Deep Temporal Features and Multi-Level Cross-Modal Attention Fusion for Multimodal Sentiment Analysis

To address the challenges of insufficient multimodal feature extraction and limited cross-modal semantic diversity and interaction in multimodal sentiment analysis, this paper introduces Deep Temporal Features and Multi-Level Cross-Modal Attention Fusion (DTMCAF). Initially, a deep temporal feature extractor is developed, creating a multimodal temporal modeling network that combines bidirectional LSTMs with multi-head self-attention to capture multimodal features. Next, hierarchical cross-modal attention mechanisms along with feature-enhancement attention modules are designed to facilitate thorough information exchange between different modalities. Additionally, gated fusion and multi-layer feature transformations are employed to strengthen multimodal representations. Lastly, a multi-component collaborative loss function is proposed to align cross-modal features and optimize sentiment representations. Comprehensive experiments conducted on the CMU-MOSI and CMU-MOSEI datasets demonstrate that the proposed method outperforms current state-of-the-art techniques in terms of correlation, accuracy, and F1 score, significantly enhancing the precision of multimodal sentiment analysis.