Accurate emotion categorization in social media is critical for applications ranging from mental-health monitoring to market intelligence, yet it remains hampered by two key challenges: the scarcity of high-quality labeled data and the complexity of multimodal content. Here, we introduce Guided Polarity Prompt Learning for Enhanced Emotion Analysis (GPPLEA), a unified framework that addresses both challenges by injecting explicit polarity guidance into pre-trained language models and by leveraging self-supervised visual representation learning. First, we construct a compact few-shot dataset via stratified sampling to preserve label distributions under extreme annotation budgets. Next, we enrich image representations through a rotation-prediction pretext task, and we cast text classification as masked-token prediction guided by a library of positive and negative exemplar prompts. Finally, we fuse image and text embeddings in a shared transformer, steered by polarity prompts that anchor the model’s attention to emotional cues. Evaluated on four benchmark multimodal datasets, GPPLEA consistently outperforms state-of-the-art few-shot and full-data baselines, achieving up to a 2.3% absolute gain in accuracy under 1% training data. Our results demonstrate that guided polarity prompting not only amplifies learning from limited labels but also preserves robust generalization in real-world social media contexts.
Wang et al. (2026) studied this question.