MultiPolar: Multimodal Learning to Measure Polarization through Social Media

• In Event: New Methods and Frameworks to Study Digital Politics

Abstract

This project offers a new measure of online political preferences by developing and testing artificial intelligence systems to better understand how user interactions convey partisan animosity and influence the overall levels of polarization found on social networks. Today, election campaigns are increasingly conducted on social media platforms, with important implications for the functioning of democracy. Scholars have shown that political polarization is very high among the most active users on websites like Twitter or Facebook, but the overall effects of social media consumption on the political preferences of the mass public are still unclear. In this research project, we begin to solve this puzzle by introducing a new multimodal joint embedding model to analyze network, text and image data in social media interactions. Our integrated framework will combine several existing approaches to study online activities, such as graph convolutional network (GCN) models for analyzing network relations, Robustly Optimized BERT Pre-training Approach (RoBERTa) for analyzing text, and Deep Residual Learning (ResNet) for analyzing images. We will integrate these machine learning tools into a comprehensive framework to estimate the political preferences of social media users on Twitter, as well as the new conservative platform TRUTH Social. We will first develop the framework using Twitter data collected during the 2020 US election. We will then set up a pipeline to collect and process several million social media posts from both Twitter and TRUTH Social during the 2022 US midterm election, to measure how political polarization is changing during this campaign. By combining information on the structure of the social networks, likes, hashtags, re-posts, images, and the content of messages, our integrated approach will estimate the partisan preferences of politically active users in these online communities. Our index will calculate the degree of political conflict on a daily basis by analyzing changes over time in the distance of partisan clusters. We will validate our measure by conducting both synthetic and real data analyses, ranging from controlled tests of particular aspects of the model to full-scale comparisons with other methods over months of data. We will also use different gold standard measures, such as the voting records of Members of Congress, the profile biography of users, primary registration records, as well as an online survey of users active on social networks, to test the accuracy of our political preference index with external data. This new measure represents the first stage of a broader research project that aims to identify the causes and consequences of online political conflict on social networks. It should generate a considerable amount of interest outside of academia, including from government officials and information technology personnel who determine policies on social media platforms. Our aim is to provide an original and rigorous tool to monitor online political conflict in order to foster more reasonable and civil exchange of viewpoints, which is absolutely essential in our democratic system. We will develop a generalized scalable index of political preferences based on social media interactions that can be used across different social media platforms with our publicly available data and code. Given that social media platforms are increasingly accused of weakening democratic norms, it is imperative that we develop a more standardized measure of online political conflict in order to understand the causes of extreme polarization. This project represents a first necessary step in this direction by proposing an effective AI-based solution to monitor the information space proactively and promote a healthy and productive democracy in online communities.

Authors

• Kellin Pelrine, McGill University

• Reihaneh Rabbany, McGill University

• André Blais, University of Montreal

• Jean-Francois Godbout, Universite de Montreal