TextGAIL: Generative Adversarial Imitation Learning for Text Generation

This is the repository for the paper: TextGAIL: Generative Adversarial Imitation Learning for Text Generation

Requirements

TorchFly (https://github.com/qywu/TorchFly) is needed for all dependencies. It is included in this repository.

Datasets

To run the code, You need to download the following datasets, and store them under the folder data.

The datasets used in the paper can be found below:

• COCO: https://github.com/geek-ai/Texygen/tree/master/data

• EMNLP2017 NEWS: https://github.com/geek-ai/Texygen/tree/master/data

• ROCStories: https://cs.rochester.edu/nlp/rocstories/

• CommonGEN: https://github.com/INK-USC/CommonGen

Then run the pre-processing script to get train.jsonl, valid.jsonl, test.jsonl for each dataset. The details can be found in the data folder.

Training

Pre-trained Weights

We use GPT-2 small (117M) and RoBERTa base (117M) as the pre-trained model weights, which are automatically downloaded and loaded.

Scripts

To obtain the best MLE models, please run scripts in the Conditional/MLE folder or Conditional/MLE first.

cd Conditional/MLE
#  choose the task.name in {CommonGEN, DailyDialog}
python main.py --config config/config.yaml task.name=CommonGEN

To train TextGAIL, please specify the MLE model after warm-up training. Then simply run the following command.

python main.py --config config/config.yaml task.name=CommonGEN task.weights_path="../../../MLE/outputs/CommonGEN/Checkpoints/iter_252_model_state.pth"

All hyper-paramters are stored in YAML format and will be automatically loaded by the script. You can adjust them according to your task.

Evaluation

Please run Generation.ipynb to have the geenerated outputs before the evaluation.

Conditional generation tasks and unconditional generation tasks have different evaluation metrics. Therefore, please refer to the Evaluation folder for details.

We have provided generated outputs along with the generated outputs for evaluation.

Citation

@inproceedings{wu2021textgail,
  author    = {Qingyang Wu and
               Lei Li and
               Zhou Yu},
  title     = {TextGAIL: Generative Adversarial Imitation Learning for Text Generation},
  booktitle = {The Thirty-Fifth {AAAI} Conference on Artificial Intelligence, {AAAI} 2021},
  pages     = {online},
  publisher = {{AAAI} Press},
  year      = {2021},
}

Contact

Feel free to contact me (wilwu@ucdavis.edu) if you have any questions.