基于因果掩码控制与反事实干预机制融合的多模态仇恨模因检测

打开文本图片集

关键词：多模态学习；仇恨模因检测；因果关系；反事实干预；注意力机制中图分类号：TP391 文献标志码：A 文章编号：1001-3695（2025）12-005-3559-07doi：10.19734/j.issn.1001-3695.2025.04.0137

Multimodal hate meme detection based on fusion of causal mask control and counterfactual intervention mechanism

XuChao，Pu Yue，Chen Yong，SunKaili† （SchoolofComputerScience，NanjingAuditUniversity，Nanjing211815，China）

Abstract：Hateful memesfunctionasemerging vectors ofviolence insocialmedia，derivingsemanticdepthfromsocio-cultural contexts.This inherentcomplexityposescorechallenges tocontentsafety governance.Existing research prioritizeshallwmodalityoptimization whileneglectingcausal dependencies insemanticalignment，resulting incross-modalsemanticdeviations. Toaddressthisgap，thispaperproposedamultimodalhatedetection model integratingcausal mask controlandcounterfactual intervention mechanisms.The modelemployedvisual-textual encoding followedbyacausal mask control moduleforexplicit cros-modaldependencymodeling.Furtheritintegratedcounterfactualinterventionmechanismtomaskscriticalfeatureuits， andquantifiedpredictionprobabilityshifts，andidentifiedcausalpathwaynodesforhatepropagation.Thisprocessgenerated counterfactualsemanticdeviationvectors toadaptivelyrecalibrateatentionweightsbasedoncontextualsemantics.Experimental results demonstrate F1 -scores of 71.8% on MUTE and 68.8% onMultiOFF datasets. The model effectively mitigates spuriouscorelationsbetweenbackground noiseand hatesemantics，significantlyenhances detectionaccuracycompared toestablished baselines.

Keywords：multimodal learning；hate meme detection；causalrelationship；counterfactual intervention；atention mechanism

0 引言

互联网中仇恨言论和极端内容的蔓延会对社会造成严重的负面影响，其载体从纯文本扩展至图像、视频及表情符号等多模态载体中。（剩余19380字）