Hybrid Approach to Fake News Detection: Leveraging BERT-Based Model with Word Embedding Features for Sentiment Classification in Social Media Content

Abstract

The proliferation of fake news in social media necessitates robust and efficient methods for its detection. This research presents a novel approach to sentiment classification for fake news detection, focusing on non-sarcastic social media content. Leveraging the powerful contextual representation capabilities of BERT and enhancing its performance through the incorporation of word embedding features, our proposed model aims to discern the subtle nuances in sentiment that are indicative of deceptive information. The designed model is meticulously crafted to capture the intricate relationships between words and their contextual meanings within the given text. To assess its efficacy, a comparative analysis is conducted with two other models: a Long Short-Term Memory (LSTM) model, a widely used sequential model for natural language processing, and a BERT model without the integration of word embedding features. The experiments involve training and evaluating the models on a comprehensive dataset of non-sarcastic social media content. Performance metrics such as accuracy, precision, recall, and F1 score are employed the research findings indicate that the proposed BERT-based model with word embedding features outperforms both the LSTM model and the BERT model without word embedding features in terms of fake news detection accuracy.