Abstract: In multimodal sentiment analysis tasks, effective fusion of sentiment information between different modalities is hindered due to three challenges: nonverbal modal information being insufficiently utilized, fine-grained associative modeling for cross-modal interactions being inadequately established, and hierarchical semantic fusion mechanisms being imperfectly designed. To address these issues, a multimodal sentiment analysis method with cross-modal feature enhancement and hierarchical MLP communication is proposed in this paper. A progressive fusion architecture is constructed, where nonverbal modal information is first enhanced through a cross-modal attention mechanism, enabling many-to-many cross-modal fine-grained interactions to be captured. Subsequently, a lightweight hierarchical MLP communication module is designed to implement hierarchical feature interactions in both horizontal and vertical dimensions, through which cross-modal deep semantic fusion is achieved. It is demonstrated by the experimental results that compared with the suboptimal model on CMU-MOSI, The Acc2 and F₁ values increase 0.89 percentage points and 0.77 percentage points compared with the suboptimal model. Furthermore, all metrics in comparative experiments on CMU-MOSEI are shown to surpass those of baseline models, with the Acc2 and F₁ values being elevated to 86.34% and 86.25%.

Key words: multimodality, sentiment analysis, cross-modal attention, hierarchical MLP communication, gating units