Multi-Dilated Transformer for Multimodal Sentiment Analysis

Multimodal sentiment analysis (MSA) aims to synthesize textual, visual, and auditory information to more accurately infer sentiment polarity. However, complex and diverse causal interactions among modalities often result in spurious correlations in unbalanced datasets, which compromise model generalization and lead to biased or inaccurate predictions. To address these challenges, we propose a novel multimodal fusion framework based on a multi-dilated transformer architecture, enhanced with causal inference strategies including within-sample and cross-sample sampling. Unlike traditional attention mechanisms, the proposed Multi-Dilated Causal Attention Block (MDCAF) employs dilated attention heads to capture both local and global modality interactions across multiple temporal and semantic scales. This enables the model to focus on key causal cues while effectively suppressing misleading correlations. Our approach improves the interpretability, robustness, and generalization capability of multimodal sentiment models, making it better suited for real-world applications where data imbalance and complex multimodal dependencies are prevalent.