A Contrastive Learning Method for Automated Fact-Checking (C2E2)

This repository includes the implementation of the C2E2 contrastive learning method for a Korean fact-check dataset.

Data

factcheck-ko

The Korean fact-checking dataset can be obtained from this repository.

• data/wiki_claims.json: human-annotated Dataset for the Factcheck
• data/train_val_test_ids.json: Lists of claim ids for train/validation/test split
• data/wiki/wiki_docs.json: Wikipedia documents corresponding to claims in wiki_claims.json
• dr/dr_results.json

Newly processed data

• pretrain/data/c2e2_data.csv
• pretrain/data/simcse_data.csv

Contrastive pretraining

1. C2E2

   cd pretrain
   python ./train.py --input_df="c2e2_data.csv" --pos_neg="c2e2"
   

2. SimCSE

   cd pretrain
   python ./train.py --input_df="simcse_data.csv" --pos_neg="simcse"
   

• The backbone model is fixed in our implementation as KPFBERT.
• You can obtain the KPFBERT-C2E2 pretrained checkpoint here.

Inference (Sentence Selection)

python sentence_selection/embedding_based_similarity.py --split="test" --gpu_number=0 --checkpoints_dir="./pretrain/checkpoints/" --max_length=512 --model="kosimcse_kpfbert_c2e2" --model_name="kpfbert_c2e2_checkpoint.pt"

Reference

For more details on the task and method, please take a look at the paper published in the Journal of KIISE (in Korean).

@article{송선영2023팩트체킹,
  title={자동화 팩트체킹을 위한 대조학습 방법},
  author={송선영 and 안제준 and 박건우},
  journal={정보과학회논문지},
  year={2023}
}