[5K1-OS-48] OS-48

9:00 AM - 9:15 AM JST(12:00 AM - 12:15 AM UTC)

[5K1-OS-48-01] Interpretable estimation of diverse emotions from diary textsLeveraging sentence embeddings and independent component analysis

〇Masataka Nakayama¹ (1. Kyoto University)

Keywords:

emotion,embedding,independent component analysis,natural language processing

Understanding the nuances of diverse daily emotions is essential for interpretable human-AI symbiosis. This study developed a method to estimate emotions from diary entries with explainable logic. We conducted an online survey with 300 participants who recorded daily diaries and selected all applicable emotions from 26 categories for 30 days. Diary texts were converted into sentence embeddings and processed using Independent Component Analysis (ICA). These components served as features for L1-regularized logistic regression models. The models achieved a mean AUC of 0.747 via 10-fold cross-validation. Furthermore, analysis of the six most frequent positive emotions revealed that the predictive independent components represented distinct and interpretable emotional profiles. Our findings demonstrate a method for AI to recognize complex emotional states while maintaining human-interpretable reasoning.