chijames / bookqa

Book QA

Codes and resutls for the TACL paper Narrative Question Answering with Cutting-Edge Open-Domain QATechniques: A Comprehensive Study.

Installation

The implementation is based on huggingface transformer 3.2 and our customizations are directly made to the original source code. We provide the exact version in transformers folder that corresponds to the results in the paper. The following scripts have only been tested on Ubuntu 18.04 with Python 3.6 & 3.7.

It is noteworthy that, based on our tests, our customization based on huggingface transformer 4.7 can provide better results. We will provide the code in the future.

conda create -n trans32 python=3.7
conda activate trans32
cd transformers
pip install -e ".[testing]"
conda install pytorch torchvision cudatoolkit=10.2 -c pytorch
pip install -r ./examples/requirements.txt
pip install pytorch-lightning==0.8.5

For evaluation, more details can be found in evaluation/ folder

Citation

If you find this repo useful, please consider citing our paper:

@article{mou2021narrative,
  title={Narrative Question Answering with Cutting-Edge Open-Domain QA Techniques: A Comprehensive Study},
  author={Mou, Xiangyang and Yang, Chenghao and Yu, Mo and Yao, Bingsheng and Guo, Xiaoxiao and Potdar, Saloni and Su, Hui},
  journal={arXiv preprint arXiv:2106.03826},
  year={2021}
}