user-simulator

Codebase for How to Build User Simulators to Train RL-based Dialog Systems, published as a long paper in EMNLP 2019. The sequicity part is developed based on Sequicity: Simplifying Task-oriented Dialogue Systems with Single Sequence-to-Sequence Architectures.

If you use the datasets or any source codes included in this repository in your work, please cite the following paper. The bibtex is listed below:

@article{shi2019build,
  title={How to Build User Simulators to Train RL-based Dialog Systems},
  author={Shi, Weiyan and Qian, Kun and Wang, Xuewei and Yu, Zhou},
  journal={arXiv preprint arXiv:1909.01388},
  year={2019}
}

Agenda-based simulator

under simulator/

Supervised-learning-based simulator

under sequicity_user/

** for the seq2seq model, because the codebase for the seq2seq module exceeds the file limit, please contact us for it. But it's a simple vanilla seq2seq, you can build your own. The code is under seq2seq/, and we use the implementation from https://github.com/IBM/pytorch-seq2seq for the seq2seq generation model. The vectors used in the training can be downloaded from https://nlp.stanford.edu/projects/glove/.

RL training with agenda-based simulator

python run_mydata_new.py

RL training with supervised-learning-based simulator

python run_mydata_seq_new.py

Interacting with trained policies

policies are under simulator/policy/

data

data.json and delex.json exceed the file limit (100MB), therefore there are two compressed files named data.json.tar.gz and delex.json.tar.bz2 under data/. Please use these two files.