A new AI system built for a research competition can spot misogynistic memes in multiple languages, but its accuracy swings wildly depending on which one.
Researchers developed GeoMVC (Geometric Interaction and Multi-View Consensus) for the CC-MMD Grand Challenge at ICMI 2026, a contest focused on catching misogyny in internet memes. Memes are hard to moderate because the hateful intent often lives in the clash between image and caption, not in either one alone. GeoMVC addresses that with a Geometric Interaction Layer that compares frozen visual and text embeddings using Hadamard products and cosine similarity, instead of simply concatenating features. To handle messy OCR output and code-mixed transliteration, where users blend scripts and languages, the system also runs predictions across three versions of the same text: raw, length-filtered, and English-translated, then combines the results into a consensus.
The approach placed second in the competition's Malayalam partition (Macro F1: 0.892) and third in Chinese (0.895), results that suggest the geometric-alignment trick genuinely helps parse culturally coded hate speech. But it ranked only fifth in Tamil, with the F1 score dropping to 0.521, a reminder that "multilingual AI" often means strong in a couple of languages, not universally capable.
The paper's own error analysis blames code-mixed sarcasm and localized transliteration for the Tamil shortfall. Translation: the model still doesn't get the joke, or the insult, equally well in every language.