Soft-Label for Multi-Domain Fake News Detection

Abstract

The spread of fake news across several fields has had serious negative impacts on the public and society. Existing studies have shown that the use of multi-domain labels can improve the accuracy of fake news detection models since news from different domains has different characteristics. However, the previous multi-domain strategy has a problem: if the data lacks domain labels, the domain knowledge learned by the model for each domain cannot be used to determine whether the news is true or fake. Therefore, a novel multi-domain fake news detection model (SLFEND) is proposed to address the above problems by using soft labels. First, using our proposed Leap GRU to skip useless words, the membership function module generates soft labels for each news item. Then, multi-domain features of the messages are extracted using the soft labels to obtain the final overall feature representation. Finally, the overall feature set of the news is fed into the discriminator, and a judgment is made as to whether the news is real or fake. On the Weibo21 and Thu datasets, our proposed SLFEND approach achieves F1 scores of 92.49% and 89.98% respectively, surpassing existing state-of-the-art research. Experiments on the Weibo21 dataset demonstrate the SLFEND model's practicality and efficacy. Experiments using the Thu dataset indicate that SLFEND can transfer knowledge to the news without multi-domain labels and add suitable multi-domain labels to more scientifically determine whether the news is real or fake.