Answers are grounded in this paper and cite their sources.
Multimodal aspect-based sentiment analysis (MABSA) is challenged by visual noise, as images in social media posts are often irrelevant to, weakly related to, or only partially informative for specific textual targets. Although existing methods attempt to fuse heterogeneous textual and visual information, their reliance on global visual representations or implicit feature fusion may be insufficient to filter weakly related visual cues, thereby limiting fine-grained aspect-level reasoning. To address these issues, this paper proposes SAGA-Net, a unified Similarity-Aware Aspect-Guided Relation Graph Network for MABSA. SAGA-Net first estimates aspect-specific visual reliability to dynamically suppress unreliable visual information before multimodal fusion. It then employs a cross-modal aspect-guided alignment module, where aspect representations serve as semantic queries to extract aspect-relevant visual evidence from region-level features. Furthermore, SAGA-Net constructs a heterogeneous aspect-text-visual graph and introduces a relation-aware graph attention network that incorporates edge weights into attention computation for explicit evidence propagation. Extensive experiments on Twitter-2015 and Twitter-2017 across three core tasks, including JMASA, MATE, and MASC, show that SAGA-Net achieves competitive performance against representative baselines. Additional ablation and diagnostic analyses further validate the effectiveness of visual reliability estimation, aspect-guided evidence selection, and structured cross-modal relation modeling.
No takes yet. Share an insight, caveat, or question.
Zhang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: