gouqi666 / RAST

RAST

This repository contains code for the paper Diversify Question Generation with Retrieval-Augmented Style Transfer

• we provide our processed_data in data_link.
• we also provide our model checkpoint in checkpoint_link.
• if you use our repository, please cite paper. If you find this code useful in your research, please consider citing:

@misc{gou2023diversify,
      title={Diversify Question Generation with Retrieval-Augmented Style Transfer}, 
      author={Qi Gou and Zehua Xia and Bowen Yu and Haiyang Yu and Fei Huang and Yongbin Li and Nguyen Cam-Tu},
      year={2023},
      eprint={2310.14503},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

To reproduce

1. download data

• squad1.1, zhou split
  • This split of squad refers to Neural Question Generation from Text- A Preliminary Study
  • data num of train/dev/test is 86,635/8,965/8,964 respectively.
• squad1.1, du split
  • This split of squad refers to Learning to Ask: Neural Question Generation for Reading Comprehension
  • data num of train/dev/test is 70484/10570/11877 respectively.
• newsqa
  • This dataset refers to NewsQA: A Machine Comprehension Dataset
  • data num of train/dev/test is 92549/5166/5126 respectively.

2. process data

• process original data

python data/process_data.py  
refer to  data/readme.md  

• convert and store corpus data into faiss vector

python rast/rag/prepare_dataset.py
refer to rast/rag/prepare_dataset.py

3. train generator with skeleton

refer to rast/qg/readme.md 

4. train vanilla generator

refer to rast/qg/readme.md 

5. train QA model

refer to rast/reward_mdoel/T5_QA/readme.md

6. train rag

refer to rast/rag/readme_v100.md