Language-Based AI Modeling of Personality Traits and Pathology from Life Narrative Interviews

Advances in artificial intelligence (AI) hold promise for improving personality disorder (PD) models, research methodology, understanding, and clinical treatment. This study models personality and personality pathology using natural language. A representative community sample of N = 1,409 older adults from St. Louis (33% Black, 65% white, 2% other) completed life narrative interviews lasting on average 20 minutes. Language from the interviews was then used to train and test language-based personality models on scores from the NEO-Personality Inventory-Revised (NEO-PI-R) and the Structured Interview for DSM-IV Personality (SIDP-IV). Criteria measures were used for multi-method construct validation of the language models including self-report measures of physical functioning and depressive symptoms and informant-report measures of personality, general health status, and social functioning. Language models were developed using fine-tuning of the parameters of the RoBERTa language model, BERTopic topic modeling, and Linguistic Inquiry and Word Count. Fine-tuned RoBERTa models predicted personality scores in testing data above r = .40, approaching what is considered a large effect size for convergent validity tests between two self-reports of the same construct. Life narrative language was more readily mapped onto the five-factor model (FFM) trait domains than onto DSM personality disorder categories, aside from moderate support for borderline pathology. The language-based FFM scores were supported by multi-method criteria correlations including informant-report personality scores in the testing data. Findings demonstrate the potential promise of language-based AI to refine conceptual frameworks of PD and provide automatic personality assessment and prediction in research and clinical practice.