Mean Teacher BERT discovers Fake News
Mean Teacher BERT discovers Fake News
Abstract
Dissemination of fake news has been ubiquitous with theproliferation of social media platforms in this digital age. If the intentis harmful, fake news could impact negatively on our society. There-fore, it is vital to automatically detect the news potentially contain fakenews before reaching to large audience. Collecting high quality of la-beled corpus for the fake news detection models is a labor-intensive anda challenging step. To overcome this challenge, in this study, we evaluatesemi-supervised training schemas for fake news detection of news articleson publicly available datasets and compare the experimental results withthe state-of-art fake news detection models. The schemas that we eval-uate are Mean-Teacher, TODO ADD YOUR SIGNIFICANT RESULT.Our source code is publicly available1
Datasets
For our paper, we use the following datasets:
- FakeNewsNet contains news articles in politics and entertainment.
- FakeHealth contains news articles in health.
- NELA, NELA-2017 and NELA-2018 are used for unlabeled samples. From the [NELA-2019]() dataset, we use the aggregated labels.
:warning: Please cite the papers of these studies if you use them.
Run the bash code in order to get experiment folds.
You are supposed to see the processed files in Data folder
The Data\Input2 is the dataset on which overall comparision have been done and it is default dataset for model.The reports also generated from dataset2
Starting
Python compiler is 3.7.9.
Install libraries in requirements.txt:
pip3 install -r requirements.txt| Parameter | Description | Default value |
|---|---|---|
| --lr | learning rate | 0.0001 |
| --epochs | epochs | 1 |
| --batch_size | Batch size | 1 |
| --maxlen | Lenght of article for tokens | 512 |
| --model | Meanteacher(MT) or Bert or PI | MT |
| --data | fakehealth or politifact or gossipcop | fakehealth |
| --pretrained_model | 'bert-base-uncased', 'bert-base-cased' | bert-base-uncased |
| --ratio | Used in overall cost for mean teacher | 0.5 |
| --alpha | EMA alpha value for mean teacher | 0.99 |
| --noise_ratio | unlabel data noise ratio for Mean teacher | 0.2 |
| --dropout | drop out value | 0.2 |
| --data_folder | folder location of data | |
| --model_output_folder | model saving location | |
| --do_train | only for BERT model | 'store_true' |
| --seed |
Command to generate the fake news:
python3 -m fake_news_generator \
--data_folder Data/ExperimentFolds/3 \
--data fakehealth \
--processed_output_folder Data/ExperimentFolds/3Command for preprocessing the data:
--data_folder Data/ExperimentFolds/3 \
--data fakehealth \
--processed_output_folder Data/Processed/3Example command to run only BERT model:
python -m BERT.main \
--model bert \
--data_folder 'Data/ExperimentFolds/3' \
--data politifact \
--model_output_folder trained_models \
--pretrained_model bert-base-cased \
--max_len 512 \
--dropout 0.1 \
--epochs 1 \
--batch_size 1 \
--lr 2e-5 \
--seed 42 \
--do_trainCommand to run Mean teacher with BERT:
python3 -m main \
--model MT \
--data_folder Data/ExperimentFolds/3 \
--data fakehealth \
--model_output_folder trained_models \
--epochs 3 \
--lr 0.0001 \
--batch_size 1 \
--dropout 0.1 \
--max_len 512 \
--alpha 0.99 \
--ratio 0.5 \
--noise_ratio 0.9Command to run Pi model with BERT:
python3 -m main \
--model PI \
--data_folder Data/ExperimentFolds/3 \
--data fakehealth \
--model_output_folder trained_models \
--epochs 3 \
--dropout 0.1 \
--lr 0.0001 \
--batch_size 1 \
--max_len 512 \Contributors
Ipek Baris
Bhupender kumar Saini