Angry White Men? Learning the Composition of Online Communities from Text

• In Event: Machine Learning and Text Analysis

Abstract

Political scientists are interested in the composition of online communities. A common assumption is that some of them are dominated by white men and serve as echo chambers for extreme far-right views that can later propagate into mainstream media or even motivate the participants to commit violent acts. A crucial problem with testing this assumption is the unknown identity of online users. Participants usually are neither required to provide their real names nor have incentives to do so. However, their online activities leave a large digital trace in the form of posted messages. In this article we use deep learning methods for predicting gender and ethnicity from text data and evaluate the performance of different approaches. We also investigate the potential algorithmic biases that could result from relying on social media posts by using alternative data sources, such as legislative speeches. And, finally, by applying our model we analyse the demographic composition of far-right online communities. The results of this study contribute to our understanding of far-right movements and extend the application of text-as-data methods by creating a model for gender and ethnicity classification based on text supplied by anonymous users.

Authors

• Tom Paskhalis, New York University

• Zhanna Terechshenko, New York University

• Joshua A. Tucker, New York University

• Jonathan Nagler, New York University