Visual Framing at Scale: A Theory-Driven Computational Framework for Analyzing Protest Imagery with Generative AI

Le résumé fourni par la source

This study presents a theory-driven, three-stage computational framework for analyzing visual framing in protest imagery. Focusing on the Black Lives Matter movement, we examine how visual elements contribute to two well-established frames in protest media coverage: the protest paradigm and solidarity framing. Leveraging GPT-4o and OpenCV, our framework extracts denotative and semiotic features—such as police presence, contestation, solidarity actions, and color contrast—and links these features to higher-order frame classifications using interpretable logistic regression models. The framework includes: (1) feature definition and validation through generative AI and a feature extraction tool, supported by human coders; (2) model training; and (3) predictive application to unseen images. Results show strong alignment between human and machine annotations, as well as high predictive accuracy in identifying the protest paradigm or solidarity frame in BLM images. We also introduce an intra-prompt stability score for the generative AI model to help mitigate hallucination and enhance the reliability of its outputs. This study offers a scalable, replicable, and interpretable approach to visual framing analysis, bridging communication theory with advanced computational tools in the study of visual political communication.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.