Zero-Shot Automated Detection of Fake News: An Innovative Approach (ZS-FND)

Abstract

The propagation of fake news has significant societal and economic impacts, particularly in the context of the exponential growth of digital and social media platforms. Traditional Machine Learning (ML) and Deep Learning (DL) models, i.e., Recurrent Neural Networks (RNN), Convolutional Neural Networks (CNN), and Long Short-Term Memory (LSTM) networks, have been employed to address Fake News Detection (FND) with varying degrees of success. However, these models often struggle with data dependency, overfitting, scalability, flexibility, and high computational costs. A Zero-Shot Learning (ZSL) approach is proposed for Fake News Detection (ZS-FND) to address these challenges. ZSL models can predict outcomes with limited training data and do not rely on domain-specific labeled data, making them well-suited for handling fake news’s diverse and evolving nature. The proposed ZS-FND model leverages the pre-trained Bidirectional Encoder Representations from Transformers (BERT) for the semantic representation of textual data and generates word vectors. ZSL considers such vectors as input for FND. The experimental results demonstrate that ZS-FND outperforms conventional ML and DL models, achieving 98.39% accuracy, 97.33% precision, 95.67% recall, 96.49% F1-score, and 0.0160 Mean Absolute Error (MAE). ZS-FND improves accuracy, precision, recall, and F1-score by 10.76%, 4.05%, 5.96%, and 5.01%, respectively. These findings highlight the potential of ZSL models in providing a more robust and efficient solution for FND.

Keywords

• Fake news detection
• BERT
• zero-shot classifier
• classification
• deep learning