Multimodal Aspect-Based Sentiment Analysis via Evidence-Guided Region Calibration

Abstract

Multimodal Aspect-Based Sentiment Analysis (MABSA) aims to identify aspect-level sentiment from user-generated image-text posts. Although visual information can provide useful context for sentiment reasoning, it is not uniformly beneficial for aspect-level prediction. Different image regions may serve distinct roles: some provide content support for post understanding, some offer sentiment-related evidence, while others introduce irrelevant or misleading signals. To address this challenge, we propose VEGA, a MABSA framework that integrates visual evidence construction with evidence-guided region calibration. Specifically, VEGA first leverages a multimodal large language model to construct two complementary types of visual evidence from each image-text pair: content-aware visual evidence and context-aware sentiment evidence. These evidence streams are then used to calibrate region-level visual representations through a dual-support cross-attention mechanism, enabling image regions to contribute according to their roles in content understanding and sentiment inference. Finally, the calibrated visual features are integrated with the original text in a unified encoder-decoder architecture for both multimodal aspect sentiment classification and joint multimodal aspect-sentiment analysis subtasks. Experiments on public benchmarks demonstrate that VEGA consistently outperforms strong task-specific baselines. Further analyses confirm the effectiveness of visual evidence construction and evidence-guided region calibration.