Some of the material in is restricted to members of the community. By logging in, you may be able to gain additional access to certain collections or items. If you have questions about access or logging in, please use the form on the Contact Page.
While social media and the internet have become popular platforms in information dissemination amid the proliferation of using social media networks, they simultaneously lead to quick spreads of fake news and misinformation. Especially, the openness of the internet has had severe repercussions for fake news even though its positive function cannot be ignored. Further, it is notable that spreading of misinformation may produce irrevocable and appalling results towards many arbitrary users in a very short time. Thus, minimizing the damages of fake news by accurately detecting them becomes the essential challenges facing this generation. Some state-of-the-art models based on linear models such as count-based methods were proposed after many people recognized the importance of fake news detection, yet these models showed poor accuracy to identify fake news without error. For the reasons set forth, I newly propose a neural network-based model consisting of three layers to detect fake news; (1) It is obvious that there is a structural hierarchy in documents, provided that words are composed of letters, sentences are composed of words, and paragraphs are composed of sentences; (2) It is more plausible to adopt an attention network to increase the accuracy of detecting fake news or misinformation, taken from the fact that the same word may have different importance on individual sentences. Therefore, I hereby propose a model named HAND (Hierarchical Attention Networks for Fake News Detection) which consists of three layers of word-level, sentence-level, and paragraph-level. The experiment examined the performance of HAN and HAND models with the fake news labeled from Kaggle, which showed both models' outperformance over the classic methods and state-of-the-art approaches. By using 18,197 articles, the experiment yielded great results of 95.92%, with more detailed descriptions presented in the Experiment and Results section.