PxCorpus: A Spoken Drug Prescription Dataset in French for Spoken Language Understanding

PxSLU : A Spoken Drug Prescription Dataset in French for Spoken Language Understanding

PxSLU is to the best of our knowledge, the first spoken medical drug prescriptions corpus to be distributed. It contains 4 hours of transcribed and annotated dialogues of drug prescriptions in French acquired through an experiment with 55 participants experts and non-experts in drug prescriptions.

The automatic transcriptions were verified by human effort and aligned with semantic labels to allow training of NLP models. The data acquisition protocol was reviewed by medical experts and permit free distribution without breach of privacy and regulation.

Overview of the Corpus

The experiment has been performed in wild conditions with naive participants and medical experts. In total, the dataset includes 1981 recordings of 55 participants (38% non-experts, 25% doctors, 36% medical practitioners), manually transcribed and semantically annotated.

License
We hope that that the community will be able to benefit from the dataset which is distributed with an attribution 4.0 International (CC BY 4.0) Creative Commons license.

How to cite this corpus

If you use the corpus or need more details please refer to the following paper: A spoken drug prescription dataset in French for spoken Language Understanding

@InProceedings{Kocabiyikoglu2022,
  author =     "Alican Kocabiyikoglu and Fran{\c c}ois Portet and Prudence Gibert and Hervé Blanchon and Jean-Marc Babouchkine and Gaëtan Gavazzi",
  title =     "A Spoken Drug Prescription Dataset in French for Spoken Language Understanding",
  booktitle =     "13th Language Resources and Evaluation Conference (LREC 2022)",
  year =     "2022",
  location =     "Marseille, France"
}

Project Structure

The projet contains the following elements

.
├── LICENSE
├── metadata.txt
├── paths.txt
├── PxSLU_conll.txt
├── readme.md
├── recordings
├── seq.in
├── seq.label
├── seq.out
├── Demo.ipynb
└── verifications.py

Recordings

The recordings directory contains the 903 recording sessions. Each session can contain several recordings. For instance,
the directory
    recordings/J7aVvWb67L
contains the records    
    recording_0.wav  recording_2.wav
which represent two attempts to record a drug prescription

All records are stored as mono channel wav files of 16kHz 16bits signed PCM

Paths

contains the list of all the wav files in the recordings directory
00MYcyVK0t/recording_0.wav
00MYcyVK0t/recording_2.wav
02Qp6ICj9Q/recording_0.wav
02Qp6ICj9Q/recording_1.wav
...

All other files (metadata.txt, seq.*) refer to this list to describe the recording.

Metadata

contains the information about the participants:

48,60+,F,non-expert
48,60+,F,non-expert
24,18–28,F,doctor
24,18–28,F,doctor
...

The first column is the participant unique id, the second is the age range, the third is the gender and the final is the category of the participant in {doctor,expert, non-expert}.

doctor correspond to a physician, (other)expert to a pharmacist or a biologist specialized in drugs while non-expert are other people not entering in these categories.

The lines are synchronised with the paths.txt lines.

Labels

the three files seq.label, seq.in, seq.out represent respectively the intent, the transcript and the entities in BIO format.

These lines are synchronised with the paths.txt lines.

Another file "PxSLU_conll.txt" is provided in a format inspired by the conll format (https://universaldependencies.org/format.html). However, this one is *not* aligned with the acoustic records file paths.txt.

Scripts

verifications.py performs the checking of the alignement of all the seq.* paths.txt and metadata files. A user of the dataset does not need to use this script unless she plan to extend the datasets with her own data.

Demo.ipynb is a jupyter notebook that a user can run to search through the dataset. It is intended to let the user have a quicker and smoother view on the dataset.

Data splits

In the data_splits folder, you can find a data split of this dataset organized as following:

- train.txt: medical experts + non experts (80%) = 1128 samples
- dev.txt: medical experts + non experts (20%) = 283 samples
- test.txt: doctors (100%) = 570 samples

Each file contains references to line numbers of the corpus. For example, first line of the test.txt is 904, seq.in file contains the utterance "nicopatch". Users can access the labels, slots, metadata using the same line number 904 in the parallel files (paths.txt,seq.out,seq.label,...).