/ChainCQG

Primary LanguagePython

 

ChainCQG: Flow Aware Conversational Question Generation

Overview

ChainCQG is a two-stage architecture that learns question-answer representations across multiple dialogue turns using a flow propagation training strategy.

Reproduction

  1. First we need to download the coqa dataset from here, then process it from Question Answer format into Question Generation format. It should be placed in the /data folder.

  2. We release both ChainCQG and other models benchmarked in the paper. For ChainCQG, please use run_generation_coqa_chaincqg.sh. For other models such as t5 or bart, please refer to the /OtherModel folder. Changing hyperparameter inside the script should be enough to try other models and settings.

if you find our work useful, please cite:

@misc{gu2021chaincqg,
      title={ChainCQG: Flow-Aware Conversational Question Generation}, 
      author={Jing Gu and Mostafa Mirshekari and Zhou Yu and Aaron Sisto},
      year={2021},
      eprint={2102.02864},
      archivePrefix={arXiv},
      primaryClass={cs.AI}
}