多模態情緒分析（Multimodal Sentiment Analysis， MSA）隨著預訓練語言模型與跨模態注意力機制的引入而取得顯著進展，然而一項根本性的限制至今仍未解決，即現有方法將維度較高的文字、維度稀疏的聲學與維度緊湊的視覺三種模態，統一透過共享或均勻投影的管線進行處理。此種一致性處理方式導致融合與加權元件在本質上偏向表徵能力較強的文字流，而聲學與視覺模態則長期處於未充分利用的狀態。本研究提出 MSE-MoE，一個以嚴格的「單一模態對應單一專家」設計原則為核心的模態特定專家混合框架，其中文字、聲學與視覺模態各由獨立的專家管線學習模態專屬的情感表徵，且各管線之間不共享任何參數。成對跨模態注意力模組先以跨模態上下文資訊豐富各專家的表徵，再由 CrossModalContextRouter 透過文字條件式交叉注意力計算樣本自適應路由權重，並依據每個樣本的模態可靠性對三個專家的輸出進行加權聚合。非對稱負載均衡損失與兩階段訓練流程共同穩定路由機制，防止因多模態特徵空間中固有的維度失衡而導致的專家崩潰問題。本研究在 CMU-MOSI 與 CMU-MOSEI 資料集上，分別於時間對齊與非對齊條件下對 MSE-MoE 進行評估，在統一評估協議下，所提框架於所有評估指標與實驗條件下均達到最佳性能表現。消融實驗進一步驗證了每個提出元件均對整體性能有所貢獻，而路由機制分析則揭示 CrossModalContextRouter 能夠學習有意義的樣本相依模態重要性權重，而非退化為僅依賴文字模態的決策路徑。;Multimodal Sentiment Analysis (MSA) has advanced substantially through the adoption of pre-trained language models and cross-modal attention mechanisms， yet a fundamental limitation persists in that existing methods process the dominant textual， sparse acoustic， and compact visual modalities through shared or uniformly projected pipelines. This uniform treatment causes the fusion and weighting components to be intrinsically biased toward the richer textual stream while the acoustic and visual modalities remain underutilized. In this work， we propose **MSE-MoE**， a Modality-Specific Expert Mixture-of-Experts framework built around a strict *one-modality， one-expert* design principle， in which independent expert pipelines for text， acoustic， and visual modalities learn modality-specific affective representations without parameter sharing. A pairwise Cross-Modal Attention module enriches each expert′s representations with cross-modal context before a CrossModalContextRouter computes sample-adaptive routing weights via text-conditioned cross-attention， aggregating the three expert outputs according to per-sample modality reliability. An asymmetric load-balancing loss and a two-stage training procedure together stabilize routing and prevent the expert collapse that would otherwise arise from the dimensional imbalance inherent in multimodal feature spaces. We evaluate MSE-MoE on CMU-MOSI and CMU-MOSEI under both aligned and unaligned temporal conditions， and the proposed framework consistently achieves the best performance across all evaluated metrics and conditions under a unified evaluation protocol. Ablation studies further confirm that each proposed component contributes positively to overall performance， and routing analysis reveals that the CrossModalContextRouter learns meaningful， sample-dependent modality importance weights rather than degenerating into a text-only decision pathway.
