Deep learning model on stance classification

The process of identifying and assigning the relationship between two bodies of text is referred to as stance classification. Given a headline and the corresponding body they are compared and their relationship is classified into one of the following two classes — unrelated or related where related is further divided into agree, disagree and discuss. In this article, data is collected from news articles which contains headlines and bodies. We call a headline and the corresponding body as a pair. Deep learning models are applied to these pairs. We applied bidirectional Long Short-Term Memory (LSTM) model and multi-layered perceptron (MLP) model and obtained accuracies of 83.5% and 78% respectively. The accuracy calculation is based on a weighted scheme. The correctly classified unrelated pair has a score of 0.25. A pair correctly classified as related yields a score of one only if the the sub-relationships of agree, disagree and discuss are correctly identified; otherwise, the score is 0.25.