Abstract
With the advent of social media and cell phones, news is now far more reaching and impactful than ever before. This comes with the exponential increase in fake news that blurs the lines of reality and holds the power to sway public opinion. To counter the impact of fake news, several research groups have developed novel algorithms that could fact check news as a human would do. Unfortunately, natural language processing (NLP) is a complicated task because of the underlying hidden meanings in human communication. In this paper, we propose a novel method that builds a latent representation of natural language to capture its underlying hidden meanings accurately and classify fake news. Our approach connects the high-level semantic concepts in the news content with their low-level deep representations so that the complex news text consisting of satire, sarcasm, and purposeful misleading content can be translated into quantifiable latent spaces. This allows us to achieve very high accuracy, surpassing the scores of all winners of the fake news challenge.
