Temporal text-guided feedback-based progressive fusion network for multimodal sentiment analysis

The goal of text-guided multimodal sentiment analysis (MSA) is to fully leverage the unique sentiment advantages of textual features. However, existing research on multimodal sentiment analysis (MSA) predominantly relies on unidirectional text-dominant frameworks, neglecting the challenges posed by modality imbalance and sentiment conflicts in scenarios involving sarcasm or irony. Furthermore, the lack of effective modeling for sentiment information at different temporal steps in multimodal features limits the ability of MSA tasks to capture fine-grained sentiment characteristics. To address these issues, this paper proposes a framework named Text-guided Feedback Multimodal Progressive Fusion Network (TFPN). Specifically, we design a multi-layer Progressive Gated Learning Unit (PGU) module for progressive fusion, leveraging its unique feedback mechanism to dynamically adjust the sentiment contributions of each modality and mitigate modality imbalance caused by excessive text dominance. In addition, we introduce the Temporal Attention Fusion (TAF) module, capable of focusing on feature information across different temporal steps. Its distinctive temporal attention mechanism ensures the effective representation of sentiment information at each time step, thereby enhancing the overall sentiment recognition capability of the model. Experimental results on three benchmark datasets demonstrate that the proposed TFPN model significantly outperforms strong baseline models. • Text-guided Multimodal Sentiment Analysis. • The multi-layer progressive module design boosts multimodal sentiment interaction. • The feedback mechanism effectively addresses the modality imbalance issue. • Focus on sentiment expression at different time steps, capturing sentiment features.