DAML

Source code for the ACL 2019 paper entitled "Domain Adaptive Dialog Generation via Meta Learning" by Kun Qian and Zhou Yu https://arxiv.org/abs/1906.03520

@article{qian2019domain,
  title={Domain Adaptive Dialog Generation via Meta Learning},
  author={Qian, Kun and Yu, Zhou},
  journal={arXiv preprint arXiv:1906.03520},
  year={2019}
}

Simulated Data Generation

Please download the code here: https://github.com/qbetterk/SimDial

git clone https://github.com/qbetterk/SimDial.git
cd SimDial
bash run.sh

This generate train/dev/test data (1500 dialogs in each domain) and adaptation data (9 dialogs in each domain). The size of data can be modified by argument "size" of script run.sh

Training with default parameters

python model.py

(optional: configuring hyperparameters with cmdline)

python model.py -mode train_maml -model tsdf-camrest -cfg lr=0.003 batch_size=32

Testing

python model.py -mode test_maml -model tsdf-camrest

or test on new domains (e.g. movie domain):

bash run_movie.sh

Before running

1. Install required python packages. We used pytorch 0.3.0 and python 3.6 under Linux operating system.

pip install -r requirements.txt

2. Make directories under PROJECT_ROOT.

mkdir vocab
mkdir log
mkdir results
mkdir models
mkdir sheets

3. Download pretrained Glove word vectors and place them in PROJECT_ROOT/data/glove.