模态缺失场景下基于生成重构和交互式自挖掘的多模态情感分析OA

In multimodal sentiment analysis tasks,real-world applications often face modality missing issues.Existing methods for missing modality generation heavily depend on automatically generated modality representations,which amplifies generation errors and limits generalization ability.To address this,this paper proposed the PRM framework.In both single-modal and dual-modal missing scenarios,the framework firstly used generative prompts and available modality information to estimate the mis-sing modality.It then designed a dual-modality-supported reconstruction mechanism that reduced single-source generation errors effectively.In the fusion phase,the framework introduced a self-mining operator to explicitly learn deep semantic features from non-missing modalities,and used a zero-slot insertion strategy to aggregate global contextual information.Experimental results show that,in both single-modal and dual-modal missing scenarios on the CMU-MOSI and CMU-MOSEI datasets,the PRM model improves accuracy and F1 by approximately 1％～3％on average.Moreover,the model demonstrates robust generaliza-tion ability in dynamic missing and cross-dataset transfer experiments,confirming its effectiveness and robustness in complex missing scenarios.