PROBING THE COGNITIVE APPRAISAL STRUCTURE OF EMOTION REPRESENTATIONS IN LARGE LANGUAGE MODELS (LLMS): A FULLY AUTOMATED GEOMETRIC ANALYSIS BASED ON CHINESE–ENGLISH BILINGUAL CORPORA

Abstract

The emotional capabilities of large language models (LLMs) have been extensively validated, yet the organizational principles and cognitive regularities underlying their internal emotion representations remain poorly understood. This study introduces Smith and Ellsworth’s (1985) cognitive appraisal theory and designs an automated experimental pipeline in which five mainstream LLMs rate Chinese and English emotional texts along six appraisal dimensions. Combining representational similarity analysis, principal component analysis, and unsupervised clustering, we systematically probed the geometric structure of the LLMs’ emotion space. The results show that the appraisal structures of all LLMs are significantly aligned with human templates; however, under the English condition, the alignment did not surpass a purely semantic baseline, whereas under the Chinese condition the LLMs’ appraisals significantly outperform the semantic baseline. The LLMs’ appraisal space exhibits a systematic “responsibility shift,” in which the responsibility dimension is disproportionately amplified on higher-order principal axes. Chinese alignment is significantly higher than English alignment for all models, and clustering structures were driven more by linguistic properties than by model architecture differences. These findings reveal that LLMs form emotion-cognitive structures that exhibit both commonalities and language-specific particularities, providing a psychology-theory-driven quantitative framework for explainable artificial intelligence (AI).