Attention-Guided Fusion for Multimodal Aspect-Based Sentiment Analysis

Abstract

Multimodal Aspect-Based Sentiment Analysis aims to determine the sentiment polarity associated with specific aspects by jointly analyzing textual content and corresponding visual information. This fine-grained task plays a crucial role in applications such as brand monitoring, customer feedback analysis, and personalized recommendation systems. Despite recent progress in the field, existing models still suffer from two major limitations: suboptimal modality fusion and weak aspect-level alignment, both of which degrade overall performance. In this paper, we propose a novel attention-guided fusion framework designed two-stage attention mechanism. First, an attention-based fusion is performed between the aspect embeddings and the text embeddings to generate aspect-aware textual representations. These enriched textual features are then fused with visual embeddings using a secondary attention layer, enabling cross-modal interactions that are critical for sentiment interpretation. We evaluate our model on the MASAD multimodal benchmark dataset, where it consistently outperforms several state-of-the-art baselines across multiple categories. The results confirm the effectiveness of our approach in capturing aspect-specific sentiment cues by jointly leveraging textual and visual information in a unified and context-aware manner.