Affective Dynamics in Text Carry Complementary but Modest Signal for Big Five Personality Prediction: A Rigorous Evaluation on the Essays Corpus

Abstract

Psychological theory links personality to how emotions unfold over time, yet text-based personality detection usually treats affect as a static summary. We segment each essay of the Essays corpus (n = 2,467) into 64-word windows, score every window with pretrained sentiment and emotion models, and derive 79 trajectory features. Under 3 × 10-fold cross-validation with corrected resampled t-tests, affect dynamics differ significantly between high and low trait groups: valence rises more steeply for high Conscientiousness, Extraversion and Agreeableness, and is lower but less volatile for high Neuroticism. Affect is largely not recoverable from MPNet embeddings (mean cross-validated R² = 0.35 for means and 0.11 for dynamics), yet stacking it with embeddings adds only 0.65 macro-F1 points (p = 0.14). The fused model (58.84% accuracy) matches a fine-tuned RoBERTa (58.91%) without GPU training. Affective dynamics are therefore complementary but modest predictors.