MSPR: Multi-stage Semantic-Guided Prompt Reasoning for Multimodal Sentiment Analysis
Abstract
Understanding sentiment in social media requires integrating heterogeneous textual and visual cues. However, affective meaning is rarely expressed explicitly and is often progressively constructed from perceptual impressions, semantic content, and higher-level emotional interpretation. Most existing methods still formulate multimodal sentiment analysis as a one-step fusion or alignment problem, making it difficult to capture this progressive nature of affect understanding. To address this issue, Multi-stage Semantic-guided Prompt Reasoning (MSPR) is proposed to model multimodal affect understanding as a three-stage process over Tone, Content, and Emotion. Specifically, offline Chain-of-Thought (CoT)-guided hierarchical captions provide progressive stage-wise supervision, grounding perceptual, semantic, and emotional cues in sequence. Building on this structured supervision, MSPR introduces prototype-guided cross-modal prompt interaction to transfer affective cues across textual, visual, and prototype representations. It further employs a prototype-guided dual-path contrastive alignment strategy to regularize the learned affective space and improve representation robustness. Extensive experiments on four public benchmarks demonstrate that MSPR consistently outperforms strong baselines and learns more structured and interpretable affect representations.