Context over Categories - Implementing the Theory of Constructed Emotion

Emotion analysis is a critical research area with applications ranging from content moderation to personalized systems. Despite its importance, many approaches rely on traditional models, such as Ekman’s universal emotions theory, which reduces emotions to static, predefined categories. This oversimplification neglects the complexity and contextual variability of human emotions. Drawing on Lisa Feldman Barrett’s Theory of Constructed Emotion introduces a novel, context-aware approach to emotion analysis. A central contribution is the development of the “context sphere”, a personalized construct derived from user behavior data. To our knowledge, this is the first operationalization of Barrett’s theory for computational methods, enabling the dynamic modeling of emotions as emergent and context-dependent phenomena. The Design Science Research methodology is employed to develop a context-aware emotion analysis pipeline, which utilizes advanced Large Language Model (LLM) prompting strategies, including Role-Play prompting and Controlled Generation, to align with constructed emotion principles. A case study in online content moderation demonstrates the feasibility of this approach, showing that the “context sphere” facilitates nuanced and contextually aware emotion analyses that surpass traditional model limitations. Furthermore, a Semi-Structured Interview evaluation with five participants revealed strong alignment, with all participants agreeing that the LLM-generated analysis closely matched their perceptions based on the “context sphere”. Although minor disagreements regarding individual words were noted, there was no consistency across participants in these discrepancies. These findings suggest that the pipeline successfully analyzes user emotions using the Theory of Constructed Emotion without relying on distinct labels. This work bridges advancements in cognitive science and computer science to propose a novel, context-aware approach to emotion analysis. Challenges in evaluating such analyses without a definitive ground truth are discussed, and future directions include refining the “context sphere,” enhancing LLM-guided methodologies for emotion analysis, and integrating different user studies. This research lays the groundwork for developing more nuanced, context-sensitive emotion analysis systems, opening new avenues for human-centered, ethical, and effective applications in technology.