Abstract:To address the limitations of current multimodal aspect-based sentiment analysis (MABSA) approaches, including susceptibility to visual noise interference and oversimplification of aspect representations, this study proposes a visual denoising and aspect-enhanced (VDAE) method. This method systematically integrates image-text relevance filtering and semantic relevance filtering to achieve visual noise suppression, in combination with multimodal semantic enhancement. VDAE consists of two components: the multi-perspective aspect enhancement module (MPAEM), which enhances aspect representations through cross-modal interactions, and the sentiment-semantic-syntactic graph convolutional network (S3GCN), which enables effective utilization of visual information by jointly modeling sentiment cues. Experimental results on datasets Twitter2015 and Twitter2017 benchmarks indicate that the proposed method achieves significant performance improvements over baseline methods across all three MABSA sub-tasks.