基于全局-局部交互与对比学习的多模态对话情感识别OA

This paper proposed a multimodal emotion recognition method based on global-local interaction and contrastive learning to address the lack of global guidance,difficulties in cross-modal semantic alignment,and modal learning imbalance in conversational emotion recognition.The method introduced a semantic-guided global-local interaction mechanism,where a global semantic hub directed deep feature fusion through adaptive attention allocation.It further constructed a text-audio-visual tri-modal contrastive learning framework to align and complement modal representations within a shared semantic space.Addi-tionally,it designed a modality balanced optimizer to monitor modal performance and dynamically adjust learning rates,mitiga-ting modal dominance.Experiments on the IEMOCAP and MELD datasets achieve accuracies of 76.09％and 69.66％,with weighted F1-scores of 76.20％and 68.79％,respectively,significantly surpassing existing approaches.The results confirm the method's effectiveness in enhancing multimodal collaboration and emotion recognition.