Abstract
Multimodal sentiment analysis (MSA) benefits from integrating diverse modalities (e.g., text, video, and audio). However, challenges remain in effectively aligning non-text features and mitigating redundant information, which may limit potential performance improvements. To address these challenges, we propose a Hierarchical Text-Guided Refinement Network (HTRN), a novel framework that refines and aligns non-text modalities using hierarchical textual representations. We introduce Shuffle-Insert Fusion (SIF) and the Text-Guided Alignment Layer (TAL) to enhance crossmodal interactions and suppress irrelevant signals. In SIF, empty tokens are inserted at fixed intervals in unimodal feature sequences, disrupting local correlations and promoting more generalized representations with improved feature diversity. The TAL guides the refinement of audio and visual representations by leveraging textual semantics and dynamically adjusting their contributions through learnable gating factors, ensuring that non-text modalities remain semantically coherent while retaining essential crossmodal interactions. Experiments demonstrate that the HTRN achieves state-of-the-art performance with accuracies of 86.3% (Acc-2) on CMU-MOSI, 86.7% (Acc-2) on CMU-MOSEI, and 80.3% (Acc-2) on CH-SIMS, outperforming existing methods by 0.8-3.45%. Ablation studies validate the contributions of SIF and the TAL, showing 1.9-2.1% performance gains over baselines. By integrating these components, the HTRN establishes a robust multimodal representation learning framework.