ACCENT: Attentive Contrastive Cross-modal Emotion Network for Multimodal Emotion Recognition

Accurately identifying human emotional states is of great significance for building intelligent human-computer interaction systems. Multimodal sentiment analysis, by jointly modeling multiple sources of information such as text, audio, and visual signals, can effectively overcome the limitations of unimodal approaches, demonstrating stronger robustness and expressiveness in complex real-world scenarios. However, existing methods still face two key challenges: on one hand, significant distributional discrepancies exist across different modalities, making semantic alignment difficult; on the other hand, there is a lack of effective mechanisms to enhance the discriminability and stability of intra-modal representations. To address these issues, this paper proposes ACCENT (Attentive Contrastive Cross-modal Emotion Network), a multimodal sentiment analysis framework that combines intra-modal contrastive learning with cross-modal attention mechanisms. Specifically, ACCENT consists of two parallel modules: first, the Intra-Modal Contrastive Learning (IMCL) module, which enhances the robustness and discriminability of each modal representation by enforcing consistency between original features and their masked counterparts in the representation space; second, the Cross-Modal Attention (CMAT) module, which constructs an attention-based interaction mechanism that enables each modality to adaptively aggregate complementary information from other modalities, thereby capturing fine-grained cross-modal dependencies. The output features from both modules are subsequently fused and fed into a classifier for final sentiment prediction. Experimental results on three public benchmarks — CMU-MOSI, CMU-MOSEI, and CH-SIMS — demonstrate that the proposed ACCENT model achieves competitive performance across multiple evaluation metrics, validating the effectiveness of the proposed approach for multimodal sentiment representation learning.