Decoding neural emotion patterns through large language model embeddings

Understanding how emotional expression in language relates to brain function remains a key challenge in neuroscience. Traditional neuroimaging provides valuable insight but is costly and limited to controlled laboratory settings. Here, we present a computational framework that explores potential links between emotional content in natural language and neuro-anatomical regions associated with affective processing. This, when validated through complementary neuroimaging, may enable scalable, imaging-free investigation of emotion-brain relationships. Our approach combines text embeddings, dimensionality reduction, and clustering to identify emotional states, which are then mapped to relevant brain regions. The framework was evaluated across three applications: (i) comparing healthy and depressed individuals, (ii) analyzing a large-scale emotion dataset, and (iii) contrasting human and large language model (LLM) outputs. Emotion intensity was quantified using a lexical scoring system sensitive to keywords, syntax, and modifiers, producing computationally plausible emotion-to-region clusters with visualization mapping. Across experiments, the framework distinguished healthy from depressed participants through distinct computational activation patterns and revealed systematic differences between human and LLM-generated texts in predicted computational engagement. A key finding emerged: depressed individuals exhibited reduced emotional diversity, showing 2.2 - 2.7 times more homogeneous emotional expression than healthy controls, suggesting that emotional rigidity may serve as a computational marker of depression. These computational patterns represent testable hypotheses, requiring further validation through neuroimaging. This work establishes a scalable, cost-effective tool for advancing both clinical and computational models of emotion, and provides a neuro-inspired benchmark for assessing how closely AI-generated language mirrors human emotional expression. • This study introduces a computational framework for mapping language emotion to brain regions without requiring neuroimaging. • The integration of embeddings and neuro-anatomical mapping differentiates through distinct activation patterns. • The framework demonstrates high spatial specificity by mapping discrete emotions to neuro-anatomically plausible regions. • Predicted regional assignment are derived from established neuroimaging coordinates requiring further independent validation. • In favor of reproducible research and to advance the field, all code used in this study is made publicly available.