Diverse Enough but with Common Views:Building a Global Stance Classifier on COVID-19

Abstract

Stance detection, which determines a user’s position on a specific topic through their generated content or interactions, has been widely studied for various domains. However, most existing work focuses on regional or community-specific topics, lacking a global perspective. In this paper, we investigate the ability to detect stance on the COVID-19 pandemic, a truly global issue transcending geographical and cultural boundaries. We compile a large, multilingual dataset of 7.9 million tweets related to COVID-19, accompanied by media content, spanning 3,516 users from 90 countries and 31 languages. Our objective is to develop an effective stance detection approach that can accurately predict users’ stances (pro-vax or anti-vax) regardless of their language or location. To achieve this, we propose a network-based method that leverages user interactions on Twitter, such as friends, likes, replies, and mentions, in addition to textual content. Despite the significant cultural diversity within our dataset, our approach demonstrates the ability to accurately predict users’ COVID-19 stance by analyzing their interaction signals and network homophily patterns. Our classification model achieves an F-score of 0.95 for both pro-vax and antivax user stances, surpassing state-of-the-art text-based methods. The findings suggest that echo-chamber effects and network homophily can extend beyond borders and languages, forming global patterns of polarization around certain topics. Our work highlights the potential of network-based approaches for stance detection on global issues and contributes insights into the challenges and opportunities of developing inclusive and robust models across diverse contexts.