Social context in political stance detection: Impact and extrapolation

Stance detection is an important task with a wide range of high-impact social applications, including opinion polling and detecting propaganda, misinformation, and hate speech. In this work, we explore the performance and extrapolation power of political stance-detection models using an existing large-scale weakly-labeled Twitter dataset collected around the 2019 South American Protests. We construct transformer-based user and tweet encoders to embed users in a low-dimensional space using their posts and interactions. We then train heterogeneous graph attention networks to predict user stances and contrast their ability to extrapolate stance predictions to different country contexts and to future events. We find that leveraging users' ego-network in political stance detection improves in-country model performance for every country we examine. More notably, we find that leveraging a user's social context greatly enhances the ability of our stance detection models to extrapolate to new country contexts and future data.