Neural Stance Detectors for Fake News Challenge

Fake news pose serious threat to our society nowadays, particularly due to its wide spread through social networks. While human fact checkers cannot handle such tremendous information online in real time, AI technology can be leveraged to automate fake news detection. The first step leading to a sophisticated fake news detection system is the stance detection between statement and body text. In this work, we analyze the dataset from Fake News Challenge (FNC1) and explore several neural stance detection models based on the ideas of natural language inference and machine comprehension. Experiment results show that all neural network models can outperform the hand-crafted feature based system. By improving Attentive Reader with a full attention mechanism between body text and headline and implementing bilateral multi-perspective mathcing models, we are able to further bring up the performance and reach metric score close to 87%.