As polarization deepens on social media, understanding how political memes shape online narratives becomes increasingly important. Prior work either relies on qualitative frameworks that are hard to scale or on computational methods lacking theoretical grounding; both treat narrative categories in isolation and focus solely on moral portrayals of identity.
We propose a theory‑driven framework with three dimensions: social identities (gender, political affiliation, ethnicity, etc.), viewpoints (ingroup vs. outgroup) and narrative roles (moral character, intelligence, status, power). The design preserves social‑science insight while remaining amenable to machine‑learning models.
Guided by this framework we built IRON-Meme, a human‑annotated benchmark containing 1,110 political memes and a substantially larger set of templates than existing resources. Each meme is labeled for identity, viewpoint and role.
Statistical analysis shows that most memes assign derogatory roles and that identities, roles and templates are systematically intertwined. We evaluated nine vision‑language models; frontier‑scale and reasoning‑capable models achieved the best scores, yet accuracy on rare identity and role labels dropped sharply. Adding template background knowledge did not improve role classification, suggesting that meme interpretation relies more on how memes are adapted for specific communicative purposes.
The IRON-Meme dataset and code are released publicly for the research community.
Review