There are source codes for Kernel Graph Attention Network for Fact Verification.
For more information about the FEVER 1.0 shared task can be found on this website.
- Python 3.X
- fever_score
- Can be found at Google Drive.
- BERT based ranker.
- Go to the
retrieval_modelfolder for more information.
- Pre-train BERT with claim-evidence pairs.
- Go to the
pretrainfolder for more information.
- Our KGAT model.
- Go to the
kgatfolder for more information.
The results are all on Codalab leaderboard. (The Rank@1 and Rank@2 use XLNet and BERT(large)).
| Rank | User | Pre-train Model | Label Accuracy | FEVER Score |
|---|---|---|---|---|
| 1 | DREAM | XLNet | 0.7685 | 0.7060 |
| 2 | abcd_zh (Ours) | RoBERTa (Base) | 0.7407 | 0.7038 |
| 3 | a.soleimani.b | BERT (Large) | 0.7186 | 0.6966 |
| 9 | GEAR_single | BERT (Base) | 0.7160 | 0.6710 |
KGAT performance with different pre-trained language model.
| Pre-train Model | Label Accuracy | FEVER Score |
|---|---|---|
| RoBERTa (Base) | 0.7407 | 0.7038 |
| BERT (Large) | 0.7361 | 0.7024 |
| BERT (Base) | 0.7281 | 0.6940 |
@article{liu2019kernel,
title={Kernel Graph Attention Network for Fact Verification},
author={Liu, Zhenghao and Xiong, Chenyan and Sun, Maosong},
journal={arXiv preprint arXiv:1910.09796},
year={2019}
}
If you have questions, suggestions and bug reports, please email: