This is the code repository for the paper

     CycleGT: Unsupervised Graph-to-Text and Text-to-Graph Generation via Cycle Training

Given a text corpus, and a graph dataset, and no parallel (text, graph) pairs, our model CycleGT aims to jointly learn text-to-graph and graph-to-text in a cycle framework.

Dependencies

• pytorch 1.4.0 cu10
• pip install -r requirements.txt

How to Run

python -u main.py

Feel free to change config.yaml to set other configurations.

Results on WebNLG

Method	BLEU	Micro F1	Macro F1
Back Translation w/ warmup	43.08+-0.39	62.3+-0.2	52.1+-0.3
Back Translation w/o warmup	44.97+-0.66	61.6+-0.2	51.4+-0.2
Supervised	44.88+-0.66	61.0+-0.3	51.0+-0.3

Training on new data

Preparing the data in json format and replace the file path in the config.yaml
And there are some constraints:

• entity mentions, we use the symbol like <ENT_0> to represent an entity in the text, and its surface form should list in the "entities" field.
• all the nodes in the graph should be mentioned in "entities"

For example,

{"relations": [[["Abilene", "Regional", "Airport"], "cityServed", ["Abilene", ",", "Texas"]]], 
"text": "<ENT_0> is served by the <ENT_1> .", 
"entities": [["Abilene", ",", "Texas"], ["Abilene", "Regional", "Airport"]]}

BTW, we don't tokenize the text, so make sure that the entity and text have been tokenized.

Citing CycleGT

If you use CycleGT, please cite CycleGT: Unsupervised Graph-to-Text and Text-to-Graph Generation via Cycle Training.

@article{guo2020cyclegt,
  author    = {Qipeng Guo and Zhijing Jin and Xipeng Qiu and Weinan Zhang and David Wipf and Zheng Zhang},
  title     = {CycleGT: Unsupervised Graph-to-Text and Text-to-Graph Generation via Cycle Training},
  journal   = {CoRR},
  volume    = {abs/2006.04702},
  year      = {2020},
  url       = {https://arxiv.org/abs/2006.04702},
  archivePrefix = {arXiv},
  eprint    = {2006.04702}
}