Abstract
In the current Internet era, there exists rapid spread of fake news, which could lead to serious problems. Many artificial intelligence approaches have been deployed to address the problem; however, fake news detection remains a challenge. To detect a fake news, an understanding of certain actors, entities and the relation of between each word in a long text is essential. Many approaches fail to incorporate these attributes in a long text. We purpose a novel BERT approach with joint learning framework that combines relational features classification (RFC) and named entity recognition (NER). Experimenting on two real-world datasets, we observe the effectiveness of our proposed approach in three evaluation metrics: such as accuracy, F1, and area under the curve (AUC) scores. The uniqueness of our joint framework provides a meaningful weight to attributes, which leads to better performance compared to other baselines.