Enhancing Text and Image Representations for Multimodal Fake News Detection via Knowledge Distillation

The widespread use of social media and the low-cost nature of online media have led to a rapid increase in news sources. Consequently, unchecked and unverified information floods the internet like a tsunami. The emergence of trending events often brings a surge of fake news, including but not limited to satire, misleading content, imposter content, false connections, and manipulated content. To address the above phenomena, this thesis further explores methods for fake news detection. We propose a model based on the BERT-ViT multimodal architecture: 1. Incorporating knowledge distillation techniques enhances text and image encoders' sentiment and intent analysis ability to naturally and effectively extract embeddings. 2. Introduce a transformer-based fusion block to help modality fusion. Experiments conducted on two social media fake news classification datasets demonstrate that our approach achieves significantly better results compared to baseline models.