Context-Specific Embeddings of Extremism: A Regression Approach to Activists and Sympathizers in White Supremacist Spaces

• In Event: Collective Behavior and Social Movements Section Roundtables
  In Refereed Roundtable (60 minute): Table 7: The Role of Digital Platforms in the Practice and Study of Movements: The Case of the U.S.

Abstract

Activists and sympathizers often inhabit the same movement spaces, yet they may construct meaning in systematically different ways. We link 1,961,836 Discord messages from 7,530 users in U.S. white supremacist servers to 2,178 documented offline acts of activism reported by 778 individuals to examine how activists and non-activists differ in both linguistic style and the semantic organization of key movement terms. Descriptively, activists contribute a disproportionate share of content and exhibit a striking linguistic profile: compared to non-activists, they use fewer slurs and less profane, overtly violent rhetoric, while orienting more strongly toward coordination, discipline, and public-facing “optics.” These patterns suggest that movement escalation may involve not only intensifying grievance but also the strategic professionalization of discourse. To move beyond frequency and measure context-specific meaning, we estimate regression-based word embeddings using a la carte (ALC) embedding regression trained on domain-specific GloVe vectors. This approach tests whether the same term is embedded in systematically different semantic neighbourhoods across groups and time. Focusing on major political shocks (including the Charlottesville “Unite the Right” rally), we show that semantic divergence is most pronounced in anticipation windows: activists embed key terms in mobilization- and action-oriented contexts, whereas non-activists embed them in bystander, spectacle, and commentary contexts. Substantively, the findings clarify how extremist organizations may convert sympathizers into participants through cultural and organizational socialization. Methodologically, the study demonstrates how embedding regression can serve as a general tool for measuring group-conditional meaning in the study of culture, ideology, and social movements.

Authors

• Alessandro Giuseppe Drago, University of Notre Dame

• Yi-Cheng Hsieh, McGill University