Facebook Reactions as Controversy Proxies: Predictive Models over Italian News

Abstract

Discussion on social media over controversial topics can easily escalate to harsh interactions. Being able to predict whether a certain post will be controversial, and what reactions it might give rise to, could help moderators provide a better experience for all users. We develop a battery of distant supervised models that use Facebook reactions as proxies for predicting news controversy, building on the idea that controversy can be modeled via the entropy of the reaction distribution to a post. We create a Facebook-based corpus for the study of controversy in Italian, and test on it the validity of our approach as well as a series of controversy models. Results show that controversy and reactions can be modelled successfully at various degrees of granularity.