Emotions Need Context: A Large Language Model Based Exploration of Emotion Relationships and the Call for a New Model

This study investigates the dimensional representation of human emotions using large language models (LLMs) through a dual-method approach: static embeddings and context-driven prompts. Findings revealed that prompting significantly enhances the quality of emotional representations, as UMAP modeling of prompt-derived results produced distinct and interpretable clusters, validated through clustering and statistical analysis. In contrast, embeddings exhibited flat elbow values, lacking groupability. Moreover, context-driven outputs uniquely correlated with human rating of Arousal in emotion, capturing nuanced emotional dimensions. Both methods showed a lack of separation between Pleasure and Dominance, consistent with some prior research, suggesting the limitations of traditional dimensional models and the need for revised frameworks that better reflect language-expressed emotions. This study highlights the importance of context in emotion modelling, and support a transition from theory-driven to data-driven approaches. The findings proposed context-sensitive emotion models to address the demands of modern AI and affective computing.