Quantifying Content Polarization on Twitter

Abstract

Social media like Facebook and Twitter have become major battlegrounds, with increasingly polarized content disseminated to people having different interests and ideologies. This work examines the extent of content polarization during the 2016 U.S. presidential election, from a unique, "content" perspective. We propose a new approach to quantify the polarization of content semantics by leveraging the word embedding representation and clustering metrics. We then propose an evaluation framework to verify the proposed quantitative measurement using a stance classification task. Based on the results, we further explore the extent of content polarization during the election period and how it changed across time, geography, and different types of users. This work contributes to understanding the online "echo chamber" phenomenon based on user-generated content.