作者
Shanliang Yang,Hao Wang,Yu Kun,Lei Wang,Cong Liu,Tao Wang,Jiebing You,Erik Cambria
摘要
Multimodal sentiment analysis seeks to emulate human-like comprehension by integrating cues from language, vision, and acoustics. However, existing models often grapple with two critical limitations: a failure to distinguish between dominant explicit sentiments and subtle implicit undertones, and an inability to robustly arbitrate the resulting inter-modal conflicts. To bridge this gap, we propose CASCADE-Net (Conflict-Aware Signal Cue Arbitration and DEcoupling Network), a novel hierarchical arbitration framework. CASCADE-Net is built upon two core components. First, a Perceptual Cue Decoupling (PCD) module employs a constrained dual-channel architecture to disentangle each modality's features into explicit and implicit representations. Second, and central to our framework, is the Cascaded Arbitration (CA) module, which introduces a predictive arbitration mechanism. It trains auxiliary predictors to forecast one modality's explicit cues from another's implicit cues, generating a quantifiable conflict signal from the prediction discrepancy. This signal then dynamically modulates a cascaded attention process, which first forms a preliminary judgment on explicit cues and then leverages the conflict signal to adaptively arbitrate and integrate implicit cues for a refined final decision. Extensive experiments on four benchmark datasets validate CASCADE-Net, which attains state-of-the-art Acc-2 of 85.84%, 85.92%, 81.15%, and 81.68% on CMU-MOSI, CMU-MOSEI, CH-SIMS, and CH-SIMS v2.0, respectively. On the conflict-rich CH-SIMS v2.0 benchmark, it exceeds the strongest baseline by 1.15 percentage points in Acc-2, confirming its effectiveness in handling complex, conflict-rich affective states.