The source code of the paper A Generative Model for Joint Natural Language Understanding and Generation published at ACL 2020.

@inproceedings{tseng2020generative,
  title={A Generative Model for Joint Natural Language Understanding and Generation},
  author={Tseng, Bo-Hsiang and Cheng, Jianpeng and Fang, Yimai and Vandyke, David},
  booktitle={Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics},
  pages={1795--1807},
  year={2020}
}

python 3
torch 1.1.0
numpy 1.13.3
nltk 3.4.5

The two used dataset are in data/ folder with different amounts of training data.

Before training, please create the folders with the following command:

bash create_folders.sh

To train the model, use the script train.sh with the following command:

bash train.sh $dataset $data_ratio $batch_size $model_dimension $seed

• dataset: e2e or weather
• data ratio: 5 / 10 / 25 / 50 / 100
• batch size: 32 / 64
• model dimension: 150 / 300

To test the model, use the script test.sh with the following command:

bash test.sh $dataset $data_ratio $batch_size $model_dimension $model_name

• dataset: e2e or weather
• data ratio: 5 / 10 / 25 / 50 / 100
• batch size: 32 / 64
• model dimension: 150 / 300
• model name: name of a trained model