LLM-Inferred Narrative Frames in Geopolitical Conflict Reporting: An Exploratory Zero-Shot Approach

This study explores the use of large language models (LLMs) for zero-shot classification in unsupervised media framing analysis. A two-stage pipeline combining BART-based frame inference and FLAN-T5 semantic filtering was used to identify narrative structures at the article and sentence levels for 96 English-language articles covering the Gaza Freedom Flotilla incident of May 2, 2025. The analysis revealed recurrent emphasis on Legal/Human-Rights, Responsibility/Attribution, and Conflict/Security frames, with associations between key entities and specific narrative themes. While the approach demonstrates the potential of LLMs for scalable framing detection without manual annotation, it also reveals critical limitations. The models tended to over-assign broad, surface-level frames and showed reduced sensitivity to contextual or moral nuances. Findings represent language-based model interpretations rather than verified editorial intent. The study contributes to computational media research by assessing the methodological opportunities and interpretive boundaries of automated narrative inference.