Please note that the health information generated by the chatbot is for general research purposes only. It is not a diagnostic tool, nor is it a substitute for medical advice or treatment for specific conditions.

Our work has been accepted at ACM-BCB. The paper is available here: https://dl.acm.org/doi/abs/10.1145/3388440.3412413.

The dataset is the initial commercial use subset taken from COVID-19 Open Research Dataset (CORD-19) and consists of 9000 scholarly articles.

For the training purposes, we have extracted the abstract and the main body of these articles and have merged them together.

For (re)extracting the data, run the command below.

python3 extract.py

We have retrained GPT-2 774M model with the COVID-19 corpus.

The model was trained using the Adam optimizer with the learning rate of 0.0001. It went through 2500 iterations with the batch size of 8.

The model is hosted and available at this Google Drive link.

Once you have downloaded the model, and put the model inside the models directory. Afterward, to run the chatbot, execute the sequence of commands shown below.

git clone https://github.com/oniani/covid-19-chatbot
cd covid-19-chatbot
python3 -m pip install -r requirements.txt
PYTHONPATH=src python3 -W ignore interact.py

For running the web application, navigate to the web-app directory and run flask run (it should be hosted on port 5000, usually).

It is also possible to run transfer learning with your own data.

Google Colaboratory (re)training example:

# Mount the drive
from google.colab import drive
drive.mount("/content/drive")

# Set up the repository
%cd "/content/drive/My Drive"
!mkdir COVID-19_CHATBOT
!rm -rf gpt-2
!git clone https://github.com/oniani/gpt-2 "/content/drive/My Drive/COVID-19_CHATBOT/gpt-2/"
%cd COVID-19_CHATBOT/gpt-2/

# Install the pretrained model and its dependencies
!python3 -m pip install -r requirements.txt
!python3 download_model.py 774M

# Install additional dependencies
!python3 -m pip install fire==0.2.1 \
                        tensorflow-gpu==1.14 \
                        tensorflow-hub==0.7.0 \
                        toposort==1.5

# Run the transfer learning training
#
# NOTE: You will need to upload `data` folder from this repository and put it
# into the `COVID-19_CHATBOT` directory
!PYTHONPATH=src python3 train.py --dataset="/content/drive/My Drive/COVID-19_CHATBOT/data" \
                                 --model_name=774M \
                                 --batch_size=8 \
                                 --optimizer=adam \
                                 --learning_rate=0.0001 \
                                 --save_time=-1 \
                                 --sample_every=-1 \
                                 --save_every=500 \
                                 --init_tpu

Special thanks to @shawwn for making GPT-2 TPU-trainable on Google Colaboratory.

The results are also available in a single file here.

The work we would like to see in the future includes retraining the model with a different dataset, tweaking the hyperparameters, and/or applying the larger GPT-2 model (1.5B parameters).

The project was developed by David Oniani and Dr. Yanshan Wang.