Explore the Style for Fake News Detection

The spread of information on the internet is caused with little to no filters or supervision, which enables the widespread dissemination of misinformation. Due to the frequency of false content, misinformation has been treated synonymously as fake news. To mitigate the fake news problem, we have explored automatic methods to sort through the vast amount of information for its correctness. The problem occurs because fake news is fabricated deliberately to include false information, which is hard to verify. Many general-purpose classifiers rely on content to determine its reliability which unfortunately often could not be verified due to a lack of information about an incident that happens in real-time. The lack of realtime information inhibits the model's ability to produce an educated prediction. In our work, we propose a method that focuses on writing style to generalize the classifiers to maintain robust performance for previously unseen topics and unseen sources. The experiment shows that our model could improve over 5% over the BERT [1] model and over 3% over the best results on documents with unknown sources; our model establishes the best results in the condition where training data is insufficient by improving 5-8% over baseline results.