Multi-domain Fake News Detection with Fuzzy Labels

Fake news commonly exists in various domains (e.g., education, health, finance), especially on the Internet, which cost people much time and money to distinguish. Recently, previous researchers focused on fake new detection with the help of a single domain label because fake news has different features in different domains. However, one problem is still solved: A piece of news may have semantics even in one domain source and these meanings have some interactions with other domains. Therefore, detecting fake news with only one domain may lose the contextual semantics of global sources (e.g., more domains). To address this, we propose a novel model, FuzzyNet, which addresses the limitations above by introducing the fuzzy mechanism. Specially, we use BERT and mixture-of-expert networks to extract various features of input news sentences; Then, we use domain-wise attention to make the sentence embedding more domain-aware; Next, we employ attention gate to extract the domain embedding to affect the weight of corresponding expert's result; Moreover, we design a fuzzy mechanism to generate pseudo domains. Finally, the discriminator module uses the total feature representation to discriminate whether the news item is fake news. We conduct our experiment on the Weibo21 dataset and the experimental results show that our model outperforms the baselines. The code is open at https://anonymous.4open.science/r/fakenewsdetection-D2F4.