YeDeming / KernelGAT

The source codes for Fine-grained Fact Verification with Kernel Graph Attention Network.

Kernel Graph Attention Network (KGAT)

There are source codes for Fine-grained Fact Verification with Kernel Graph Attention Network.

For more information about the FEVER 1.0 shared task can be found on this website.

Requirement

• Python 3.X
• fever_score
• Pytorch
• pytorch_pretrained_bert
• transformers

Data and Checkpoint

• All data and BERT based chechpoints can be found at Ali Drive.
• RoBERTa based models and chechpoints can be found at Ali Drive.

Retrieval Model

• BERT based ranker.
• Go to the retrieval_model folder for more information.

Pretrain Model

• Pre-train BERT with claim-evidence pairs.
• Go to the pretrain folder for more information.

KGAT Model

• Our KGAT model.
• Go to the kgat folder for more information.

Results

The results are all on Codalab leaderboard.

User	Pre-train Model	Label Accuracy	FEVER Score
GEAR_single	BERT (Base)	0.7160	0.6710
a.soleimani.b	BERT (Large)	0.7186	0.6966
KGAT	RoBERTa (Large)	0.7407	0.7038

KGAT performance with different pre-trained language model.

Pre-train Model	Label Accuracy	FEVER Score
BERT (Base)	0.7281	0.6940
BERT (Large)	0.7361	0.7024
RoBERTa (Large)	0.7407	0.7038
CorefBERT (RoBERT Large)	0.7596	0.7230

Citation

@inproceedings{liu2020kernel,
  title={Fine-grained Fact Verification with Kernel Graph Attention Network},
  author={Liu, Zhenghao and Xiong, Chenyan and Sun, Maosong and Liu, Zhiyuan},
  booktitle={Proceedings of ACL},
  year={2020}
}

Contact

If you have questions, suggestions and bug reports, please email:

[e-mail]