Using Back-Propagation based Highlights as an Explanation in Human Fake News Detection

Abstract

Few clicks and one can publish a piece of text online, and such ease has allowed Fake news to spread quickly. It can wreak havoc, even to the extent of wiping billions of dollars from the stock market or instigating the lynching of innocent people. Past research shows that humans are marginally slightly better than chance at detecting Fake news. Therefore, researchers are actively exploring the prospect of using using black-box models to detect Fake news, but this could result in unnecessary censorship as we cannot directly interrogate a black-box model to ask why it made a particular decision. There are existing techniques that allow inspecting the output of the model, but little or no research has been done on their applicability. Therefore, this thesis examines the usability of explanation developed by O’brien, Latessa, Evangelopoulos, and Boix (2018) with a between-subject experiment where average accuracy of fake news detection was compared between groups that received such explanations against group that did not receive any explanations. The result shows that providing any explanation improves the mean accuracy but there is no statistically significant difference in accuracy between groups that received explanations and groups that did not receive any explanations.