Large Language Models and Social Movement Framing: A Hybrid Inductive-Deductive Approach to Cross-National Media Analysis

• In Event: Using AI as a Tool for Social Analysis: Prospects and Challenges

Abstract

The framing perspective offers powerful conceptual tools but remains difficult to operationalize at scale without sacrificing interpretive validity. Qualitative coding retains contextual depth yet lacks scalability, while computational approaches such as topic modeling often conflate latent semantic clusters with theoretically defined frames.

This study proposes a formal AI-assisted workflow for large-scale social movement framing analysis: a hybrid inductive-deductive design that embeds large language models (LLMs) in a validated coding architecture. Using a corpus of 11,275 news articles covering the 2019 Hong Kong Anti-Extradition Movement in China, the UK, and the US, I first generate candidate frames inductively and formalize them into a theory-driven codebook distinguishing master frames, collective action frames, and diagnostic, prognostic, and motivational tasks. After establishing substantial intercoder reliability, I operationalize the validated codebook through structured prompting and systematically evaluate human-LLM alignment across GPT-4, Claude, and LLaMA-3. Following validation, the best-performing model is deployed to classify the full corpus, producing frame assignments at the population scale. This design treats LLMs as constrained classification instruments rather than generative black boxes.

Substantively, the analysis reveals structured cross-national divergence in framing and a systematic dominance of prognostic over diagnostic frames, suggesting that media narratives repositioned the Hong Kong protests from a localized policy dispute to a symbolic arena of geopolitical competition.

Methodologically, the study advances a formal framework for translating abstract sociological constructs into operational decision rules suitable for AI implementation. It demonstrates how interpretive categories can be measured at the population scale without collapsing into undifferentiated semantic clusters, introduces a replicable strategy for distinguishing topics from frames, and proposes a method for identifying framing bias as a measurable property of discourse. The paper also addresses key challenges in AI-assisted research, including construct validity, reproducibility, model drift, and the redistribution of interpretive authority between researcher and algorithm.

Author

• Hanning Wang, University of Pittsburgh