Measuring the Protest Paradigm: LLM coding and machine learning approaches to Selection and Framing

The protest paradigm has been a central concept in research on media and social movements, documenting systematic patterns of protest marginalization through emphasis on violence, disruption, and official authority. Yet existing studies vary in how they operationalize the paradigm, raising a critical empirical question: do measures designed to capture media attention, the selection and salience of textual elements, converge with measures designed to capture framing, the interpretive structures that construct meaning? This study disentangles these dimensions by comparing two computational approaches: (1) document-level frame identification using a generative large language model (LLM), producing both content and framing scores, and (2) a sentence-level classifier to identify the same categories and aggregated via article structure and attention metrics into a weighted composite measure at the document level. Using news wire data from Latin America and Southern Europe (2000–2025), we examine nine protest-paradigm categories. Results reveal strong convergence for lexically concrete, recurrent frames (e.g., Violence, Law and Order, Decay of Morals) and weak or inconsistent alignment for interpretive or context-dependent frames (e.g., Troublemakers, Nuisance, Righteous Struggle). These findings highlight systematic methodological differences: LLMs capture implicit, cross-sentence framing, while sentence-level classifiers can validate only structurally explicit frames. The study underscores the importance of transparent aggregation procedures and human validation, and demonstrates that conclusions about protest coverage, media bias, and framing are contingent on the operationalization employed. By clarifying the boundary between attention and framing, this research contributes both to computational media analysis and to the conceptual refinement of the protest paradigm.