Answers are grounded in this paper and cite their sources.
Multimodal Sentiment Analysis (MSA) aims to infer human emotions by jointly modeling heterogeneous signals such as text, audio, and visual data. However, real-world constraints such as sensor failures, privacy restrictions, and bandwidth limitations often lead to missing modalities, making robust sentiment reasoning difficult. Existing methods either overemphasize shared representations while suppressing modality-specific emotional cues, or introduce noisy pseudo-features through imputation. To address these challenges, we propose HDE-LSR, a framework that integrates Hybrid Decoupled Experts and Latent Space Reconstruction to enhance robust multimodal sentiment reasoning under missing modality conditions. The Hybrid Decoupled Experts (HDE) module adopts a two-stage training strategy and uses dynamic routing to adaptively weight private and shared experts under different missing conditions, thereby improving representation discriminability and mitigating the overemphasis on shared representations. The Latent Space Reconstruction (LSR) module further recovers missing modality-private semantics from shared latent anchors, effectively reconstructing authentic features without introducing noisy imputation. By balancing cross-modal shared semantics and modality-specific details, HDE-LSR improves the authenticity and consistency of reconstructed features. Comprehensive experimental results on CMU-MOSI and CMU-MOSEI demonstrate that HDE-LSR achieves average ACC/F1 scores of 81.9/81.7% and 83.2/83.2%, respectively, substantially outperforming MoMKE in average performance and under most missing-modality settings.
No takes yet. Share an insight, caveat, or question.
Yang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: