Combating Fake News with Stance Detection using Recurrent Neural Networks

Abstract

Comparing the attitude that different news organizations have towards a certain claim or topic is an important part of the procedure used by human fact-checkers today for assessing the veracity of a news story reporting about the issue. In this thesis we focus on automating the challenging task of stance detection in the news domain, specifically determining the relative stance of an article towards a claim stated in an associated headline, making use of the labelled dataset delivered for supervision in the first stage of the Fake News Challenge. While the most successful approaches in this domain have used complex ensemble classifiers employing large sets of hand-engineered features, their performance is just marginally better than a simple bag-of-words model deriving only lexical similarity. Our approach makes use of recurrent neural networks that read headlines and articles word-by-word. The models we implement are comparable to the state-of-the-art systems, however, observing that severe overfitting occurs.