Multimodal Fusion with BERT and Attention Mechanism for Fake News Detection

Fake news detection is an important task for in- creasing the reliability of the information on the internet since fake news is spreading fast on social media and has a negative effect on our society. In this paper, we present a novel method for detecting fake news by fusing multi-modal features derived from textual and visual data. Specifically, we proposed a scaled dot- product attention mechanism to capture the relationship between text features extracted by a pre-trained BERT model and visual features extracted by a pre-trained VGG-19 model. Experimental results showed that our method improved against the current state-of-the-art method on a public Twitter dataset by 3.1% accuracy.