataset: CHIA dataset of clinical trial criteria.
- Translation Process: Sentences were translated using Helsinki-NLP's opus-mt-en-fr model.
Annotation Process
Annotations were generated using a fine-tuned XLM-RoBERTa model trained on the English CHIA dataset. The pipeline ensures consistency in BIO format even for subword tokenization.
Personal and Sensitive Information
The dataset does not include personal identifiers. It represents anonymized text from clinical trial eligibility criteria.
Bias, Risks, and Limitations
The dataset relies on machine-translated text and model-generated annotations, which may introduce biases or inaccuracies. It is recommended to evaluate model predictions against human-annotated benchmarks for critical applications.