This is the code for the AAAI-24 Paper:Heterogenous Graph Reasoning for Fact Checking over Texts and Tables.
-
Download Roberta pretrained model and put related files into directory
roberta-large-snli_mnli_fever_anli_R1_R2_R3-nli/
-
Prepare dataset and Wikidump. Link. As a result, you should rename the wikidump to
feverous_wikiv1.db
as well as the dataset splits:train.jsonl
,dev.jsonl
andtest.jsonl
. -
Follow DCUF's work and generate retrieved evidence files for the three splits. See [GitHub Link](https://github.com/lanlanabcd/dual_ channel_feverous) for more instruction on generating retrieved evidence file. After the retrieval process, the results should be three files named as
dev.combined.not_precomputed.p5.s5.t3.cells.jsonl
,train.combined.not_precomputed.p5.s5.t3.cells.jsonl
andtest.combined.not_precomputed.p5.s5.t3.cells.jsonl
. -
Run
python train.py
to run the training process. It includes a preprocess procedure for the dataset splits and evaluation on dev split. -
To evaluate the performance on test split, online judge is needed. You can also check our submission result on the online judge system. To generate file to submit run the following code:
python testresult.py; python generatesubmitfile.py --input_file tosubmit.csv
- python3
- jsonlines
- pytorch
- torch-geometric
Please cite our paper if you use the code:
@inproceedings{gong2024heterogeneous,
title={Heterogeneous Graph Reasoning for Fact Checking over Texts and Tables},
author={Gong, Haisong and Xu, Weizhi and Wu, Shu and Liu, Qiang and Wang, Liang},
booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
volume={38},
number={1},
pages={100--108},
year={2024}
}