基于对抗训练与多模态自适应特征融合的情感分析

打开文本图片集

关键词：自适应权值；对抗训练；多分类；交叉注意力；掩码特征增强中图分类号：TP391 文献标志码：A 文章编号：1001-3695（2025）12-006-3566-08doi：10.19734/j. issn.1001-3695.2025.05.0138

Sentiment analysis multimodal adaptive feature fusion based on adversarial training

Feng Guanga，Huang Rongcanb，Zhou Yuanhuaa，Xiang Fengʰ，Yang Yanru Zheng Runtingʰ，Liu Tianxiangʰ，Li Weichenʰ （utomation.holompueienceadhogGuandogUniesityfTchologyuangzou5n）

Abstract：Multimodalsentimentanalysis plays akeyroleinonline classroom-basedsmarteducation，whichhasatracted increasingatentioninrecentyears.However，currentmethodsfail tullexploit inter-modalcomplementarity，overlook the dominantrole thetextual modality，and strugle with noiseandrobustnessissues.To adressthese problems，this paper proposedamultimodaladaptivefeaturefusionapproachbasedonadversarial taining.Itusedaudioandvideoaslow-levelfeatures，enhancedtheirinformationdensitythroughcross-atentioninteraction，andappliedadaptiveweightingtodynamicayrefinethe interactionresults.Itintroducea masking mechanism to preserve keysentencerepresentations inthetext modality， andusedthecross-modalatentionturtheroptimizeudioandvideo.Inadition，itdesignedanadversarialframeworkwhere thefeatureextractorandmodal discriminatorjointlyalignedandfusedmultimodalfeatures inashared space.Experimntalresults show that the proposed model achieves superior performance on the MOSI and MOSEI datasets.

KeyWords：adaptive weights；adversarial training；multi-categorization；cross-atention；mask featureenhancement

0 引言

在人工智能与深度学习技术迅猛发展的时代，情感识别已成为人机交互领域的核心研究方向，并在智能客服、虚拟助手、智能推荐等应用场景中发挥着举足轻重的作用[]。（剩余20038字）