GPTNERMED

About

GPTNERMED is a novel open synthesized dataset and neural named-entity-recognition (NER) model for German texts in medical natural language processing (NLP).

Key features:

Supported labels: Medikation, Dosis, Diagnose
Open silver-standard German medical dataset: 245107 tokens with annotations for Dosis (#7547), Medikation (#9868) and Diagnose (#5996)
Synthesized dataset based on GPT NeoX
Transfer-learning for NER parsing using gbert-large, GottBERT-base or German-MedBERT
Open, public access to models

Online Demo: A demo page is available: Demo, or use the HuggingFace links given below.

See our published paper at https://doi.org/10.1016/j.jbi.2023.104478.

Our pre-print paper is available at https://arxiv.org/pdf/2208.14493.pdf.

NER demonstration:

Models

The pretrained models can be retrieved from the following URLs:

gbert-based: model link
GottBERT-based: model link
German-MedBERT-based: model link

The models are also available on the HuggingFace platform:

gbert-based: HuggingFace link
GottBERT-based: HuggingFace link
German MedBERT-based: HuggingFace link

HuggingFace Dataset: The dataset is also available as a HuggingFace Dataset.
You can load the model as follows:

# You need to install datasets first, using: pip install datasets
from datasets import load_dataset
dataset = load_dataset("jfrei/GPTNERMED")

Scores

Note: Metric scores are evaluated by character-wise classification.

Out of Distribution Dataset (provided in OoD-dataset_GoldStandard.jsonl):

Model	Metric	Drug = Medikation
gbert-large	Pr	0.707
	Re	0.979
	F1	0.821
GottBERT-base	Pr	0.800
	Re	0.899
	F1	0.847
German-MedBERT	Pr	0.727
	Re	0.818
	F1	0.770

Test Set:

Model	Metric	Medikation	Diagnose	Dosis	Total
gbert-large	Pr	0.870	0.870	0.883	0.918
	Re	0.936	0.895	0.921	0.919
	F1	0.949	0.882	0.901	0.918
GottBERT-base	Pr	0.979	0.896	0.887	0.936
	Re	0.910	0.844	0.907	0.886
	F1	0.943	0.870	0.897	0.910
German-MedBERT	Pr	0.980	0.910	0.829	0.932
	Re	0.905	0.730	0.890	0.842
	F1	0.941	0.810	0.858	0.883

Setup and Usage

The models are based on SpaCy. The sample code is written in Python.

model_link="https://myweb.rz.uni-augsburg.de/~freijoha/GPTNERMED/GPTNERMED_gbert.zip"

# [Optional] Create env
python3 -m venv env
source ./env/bin/activate

# Install dependencies
python3 -m pip install -r requirements.txt

# Download & extract model
wget -O model.zip "$model_link"
unzip model.zip -d "model"

# Run script
python3 GPTNERMED.py

Citation

Cite our work with BibTex as written below or use the citation tools from the paper.

@article{FREI2023104478,
title = {Annotated dataset creation through large language models for non-english medical NLP},
journal = {Journal of Biomedical Informatics},
volume = {145},
pages = {104478},
year = {2023},
issn = {1532-0464},
doi = {https://doi.org/10.1016/j.jbi.2023.104478},
url = {https://www.sciencedirect.com/science/article/pii/S1532046423001995},
author = {Johann Frei and Frank Kramer},
keywords = {Natural language processing, Information extraction, Named entity recognition, Data augmentation, Knowledge distillation, Medication detection},
abstract = {Obtaining text datasets with semantic annotations is an effortful process, yet crucial for supervised training in natural language processing (NLP). In general, developing and applying new NLP pipelines in domain-specific contexts for tasks often requires custom-designed datasets to address NLP tasks in a supervised machine learning fashion. When operating in non-English languages for medical data processing, this exposes several minor and major, interconnected problems such as the lack of task-matching datasets as well as task-specific pre-trained models. In our work, we suggest to leverage pre-trained large language models for training data acquisition in order to retrieve sufficiently large datasets for training smaller and more efficient models for use-case-specific tasks. To demonstrate the effectiveness of your approach, we create a custom dataset that we use to train a medical NER model for German texts, GPTNERMED, yet our method remains language-independent in principle. Our obtained dataset as well as our pre-trained models are publicly available at https://github.com/frankkramer-lab/GPTNERMED.}
}