A Unified Training Process for Fake News Detection Based on Finetuned Bidirectional Encoder Representation from...

Area of Science:

• Natural Language Processing
• Machine Learning
• Artificial Intelligence

Background:

• The rapid growth of social media necessitates efficient fake news detection.
• Existing models often perform poorly due to reliance on single datasets.
• Training models on combined datasets is crucial but time-consuming and complex.

Purpose of the Study:

• To introduce a unified training strategy for fake news detection using a pretrained transformer model.
• To address the challenges of extensive training time and parameter optimization in combined datasets.
• To develop a robust fake news classifier with improved performance and efficiency.

Main Methods:

• A unified training strategy was developed using a pretrained transformer model.
• The model was trained on a combined dataset including ISOT and Kaggle datasets.
• Input samples were preprocessed by removing words shorter than three letters to reduce training time.
• Performance was analyzed by varying the number of encoder blocks.

Main Results:

• The unified training strategy achieved 97% accuracy and an F1 score of 0.97.
• The proposed model outperformed existing methods like Random Forests, CNNs, and LSTMs.
• Training time was reduced by 1.5 to 1.8 times through input data preprocessing.
• Reducing encoder blocks led to decreased model performance.

Conclusions:

• The unified training strategy offers an efficient and effective solution for fake news detection.
• Pretrained transformer models combined with a unified strategy provide a robust approach for handling diverse datasets.
• Further research can explore optimizing model compactness without sacrificing performance.