GeoMVC detects misogyny in memes across 3 languages with 89% F1
AI that understands cultural sarcasm in memes achieves top ranks in Malayalam and Chinese.
GeoMVC tackles the challenge of detecting misogynistic intent in internet memes, where hateful messages often hide behind the semantic clash of images and text. The system introduces a Geometric Interaction Layer that replaces static feature concatenation with Hadamard products and cosine similarity between frozen visual and textual embeddings, learning how the two modalities interact. To handle noisy OCR and code-mixed transliterations (e.g., mixing English with Dravidian languages), a Multi-View Consensus strategy aggregates predictions across raw, length-filtered, and English-translated text views, reducing distribution shifts.
Submitted by Md. Ajwad Hossain for the CC-MMD Grand Challenge at the 28th ACM ICMI 2026, GeoMVC achieved strong results: F1 of 0.892 in Malayalam (Rank 2) and 0.895 in Chinese (Rank 3) on Task A. However, performance dropped sharply in Tamil (F1 0.521, Rank 5), revealing persistent challenges with localized transliteration and culturally grounded sarcasm. The paper emphasizes that while geometric interaction modeling captures cross-modal alignment, code-switching and context-dependent humor remain open problems for moderation systems.
- Geometric Interaction Layer uses Hadamard products and cosine similarity to model cross-modal alignment, outperforming static concatenation.
- Multi-View Consensus strategy merges predictions from raw, length-filtered, and English-translated text to handle OCR noise and code-mixing.
- GeoMVC scored F1 0.892 on Malayalam and 0.895 on Chinese, but only 0.521 on Tamil due to transliteration and sarcasm challenges.
Why It Matters
Automated moderation must handle cultural nuance and code-mixing – a gap GeoMVC highlights for future AI safety systems.