A cognitive-affective chain-driven framework for emotion understanding

Abstract

When users seek help on online health platforms, their expressions are diverse, unstructured, and cognitively complex, posing significant challenges for fine-grained users’ emotion understanding. Existing approaches typically rely on commonsense associations to statically model relationships between emotions and events, overlooking dynamic cognitive processes underlying these connections. To address this shortcoming, we proposed the Cognitive–Affective Chain framework, grounded in Social Support Theory, the Theory of Mind, and the James–Lange Theory of Emotion, to analyze users’ expression from a cognitive perspective. Based on this, we defined a novel task, Cognitive-aware Contextual Emotion Understanding (CCEU), which adapts Aspect-Based Sentiment Analysis to better capture the multidimensional and cognition-driven emotional content. To ensure fair and meaningful evaluation of large language models (LLMs) in cognitively demanding tasks, we introduced the Hybrid Generation and Classification Score (HGCS), a metric combining generation quality and classification reliability. Experimental results showed that LLMs can outperform baselines on HGCS by 15.56 %, even when F1 score drops by 2.12 %, demonstrating that HGCS can better reflect the capabilities of generative models in complex emotion understanding. Next, inspired by Dual Process Theory, we designed prompting strategies that simulate human-like reasoning, improving LLMs’ performance in CCEU task. However, behavioral analysis revealed a bias toward information support over emotional support, exposing the gap between machine inference and human empathy. Taking depression as an example, this study established a cognitively grounded paradigm for emotion modeling in mental health support, also contributing to the development of fair, socially responsive, and cognitively aligned AI systems.