A Semantically-Driven Multimodal Sentiment Analysis Framework With Temporal and Synergistic Attention

Multimodal sentiment analysis aims to identify emotional tendencies from text, audio, and visual data, but existing methods often struggle with weak temporal modeling within modalities and shallow cross-modal fusion. The proposed temporal modeling and synergistic attention–based multimodal sentiment analysis framework can address these issues. Word-level features are first extracted from all modalities, then modeled using a state-gated long short-term memory network combined with multi-head attention to capture temporal emotional dynamics while filtering noise. A hierarchical collaborative attention mechanism is further designed to enable deep, fine-grained cross-modal semantic interactions. Experiments on the Carnegie Mellon University multimodal corpus of sentiment intensity and multimodal opinion sentiment and emotion intensity datasets show that the modeling and synergistic attention-based multimodal sentiment analysis framework achieves an F1 score of 87.3% and an mean absolute error of 0.426, it achieves a 1.2–1.5% improvement while simultaneously reducing mean absolute error to its lowest value, outperforming existing state-of-the-art approaches and demonstrating its effectiveness in modeling complex multimodal emotions.