Adversarial Ranking for Language Generation

Introduction

This is a tensorflow implementation of Adversarial Ranking for Language Generation by Kevin Lin^*, Dianqi Li^*, Xiaodong He, Zhengyou Zhang, Ming-Ting Sun, NIPS 2017.

Environment

The code is based on python2.7 and tensorflow 1.2 version. The code is developed and tested using one NVIDIA M40 GPU.

Run

python main.py

Note

• For a fair comparison, we used same LSTMs units, pre-train and test configurations in SeqGan. If you use tf.contrib.rnn.LSTMCell instead of their LSTMs implementation, you will obtain different training results.
• save/target_params.pkl is the parameter for the oracle model from SeqGan. log folder stores the log of your model training.
• Ideally, you will receive a nll loss between 8.00-8.50. However, adversarial training sometimes depends on the quality of pre-train model.
• More evaluation metrics for adversarial text generation can be refered to: paper and repo.

Citing RankGan

if you find RankGan is useful in your research, please consider citing:

@inproceedings{lin2017adversarial,
  title={Adversarial ranking for language generation},
  author={Lin, Kevin and Li, Dianqi and He, Xiaodong and Zhang, Zhengyou and Sun, Ming-Ting},
  booktitle={Advances in Neural Information Processing Systems},
  pages={3155--3165},
  year={2017}
}

Acknowledgements

This code is based on SeqGan. Many thanks for the authors!