Abstract: Misinformation proliferates on the Internet,undermining the normal functioning of various industries.Detecting falsehoods accurately has therefore become an urgent challenge.Existing research on this task focuses primarily on three aspects:account traits,textual content,and multimodality.However,most methods overlook the key attribute of misinformation diffusion the novelty of content.They analyze the veracity of target claims in isolation,failing to capture public opinion dynamics.To address this issue,this paper proposes a cross-document misinformation detection framework called contrastive graph learning(CAL).CAL focuses on content novelty and comprises two key components:a contrastive learning module and a heterogeneous graph module.The former expands the representational difference between factual and false claims,and the latter encompasses five entity types:words,events,event sets,sentences,and documents.It injects semantic features of the public discourse into entity embeddings.We evaluate CAL on the IED,TL17,and Crisis datasets at both document and event levels.CGL achieves state-of-the-art performance,which verifies the efficacy of its design.It provides a robust solution for combating misinformation by mode-ling novelty and environmental context.

Key words: Cross-document misinformation detection, Contrastive learning, Heterogeneous graph, Event-level detection