Beyond Polarity: Continuous Affect-Enhanced Multimodal Aspect-Based Sentiment Classification

摘要

Multimodal aspect-based sentiment classification (MABSC) requires aspect-level sentiment inference from textual-image data that jointly convey opinions. Yet most existing approaches primarily exploit discrete polarity patterns and generic visual embeddings, making them less effective when the affect is subtle, implicit, or expressed through imagery. In this work, we propose \textbf{\textit{VADE}}, a Valence–Arousal–Dominance~(\textbf{\textit{VAD}})-\textbf{\textit{E}}nhanced MABSC framework that brings continuous VAD signals into multimodal sentiment reasoning and learns emotion-sensitive image representations. Specifically, we design a VAD encoder to extract continuous affect cues from text for aspect-level sentiment reasoning. Furthermore, we fine-tune a CLIP-based image encoder on affect-enriched image–text pairs to obtain visual representations that are more sensitive to sentiment cues. To support the fine-tuning process, we construct an affect-enriched image–text dataset \textbf{\textit{Senti-COCO}} by rewriting MSCOCO captions with a multimodal large language model, which yields large-scale image-text pairs with richer affective expressions. Experiments on two mainstream datasets, Twitter-15 and Twitter-17, show that VADE achieves a new state-of-the-art performance, demonstrating the effectiveness of incorporating VAD signals for MABSC.