基于统一对齐与多阶段融合机制的多模态情感分析模型OA

To address the modality heterogeneity,time-varying contributions,and limited semantic abstraction in multimodal sentiment analysis,this paper proposed a three-stage closed-loop fusion model(MIFA).Specifically,it followed an align-ment-regulation-abstraction fusion path that comprised unified semantic alignment,dynamic fusion regulation,and hierarchical semantic abstraction.Firstly,this model used unified semantic alignment to express heterogeneous modalities in a shared space.Next,it used contextual gating and channel modulation to estimate modality/channel weights and refine representa-tions.Finally,it used a hierarchical residual abstraction to enhance discriminative features,forming an end-to-end closed loop.Experiments on CMU-MOSI and CMU-MOSEI show that MIFA achieves Acc2 of 86.43％/86.03％and F,of 86.42％/85.81％,Acc7 reaches 45.04％/50.41％,and MAE is 0.689/0.532,outperforming mainstream baselines.These results indicate that MIFA stabilizes cross-modal alignment and adaptively regulates information flow,improves sentiment classification and intensity regression,and offers practical potential for complex multimodal scenarios.