Emotion recognition in cross-linguistic legal context: a comparison between human-based and computational approaches

Emotion shapes credibility assessments, judicial decision-making, and perceptions of procedural justice, yet its reliable detection in courtroom settings remains a significant methodological challenge. This study provides a multimethod evaluation of emotion-recognition approaches in cross-linguistic legal discourse using a corpus of 20 Chinese and English courtroom video clips. Eighty-eight bilingual observers provided continuous valence ratings, completed post-task Positive and Negative Affect Schedule (PANAS) assessments, and contributed electrodermal activity (EDA) and heart rate variability (HRV) recordings. Observer-perceived emotional-expression ratings were benchmarked against three analytic streams—computational, psychometric and physiological—yielding three preliminary findings. First, computational sentiment tools varied widely: cloud-based NLP services and general-purpose large language models exhibited stronger rank-order alignment with observer ratings, whereas lexicon-based and language-specific tools performed inconsistently. Second, the psychometric measure PANAS, evaluated with logistic regression, showed limited discrimination capacity. Third, among deep-learning models trained on physiological signals using Long Short-Term Memory (LSTM) architectures, the EDA-HRV fusion model yielded the highest accuracy (0.773) among several comparably performing configurations, suggesting that autonomic measures may capture affective states in courtroom interactions through the emotional contagion mechanism. Cross-linguistic analyses indicated generally stronger tool–human correspondence for Chinese clips, highlighting how linguistic and cultural factors may shape emotion detectability in legal contexts. These findings carry both methodological and practical implications for the development of specialized emotion-recognition tools in forensic and legal settings.