Stance Detection for Fake News Identification

The latest election cycle generated sobering examples of the threat that fake news poses to democracy. Primarily disseminated by hyper-partisan media outlets, fake news proved capable of becoming viral sensations that can dominate social media and influence elections. To address this problem, we begin with stance detection, which is a first step towards identifying fake news. The goal of this project is to identify whether given headline-article pairs: (1) agree, (2) disagree, (3) discuss the same topic, or (4) are not related at all, as described in [1]. Our method feeds the headline-article pairs into a bidirectional LSTM which first analyzes the article and then uses the acquired article representation to analyze the headline. On top of the output of the conditioned bidirectional LSTM, we concatenate global statistical features extracted from the headline-article pairs. We report a 9.7% improvement in the Fake News Challenge evaluation metric and a 22.7% improvement in mean F1 compared to the highest scoring baseline. We also present qualitative results that show how our method outperforms state-of-the art algorithms on this challenge.