Evidence-Aware Multimodal Chinese Social Media Rumor Detection

Abstract

The rapid proliferation of the internet has expedited the dissemination of multimedia rumor content on social media. Nevertheless, existing multimodal rumor detection approaches predominantly concentrate on the intrinsic features of the multimodal rumor content, lacking factual evidence support, which constrains their generalizability and rationality. Inspired by the task of Natural Language Inference (NLI), we propose a multimodal rumor-evidence-aware method in this paper. Given a multimodal rumor instance, we predict the extent of evidence support in the given instance for the multimodal rumor content via an evidence-aware model. To prove the effectiveness of this method, we have collected a Chinese Social Media Rumor dataset with Evidence (CSMRE) comprising 12,371 instances. Experimental results on CSMRE datasets show the superior performance of our method.