Enhancing multimodal sentiment analysis with adaptive causality enriched fusion network and dynamic attention technique

Multimodal Sentiment Analysis (MSA) has emerged as a pivotal area of study, integrating information from text, image, video, and audio modalities to derive comprehensive emotional insights. While significant progress has been made, present techniques often treat these modalities equally, over-looking the dominant role of text in sentiment inference and failing to address redundant and irrelevant data generated during multimodal fusion. This paper discusses the Adaptive Causality-Enriched Fusion Network (ACEFN), designed to develop the robustness and generalization of sentiment prediction. ACEFN employs three innovative modules, such as the Contextual Se-mantic Refinement Module (CSRM), which maintains cross-modal coherence and reduces semantic drift, the Causal Inference Fusion Layer (CIFL), which identifies and integrates modality-specific causal pathways and the Dynamic Attention Filtering Unit (DAFU), which adaptively mitigates noisy or irrelevant features in real time. Extensive experimental evaluations were done on standard datasets CMU-MOSI, which determine the effectiveness of the ACEFN technique with an accuracy of 91.23%, correspondingly. Additionally, ACEFN exhibits robust performance in both regression and classification settings, with significant improvements under out-of-distribution conditions, underscoring its ability to generalize to real-world applications. By addressing the experiments on modality bias and redundant fusion, this study establishes ACEFN as a cutting-edge solution in MSA, paving the way for developing sentiment analysis in diverse multimodal frameworks.