This is our first attempt to make state-of-the-art single-turn QA models conversational. Feel free to build on top of our code to build an even stronger conversational QA model.

For more details, please see: FlowQA: Grasping Flow in History for Conversational Machine Comprehension

perform the following:

pip install -r requirements.txt

to install all dependent python packages.

download necessary files using:

./download.sh

preprocess the data files using:

python preprocess_QuAC.py
python preprocess_CoQA.py

run the training code using:

python train_QuAC.py
python train_CoQA.py

For naming the output model, you can do

python train_OOOO.py --name XXX

Remove any answer marking by:

python train_OOOO.py --explicit_dialog_ctx 0

OOOO is the name of the dataset (QuAC or CoQA).

Do prediction with answer thresholding using

python predict_OOOO.py -m models_XXX/best_model.pt --show SS

XXX is the name you used during train.py.
SS is the number of dialog examples to be shown.
OOOO is the name of the dataset (QuAC or CoQA).