Abstract
Multimodal Aspect-Based Sentiment Analysis (MABSA) enhances sentiment detection by integrating text with complementary modalities, such as images, for a more comprehensive understanding of aspect-level sentiment. Despite notable progress, existing MABSA methods still lack strong aspect awareness, particularly in sentences with multiple aspects and conflicting sentiments or when visual cues are misaligned with the text. Together, these issues hinder accurate modeling of aspect-focused and cross-modal dependencies, limiting the effectiveness of MABSA. To address these challenges, we propose DualKanbaFormer, a novel framework with parallel Textual and Visual KanbaFormer modules tailored for MABSA. At its core, the model introduces Aspect-Driven Sparse Attention (ADSA) with an aspect-aware attention gating mechanism, which dynamically regulates the contributions of Scope, Focus, and Proximity attention branches conditioned on the target aspect. This ensures precise alignment between aspects and opinion signals. In addition, the selective state space model Mamba strengthens aspect–opinion alignment by modeling long-range dependencies across modalities. To enhance representational expressiveness and training stability, Kolmogorov–Arnold Networks (KANs) and Dynamic Tanh (DyT) replace conventional feed-forward and normalization layers. Finally, a CNN-based multimodal gated fusion module adaptively regulates cross-modal interactions, ensuring robust alignment between textual and visual features. Experiments on two benchmark datasets demonstrate that DualKanbaFormer outperforms several state-of-the-art models.
Funding Information
  • National Natural Science Foundation of China (62577006)

This publication has 27 references indexed in Scilit: