This is an implementation of our paper "KART: Parameterization of Privacy Leakage Scenarios from Pre-trained Language Models."

• Python 3.6.4
• Make sure that $HOME is set to environment variable $PYTHONPATH.

We simulate privacy leakage from clinical records using MIMIC-III-dummy-PHI

MIMIC-III-dummy-PHI is made by embedding pieced of dummy protected health information (PHI) in MIMIC-III corpus.

To install using venv module, use the following commands:

# Clone Repository
cd ~
git clone [email protected]:yutanakamura-tky/kart.git
cd ~/kart

# Install dependencies
python -m venv .venv
source .venv/source/activate
pip install -r requirements.txt

To install using Poetry, use the following commands:

# Install Poetry
curl -sSL https://raw.githubusercontent.com/python-poetry/poetry/master/get-poetry.py > ~/get-poetry.py
cd ~
python get-poetry.py --version  1.1.4
poetry config virtualenvs.in-project true

# Clone Repository
cd ~
git clone [email protected]:yutanakamura-tky/kart.git
cd ~/kart

# Activate virtual environment & install dependencies
poetry shell
poetry install

First, get NOTEEVENTS.csv.gz in MIMIC-III version 1.4 here and place it in corpus directory.

mv /path/to/NOTEEVENTS.csv.gz ~/kart/corpus
cd ~/kart/corpus
gunzip NOTEEVENTS.csv.gz

Then run make_mimic_iii_dummy_phi.sh. Make sure that you are in the virtual environment:

cd ~/kart/src
bash make_mimic_iii_dummy_phi.sh

cd ~/kart/src
bash make_pretraining_data.sh

To pre-train BERT model from scratch, use this command:

cd ~/kart/src
bash pretrain_bert_from_scratch.sh

To pre-train BERT model from BERT-base-uncased model, use this command:

cd ~/kart/src
# Download BERT-base-uncased model by Google Research
bash get_google_bert_model.sh
bash pretrain_bert_from_bert_base_uncased.sh

Please cite our arXiv preprint:

@misc{kart,
Author = {Yuta Nakamura and Shouhei Hanaoka and Yukihiro Nomura and Naoto Hayashi and Osamu Abe and Shuntaro Yada and Shoko Wakamiya and Eiji Aramaki},
Title = {KART: Parameterization of Privacy Leakage Scenarios from Pre-trained Language Models},
Year = {2020},
Eprint = {arXiv:2101.00036},
}