This is code for the paper "SynTOD: An Augmented Synthetic Response Approach for Robust End-to-end Task-Oriented Dialogue System"
To download data.zip properly, git lfs(Large File Storage) extension must be installed.
git lfs install
git lfs pull
git checkout -f HEAD
Our python version is 3.6.9.
The package can be installed by running the following command.
pip install -r requirements.txt
python -m spacy download en_core_web_sm
For the experiments, we use MultiWOZ2.1 and MultiWOZ2.2. You can download at MultiWOZ
We use the preprocessing scripts implemented by Zhang et al., 2020. Please refer to here for the details.
- T5-small (https://huggingface.co/t5-small)
- PPTOD-small (https://github.com/awslabs/pptod)
Our implementation supports a single GPU with 2 phases training. Please use smaller batch sizes if out-of-memory error raises.
First, we train model with original data by following script (please refer to config.py to see the detailed configs):
CUDA_VISIBLE_DEVICES="0" python main.py -run_type train -model_dir ckpt2.1/t5-small -backbone t5-small -version 2.1
The next step, the best checkpoint is ultilized to generate synthetic data:
CUDA_VISIBLE_DEVICES="0" python main.py -run_type predict -ckpt ckpt2.1/t5-small/ckpt-epoch4 -batch_size 64 -pred_data_type test -output out.json -version 2.1
After that, combine systhetic data with original data:
python data/merge.py
Finally, set ADD_SYNTHETIC_DATA = True in quick_config.py ,
train model with synthetic data and original data:
CUDA_VISIBLE_DEVICES="0" python main.py -run_type train -model_dir ckpt2.1/t5-small-synthetic -backbone t5-small -version 2.1
The checkpoints will be saved at the end of each epoch (the default training epoch is set to 10).
python main.py -run_type predict -ckpt $CHECKPOINT -output $MODEL_OUTPUT -batch_size $BATCH_SIZE
All checkpoints are saved in $MODEL_DIR with names such as 'ckpt-epoch10-step7053'.
The result file ($MODEL_OUTPUT) will be saved in the checkpoint directory.
To reduce inference time, it is recommended to set large $BATCH_SIZE. In our experiemnts, it is set to 16 for inference.
You can download our trained model here.
Our code hugely based on (https://github.com/bepoetree/MTTOD) for "Improving End-to-End Task-Oriented Dialogue System with A Simple Auxiliary Task" and (https://github.com/Tomiinek/MultiWOZ_Evaluation) for evaluation.