README.md

metadata

license: cc-by-4.0
language:
  - fr
tags:
  - inception
  - uima
  - annotation
task_categories:
  - token-classification
configs:
  - config_name: document_metadata
    data_files:
      - split: train
        path: data/document_metadata/document_metadata_train.parquet
      - split: dev
        path: data/document_metadata/document_metadata_dev.parquet
  - config_name: spans
    data_files:
      - split: train
        path: data/spans/spans_train.parquet
      - split: dev
        path: data/spans/spans_dev.parquet

Dataset Card for PARHAF-conclusion-annotated

Reporting Issues & Contributing

If you encounter any errors or inconsistencies in this dataset, please report them in the discussion section of the "Community" tab on Hugging Face.

For more substantial contributions or collaboration opportunities, feel free to contact us directly.

Dataset Description

• Points of Contact: NÉVÉOL Aurélie, ZANELLA Laura, ZWEIGENBAUM Pierre

Dataset Summary

PARHAF-conclusion-annotated is a subpart of the PARHAF corpus, an open French corpus of human-authored clinical reports of fictional patients.

It was designed for developing and evaluating NLP systems that generate and extract conclusion sections from French clinical reports. Each document is annotated with a conclusion span (character offsets into the full report text), identifying the conclusion section to be extracted.

This dataset contains training data only. The test corpora, that will be oficially used for the conclusion generation task, are currently under development. The test set will remain under embargo to enable future evaluations under controlled conditions, limiting the risk of LLM contamination through prior data exposure. Please contact us for access to the test data.

A subset of 108 documents (~3 %) from the training corpus was annotated manually and used to evaluate a rule-based automatic conclusion extraction pipeline. Teams are free to use the full training and evaluation sets or define a different train/dev split configuration. For more details about that use case, please contact us.

Each patient record was:

• written by a senior medical resident
• reviewed by another senior medical resident, from the same specialty
• automatically annotated by a rule-based system
• the evaluation subset was manually curated by an annotator

Data Statistics

DATASET SUMMARY

Indicator	Value
Complete Dataset
Number of JSON files	3772
Total annotations	7760
Average document length	3866.83 characters

Complete dataset distribution by type

Type	Annotations	% of total
DocumentMetadata	3880	50.00%
Span	3880	50.00%

Complete dataset distribution by populated field

Type.Field	Occurrences	% of total annot.
DocumentMetadata.Nomenclature	3880	50.00%
Span.Conclusion	3880	50.00%

Train / Dev Split Results

Indicator	TRAIN	DEV
Number of files	3664	108
Number of annotations	3664	108
Percentage of dataset	97,13%	2,86%
Avg. document length (chars)	3867	3854

Train / Dev distribution by type

Type	TRAIN	DEV	% train of type
DocumentMetadata	3664	108	97,13%
Span	3664	108	2.86%

Train / Dev distribution by type

Type.Field	TRAIN	DEV	% train of field
DocumentMetadata.Nomenclature	3664	108	97,13%
Span.Conclusion	3664	108	2.86%

Train / Dev distribution by value

Type.Field.Value	TRAIN	DEV
DocumentMetadata.Nomenclature.ConclusionSpan	3664	108
Span.Conclusion.Conclusion	3664	108

Data Origin

The clinical reports are extracted from the PARHAF corpus. Please refer to PARHAF documentation for more information about this corpus.

The documents were selected to represent a diversity of report types and specialties, following the distribution of reports with conclusions in the PARHAF corpus.

Languages

• fr_FR

Dataset Structure

We distribute both a Hugging Face dataset and a standalone version of the corpus. The standalone dataset consists of a JSON file per patient report, in UIMA CAS JSON format. This format constitutes the canonical version of the corpus.

Both formats therefore contain identical information and differ only in storage layout.

One dataset instance corresponds to one report.

Hugging Face dataset

This snippet shows how to extract and iterate over medical report information per patient using the datasets library.

import pandas as pd
from datasets import load_dataset

dfs = {cfg: load_dataset(
    "HealthDataHub/PARHAF-conclusion-annotated",
    cfg,
    split="train"
).to_pandas() for cfg in ["document_metadata", "spans"]}

for patient_raw in dfs["document_metadata"].itertuples():
    report_id = patient_raw.report
    text = patient_raw.full_text
    annot_type = patient_raw.annotation_type
    nomenclature = patient_raw.attribute_Nomenclature
    report_spans = dfs["spans"][dfs["spans"]["report"] == report_id]
    ...

Data Fields

Path	Type	Description	Possible values
document_metadata
report	string	Unique document identifier	—
full_text	string	Full text of the medical report	—
annotation_type	string	Annotation layer name	"DocumentMetadata"
attribute_Nomenclature	string	Document-level annotation label	"ConclusionSpan"
spans
report	string	Document identifier	—
span_id	integer	Annotation identifier	—
span_type	string	Annotation type name	"Span"
begin	integer	Start character offset in full_text	—
end	integer	End character offset in full_text	—
span_text	string	Text of the annotated conclusion	—
attribute_Conclusion	string	Annotation feature value	"Conclusion"

Data Splits

Only the training set is released here. The remaining portion of the corpus will be temporarily embargoed to enable future evaluations under controlled conditions, thereby limiting the risk of large language model contamination through prior exposure to the data. You can evaluate your system on the test set through the CodaBench platform.

Annotation Guidelines

You can find the detailed annotation protocol here: guidelines/guidelines.md

Licensing Information

• CC BY 4.0
• Etalab 2.0

Citation Information

[More Information Needed]