Detecting Multimodal Fake News with Gated Variational AutoEncoder

This paper focuses on the challenge of automatically detecting multimodal fake news on social media. Although multimodal fake news classifiers exist, we show that prior works fail to reflect certain real-world practicalities. For example, news captions often contain highly irrelevant information that introduces noise to the overall message contained within the post. Existing classifiers do not properly address this, resulting in misclassifications. To address this limitation and suppress noise, we propose GatedVAE (Gated Variational AutoEncoder), which enables VAE with the gating mechanism. Experimental results demonstrate the efficacy of our approach: GatedVAE is able to suppress noise and learn an effective multimodal representation. It outperforms state-of-the-art models by 3.7% and 2.4% (F1) on Twitter and Weibo datasets, respectively. Our ablation study highlights the importance of the gating mechanism and the methods we adopt to alleviate overfitting. We further show that, in addition to dynamically controlling the pass of noisy input, the gate also helps to uncover modality importance in multimodal fake news detection.