CADET: Cycle-Consistent Adversarial Disentanglement With Text-Guided Enhancement for Multimodal Sentiment Analysis

Multimodal sentiment analysis (MSA) integrates language, vision, and audio for robust emotion recognition. A core challenge lies in balancing cross-modal consistency and modality-specific private information: While consistency ensures stable semantic alignment, it often suppresses fine-grained modality cues. Conversely, excessive emphasis on private information enriches detailed representations, but may introduce noise, particularly in the audio and visual domains. To address this, we propose Cycle-Consistent Adversarial Disentanglement with Text-Guided Enhancement (CADET), a unified framework for disentangled multimodal representation learning. CADET incorporates three complementary modules.First, an adversarial autoencoder with cross-modal contrastive learning separates shared and private representations and enforces instance-level semantic compactness to improve representation discriminability. Second, a cycle-consistent generative adversarial module introduces bidirectional reconstruction to preserve modality fidelity and guarantee semantic reversibility.Third, a two-stream text-guided enhancement mechanism leverages the dominant role of language to refine audiovisual features through cross-modal attention, thereby strengthening finegrained semantic representation within the private subspace.Experiments on CMU-MOSI and CMUMOSEI demonstrate that CADET consistently outperforms state-of-the-art baselines, validating the effectiveness of disentanglement, cycle consistency, and text-guided enhancement in advancing multimodal sentiment analysis.