The Good, the Bad, and the Language: Capturing Implicit Social Evaluation of Personality Traits via the Fill‑Mask Association Test

OBJECTIVE: This study investigated whether linguistic evaluation of personality traits-the association of trait words with good versus bad in natural language-constitutes a diffuse halo or is selectively amplified for dimensions central to social functioning. METHOD: Using the Fill-Mask Association Test (FMAT), the present study extracted log probability ratios (LPR) of good vs. bad from 12 BERT models across three sentence contexts for 733 personality adjectives. These LPR scores were compared with human normative ratings (valence, social desirability, and observability) and with Big Five structure derived from three independent lexical taxonomies. RESULTS: Language-based LPR scores correlated strongly with human valence (r = 0.926) and social desirability (r = 0.931), but weakly with observability (r = 0.235). Across three independent lexical taxonomies, Agreeableness and Conscientiousness exhibited larger evaluative gaps than Extraversion, which showed the smallest gap. CONCLUSIONS: Linguistic evaluation is not a uniform halo but is functionally calibrated to traits relevant to social cooperation and moral obligation. The FMAT provides a scalable, reproducible tool for extracting implicit evaluative maps from language.