Multi-Level Multi-Modal Cross-Attention Network for Fake News Detection

With the development of Mobile Internet, more and users publish multi-modal posts on social media platforms. Fake news detection has become an increasingly challenging task. Although there are many works using deep schemes to extract combine textual visual representation in post, most existing methods do not sufficiently utilize complementary information containing semantic concepts entities complement enhance each modality. Moreover, these model incorporate rich multi-level semantics text improve fake tasks. In this paper, we propose a novel end-to-end Multi-level Multi-modal Cross-attention Network (MMCN) which exploits jointly integrates relationships duplicate different modalities (textual modality) multimedia unified framework. Firstly, Pre-trained BERT ResNet models employed generate high-quality representations for words image regions, respectively. A cross-attention network is then designed fuse feature embeddings regions by simultaneously considering data modalities. Specially, due layers transformer architecture have representations, employ encoding capture presentations posts. Extensive experiments two public datasets (WEIBO PHEME) demonstrate that compared with state-of-the-art models, proposed MMCN advantageous performance.