ICTMCG / Pref-FEND

Official repository to release the code and datasets in the paper, "Integrating Pattern- and Fact-based Fake News Detection via Model Preference Learning", CIKM 2021.

Pref-FEND

This is the official repository of the paper:

Integrating Pattern- and Fact-based Fake News Detection via Model Preference Learning.

Qiang Sheng*, Xueyao Zhang*, Juan Cao, and Lei Zhong.

Proceedings of the 30th ACM International Conference on Information and Knowledge Management (CIKM 2021)

PDF / Poster / Code / Chinese Blog

Datasets

The experimental datasets where can be seen in dataset folder, including the Weibo Dataset, and the Twitter Dataset. Note that you can download the datasets only after an "Application to Use the Datasets for Pattern- and Fact-based Joint Fake News Detection" has been submitted.

Code

Key Requirements

python==3.6.10
torch==1.6.0
torchvision==0.7.0
torch-geometric==1.7.0
torch-sparse==0.6.9
transformers==3.2.0

Preparation

Step1: Stylistic Tokens & Entities Recognition

Notice that Step1 is not necessary, because we have supplied the recognized results in dataset json files in dataset folder.

If you would like to know about the details of the recoginition procedure, you can refer to the preprocess/tokens_recognition folder.

Step2: Tokenize

cd preprocess/tokenize

As the run.sh shows, you need to run:

python get_post_tokens.py --dataset [dataset] --pretrained_model [bert_pretrained_model]

Step3: Heterogeneous Graph Initialization

cd preprocess/graph_init

As the run.sh shows, you need to run:

python init_graph.py --dataset [dataset] --max_nodes [max_tokens_num]

Step4: Preparation of the Fact-based Models

Notice that Step3 is not necessary if you wouldn't use fact-based models as a componet for Pref-FEND.

Tokenize

cd preprocess/tokenize

As the run.sh shows, you need to run:

python get_articles_tokens.py --dataset [dataset] --pretrained_model [bert_pretrained_model]

Retrieve by BM25

cd preprocess/bm25

As the run.sh shows, you need to run:

python retrieve.py --dataset [dataset]

Step5: Preparation for some special fake news detectors

Notice that Step5 is not necessary if you wouldn't use EANN-Text or BERT-Emo as a componet for Pref-FEND.

EANN-Text

cd preprocess/EANN_Text

As the run.sh shows, you need to run:

python events_clustering.py --dataset [dataset] --events_num [clusters_num]

BERT-Emo

cd preprocess/BERT_Emo/code/preprocess

As the run.sh shows, you need to run:

python input_of_emotions.py --dataset [dataset]

Training and Inferring

cd model
mkdir ckpts

Here we list all of our configurations and running script in run_weibo.sh and run_twitter.sh. For example, if you would like to run BERT-Emo (Pattern-based) + MAC (Fact-based) with Pref-FEND on Weibo, you can run:

# BERT_Emo + MAC (Pref-FNED)
CUDA_VISIBLE_DEVICES=0 python main.py --dataset 'Weibo' \
--use_preference_map True --use_pattern_based_model True --use_fact_based_model True \
--pattern_based_model 'BERT_Emo' --fact_based_model 'MAC' \
--lr 5e-6 --batch_size 4 --epochs 50 \
--save 'ckpts/BERT_Emo+DeClarE_with_Pref-FEND'

then the results will be saved in ckpts/BERT_Emo+DeClarE_with_Pref-FEND.

Citation

@inproceedings{Pref-FEND,
  author    = {Qiang Sheng and
               Xueyao Zhang and
               Juan Cao and
               Lei Zhong},
  title     = {Integrating Pattern- and Fact-based Fake News Detection via Model
               Preference Learning},
  booktitle = {{CIKM} '21: The 30th {ACM} International Conference on Information
               and Knowledge Management, Virtual Event, Queensland, Australia, November
               1 - 5, 2021},
  pages     = {1640--1650},
  year      = {2021},
  url       = {https://doi.org/10.1145/3459637.3482440},
  doi       = {10.1145/3459637.3482440}
}