Fake News Detection via Multi-Modal Topic Memory Network

With the development of the Mobile Internet, more and more people create and release multi-modal posts on social media platforms. Fake news detection has become an increasingly challenging task. Although many current works focus on constructing models extracting abstract features from the content of each post, they neglect the intrinsic semantic architecture such as latent topics, etc. These models only learn patterns in content coupled with certain specific latent topics on the training set to distinguish real and fake posts, which will suffer generalization and discriminating ability decline, especially when posts are associated with rare or new topics. Moreover, most existing works using deep schemes to extract and integrate textual and visual representation in post have not effectively modeled and sufficiently utilized the complementary and noisy multi-modal information containing semantic concepts and entities to complement and enhance each modal. In this paper, to deal with the above problems, we propose a novel end-to-end Multi-modal Topic Memory Network (MTMN), which obtains and combines post representations shared across latent topics together with global features of latent topics while modeling intra-modality and inter-modality information in a unified framework. (1) To tackle real scenarios where newly arriving posts with different topic distribution from the training data, our method incorporates a topic memory module to explicitly characterize final representation as post feature shared across topics and global features of latent topics. These two kinds of features are jointly learned and then combined to generate robust representation. (2) To effectively integrate multi-modality information in posts, we propose a novel blended attention module for multi-modal fusion, which can simultaneously exploit the intra-modality relation within each modal and the inter-modality relation between text words and image regions to complement and enhance each other for high-quality representation. Extensive experiments on two public real-world datasets demonstrate the superior performance of MTMN compared with other state-of-the-art algorithms.