Code for the paper Diverse and Faithful Knowledge-Grounded Dialogue Generation via Sequential Posterior Inference. [PDF]

The paper will be presented at ICML 2023. This code has been written using PyTorch. If you use source codes in this repository in your work, please cite the following papers:

@inproceedings{pmlr-v202-xu23j,
	author = {Xu, Yan and Kong, Deqian and Xu, Dehong and Ji, Ziwei and Pang, Bo and Fung, Pascale and Wu, Ying Nian},
	booktitle = {Proceedings of the 40th International Conference on Machine Learning},
	editor = {Krause, Andreas and Brunskill, Emma and Cho, Kyunghyun and Engelhardt, Barbara and Sabato, Sivan and Scarlett, Jonathan},
	month = {23--29 Jul},
	pages = {38518--38534},
	pdf = {https://proceedings.mlr.press/v202/xu23j/xu23j.pdf},
	publisher = {PMLR},
	series = {Proceedings of Machine Learning Research},
	title = {Diverse and Faithful Knowledge-Grounded Dialogue Generation via Sequential Posterior Inference},
	url = {https://proceedings.mlr.press/v202/xu23j.html},
	volume = {202},
	year = {2023},
	bdsk-url-1 = {https://proceedings.mlr.press/v202/xu23j.html}
}

Install the environment with the following command lines:

conda env create -f dial.yml python=3.10
conda activate dial-env

To run both training and prediction of SPI, you need to prepare the data for experiments.

1. Download datasets from the official links.

2. Put the downloaded datasets under data folder, named as wizard_of_wikipedia and holle, respectively.

3. Preprocess data by unwraping dialogues and converting the dialogues into data samples. Run the following command line under the main folder:

python src/data_utils/wow_proc.py --preproc_dir data/processed_wow
python src/data_utils/holle_proc.py --preproc_dir data/processed_holle

4. The processed data will be stored under the preproc_dir. Do not modify the above path, or you will have to modify the hard-coded path in src/data_utils/wizard_of_wikipedia.py and src/data_utils/holle.py.

Alternatively, you can also download the pre-processed data directly from here.

For reproducibility, you can access our pre-trained weights from here. You can also train the models yourself:

sh run_spi.sh

In run_spi.sh, the command line for training four different models is provided. Please use them based on your needs.

Given one checkpoint, we can evaluate our model with the following command line. The script will 1) compute the perplexity of generating the gold responses and 2）generate responses given the data samples in the test set.

sh predict_spi.sh

Our code without cleansing and pre-training model weights are available here tentatively.