Bridging the gap, not forcing the tie: dual-space alignment and fusion framework for toxic memes detection

Yulin Lei, Jin Yang, Huijia Liang, Tianrui Li

Abstract: Highlights•Geometry-area alignment preserves modality-specific semantics and structure.•Counterfactual weighting reduces bias from dominant modalities.•Adaptive multi-task fusion integrates semantic, alignment, and decision cues.•ALFUS shows strong cross-lingual performance on toxic meme datasets.

External IDs:doi:10.1016/j.inffus.2025.103992