GraphDialog

This is the Tensorflow implementation of the paper: GraphDialog: Integrating Graph Knowledge into End-to-End Task-Oriented Dialogue Systems. EMNLP 2020.

This code has been written using Tensorflow >= 2.0.0. If you find the source codes or the datasets included in this software useful to your work, please consider citing our paper. The bibtex is:

@inproceedings{yang-etal-2020-graphdialog,
    title = "{G}raph{D}ialog: Integrating Graph Knowledge into End-to-End Task-Oriented Dialogue Systems",
    author = "Yang, Shiquan and Zhang, Rui and Erfani, Sarah",
    booktitle = "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
    year = "2020",
    url = "https://www.aclweb.org/anthology/2020.emnlp-main.147",
    pages = "1878--1888"
}

Architecture

Requirements

• Tensorflow 2.0.0a0
• Spacy 2.2.1
• cudatoolkit 10.0.130
• cudnn 7.6.0
• tqdm 4.36.1
• nltk 3.4.5
• numpy 1.17.2
• python 3.7.4

Training

We created myTrain.py to train the models. For SMD dataset, you can run:

python myTrain.py -maxdeps=5 -revgraph=0 -lr=0.001 -hdd=128 -dr=0.2 -bsz=32 -l=1 -ds=kvr

For MultiWOZ 2.1 dataset, you can run:

python myTrain.py -maxdeps=5 -revgraph=0 -lr=0.001 -hdd=128 -dr=0.2 -bsz=128 -l=1 -ds=multiwoz

While training, the model with the best validation results is stored. If you want to reuse a model, please add -path=path_name_model to the call. The model is evaluated by BLEU and Entity F1.

Evaluation

We created myTest.py to restore the checkpoints and test the models. For SMD dataset, you can run:

python myTest.py -path=<path_to_saved_model> -ds=kvr -maxdeps=5 -graphhdd=128 -nheads=1 -rec=1 -graph_layer=1

For MultiWOZ 2.1 dataset, you can run:

python myTest.py -path=<path_to_saved_model> -ds=multiwoz -maxdeps=5 -graphhdd=128 -nheads=1 -graph_layer=1

Reproducibility

We've attached the checkpoints to facilitate the reproduction of the results in the paper. For example, for SMD dataset, you can run:

python myTest.py -path=save/GraphDialog-KVR/KVRHDD128BSZ32DR0.2L3lr0.001ENTF1-0.4768/ckpt-16 -ds=kvr -rec=1 -maxdeps=5 -graphhdd=128 -nheads=1 -graph_layer=1

For MultiWOZ 2.1 dataset, you can run:

python myTest.py -path=save/GraphDialog-MULTIWOZ/multiwozHDD128BSZ128DR0.2L1lr0.001ENTF1-0.1513/ckpt-9 -ds=multiwoz -maxdeps=7 -graphhdd=128 -nheads=1 -graph_layer=1

Others

If you have any questions, please feel free to send us emails (shiquan@student.unimelb.edu.au). We are happy to help you:blush: !