Rethinking Supervised Learning and Reinforcement Learning in Task-Oriented Dialogue Systems

This is the codebase for paper: "Rethinking Supervised Learning and Reinforcement Learning in Task-Oriented Dialogue Systems".

Requirements

python 3.6

pytorch 1.0

CPU

Data

unzip file under data directory.

There are two slightly different state trackers in this repository, tracker_new.py and tracker_old.py. The result in the paper is based on tracker_old.py. Just replace file tracker.py with the one that you want to use.

DiaMultiDense and DiaAdv

The params for DiaAdv can be found in file: diaadv_emnlp_reproduce. The first five lines are for DiaMultiDense while the left five lines are for DiaAdv. Other hyperparameters can be found in the file: utils.py

python -u main.py $line_params

replace $line_params with one line from file diaadv_emnlp_reproduce.

DiaSeq

The params for DiaSeq can be found in file: diaseq_emnlp_reproduce. Other hyperparameters can be found in the file: utils.py

python -u main_seq.py $line_params

replace $line_params with one line from file diaseq_emnlp_reproduce.

If you use the code for dialogue policy learning, feel free to cite our publication Rethinking Supervised Learning and Reinforcement Learning in Task-Oriented Dialogue Systems:

@article{li2020rethink,
  title={Rethinking Supervised Learning and Reinforcement Learning in Task-Oriented Dialogue Systems},
  author={Li, Ziming and Kiseleva, Julia and de Rijke, Maarten},
  journal={Findings of EMNLP 2020},
  year={2020}
}

This code is based on the source code of Ryuichi's GDPL work.