Research · 1 Jul 2026 · 5 min
Using an LLM to translate, not to judge
The design decision behind GEMF: turn visual and pragmatic meaning into structured evidence before supervised prediction.
01
The label is not in one modality
A meme can look harmless if its image and caption are read independently. GEMF starts from the opposite assumption: the useful signal often appears in the relationship between both, together with tone, implication and cultural context.
02
Mediator, not oracle
Gemini does not output the final class. It converts visual and pragmatic meaning into a structured textual representation. Supervised encoders then combine it with OCR, EEG and Ekman emotion features. This keeps the generative model in an evidence-producing role rather than treating it as an unaccountable judge.
03
Disagreement is part of the target
The project learns from soft labels because annotator disagreement can describe genuine ambiguity. That choice also makes calibration and thresholding first-class modelling decisions. The repository documents where this worked and where hard-label thresholds did not transfer reliably.
Takeaway
What I take from it
Use a general model to make hidden meaning explicit; keep the final decision in an evaluated, task-specific system.