Cosine Siamese Models for Stance Detection

Fake news detection has received much attention since the 2016 United States Presidential election, where election outcomes are thought to have been influenced by the unregulated abundance of fake news articles. The recently released Fake News Challenge (FNC) aims to address the fake news problem by decomposing the problem into distinct NLP tasks, the first of which is stance detection. We employ a neural architecture consisting of two homologous subnetworks for headline and body processing and a subsequent node for headline/body comparison and stance prediction. Headlines and bodies are represented with a weighted bag-of-words combination of word vectors passed through a ReLU, where the weights are learned. Stance is quantified by computing the cosine similarity of these weighted bag-of-words representations, and the score is regressed to a relaxed, continuous label space in which the true discrete labels are posited to lie. Our model, which outperforms other recurrent methods, achieves an FNC score of 0.891 out of 1.00, a 0.10 increase from the published 0.79 FNC baseline. The cosine similarity function induces a natural geometry among the learned headline and body representations, with unrelated inputs generally orthogonal to each other and agreeing inputs nearly collinear. Our findings implicate the importance of the optimization objective, as opposed to the architecture of the subnetwork models, to success in stance detection, echoing recent work demonstrating the competitiveness of weighted bag-of-words models for textual similarity tasks.