“Think Before You Speak”: Improving Multi-Action Dialog Policy by Planning Single-Action Dialogs IJCAI-22 long oral presentation
This is the codebase for the proposed multi-action dialog policy model PEDP and all the SL-based baseline models (gCAS, DiaMultiClass, DiaMultiDense, and DiaSeq).
For other models, we refer readers to the official implementations. (GDPL and DiaAdv)
Please refer to
environment.yml
and prepare the environment with Anaconda.
We report results on:
-
MultiWOZ. Please download the data from here and unzip under
./data
directory. -
SGD. Please download the data from here and unzip under
./sgd_data
directory.
To reproduce the results on MultiWOZ, execute:
python -u main.py --pedp --residual
To reproduce the results on SGD, execute:
python -u main.py --pedp --residual --sgd
To train other models, execute:
python -u main.py --[model_name]
More hyper-parameters are assigned in args.py
and can be modified using --para=value
.
Dataset schema is defined in config_multiwoz.py
and ./sgd_data/config.py
for MultiWOZ and SGD, respectively.
You can use tensorboard to check the results. Execute:
tensorboard --bind_all --logdir=./log/tb/[file name]
and open the corresponding website.