HDNO

This is the codebase for ICLR 2021 paper: Modelling Hierarchical Structure between Dialogue Policy and Natural Language Generator with Option Framework for Task-oriented Dialogue System.

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

@article{wang2020modelling,
  title={Modelling hierarchical structure between dialogue policy and natural language generator with option framework for task-oriented dialogue system},
  author={Wang, Jianhong and Zhang, Yuan and Kim, Tae-Kyun and Gu, Yunjie},
  journal={arXiv preprint arXiv:2006.06814},
  year={2020}
}

Repository Structure

• configs: hyperparameters for different experiments
• data: dialogue dataset (MultiWowz 2.0 & MultiWoz 2.1)
• latent_dialog: source code
• human_evaluator.py: to evalute groud truth's performance
• supervised.py: main entry for Supervised Learning
• reinforce.py: main entry for Reinforcement Learning

Requirements

1. install conda environment

conda create -n hdno python=3.6
conda activate hdno

2. install requirements

pip install -r requirements.txt     

Prepare the data

Before any operations below, please prepare your data following the script:

unzip data/multiwoz_2.0.zip -d data
unzip data/multiwoz_2.1.zip -d data

Reproduce the results

We give a script that can evaluate HDNO and show the results shown on the paper, based on the pretrained models we provide on MultiWoz 2.0 and MultiWoz 2.1:

sh reproduce.sh

Train your own models

For the convenience for freely training models, we give simple bash scripts to do it.

1. pretraining

sh train.sh sl woz2.0 # For MultiWoz 2.0
sh train.sh sl woz2.1 # For MultiWoz 2.1

2. hierarchical reinforcement learning (HRL)

sh train.sh rl woz2.0 # For MultiWoz 2.0
sh train.sh rl woz2.1 # For MultiWoz 2.1

3. evaluating trained model

sh test.sh sl woz2.0 5 # For MultiWoz 2.0 pretrained model
sh test.sh sl woz2.1 5 # For MultiWoz 2.1 pretrained model
sh test.sh rl woz2.0 2 # For MultiWoz 2.0 HRL model 
sh test.sh rl woz2.1 5 # For MultiWoz 2.1 HRL model

Main results

• The table shows the main test results of HDNO on MultiWoz 2.0 and MultiWoz 2.1 evaluated with the automatic evaluation metrics.
  

• The diagram demonstrates latent dialogue acts of HDNO clustered in 8 categories.