HyperPT : detection and classification of hyperpartisan news articles

Abstract

The modern hyper-connected world brings with it an unprecedented rise in fake and hyperpartisan news, with anyone connected online harnessing the power of producing such fabricated information. Hyperpartisan news can be defined as extremely one-sided or biased news towards or against an entity. It differs from fake news by often exaggerating and sensationalising real-life events. With the spread of such malicious information, the otherwise subjective opinion of vulnerable consumers is compromised, twisted and possibly manipulated by some ulterior agenda - resulting in unprecedented and damaging outcomes as already seen in now worldwide known incidents. We hence give our contribution to addressing this issue by introducing HyperPT, a classification system for the automatic detection of hyperpartisan news articles. Throughout this study we experiment with a number of data representations, classification algorithms and external article features with the aim of creating an accurate and reliable classification system. In doing so gaining further insight into the nature of the hyperpartisan news article. From our experiments we conclude on an SVM-based classification system working on article features represented as deep contextualised ELMo embeddings. Moreover, we test the addition of sentiment within the classification while also experimenting with different news article lengths. Explainability A.I. is used to interpret the model’s decision-making and determine the influence of the article features on the classification. Finally we compare our system with the current state-of-the-art, achieving a mean accuracy score of 0.8220 to the other’s 0.8404. In doing so we hence present an alternative system for the classification of hyperpartisan news articles.