PUBLISHED · NATURE MENTAL HEALTH · 2026
Natural language processing of youth speech predicts psychopathology across adolescence
Chase Antonacci, Jessica P. Uy, Kaitlyn Kwan, Eugenia Giampetruzzi, Sabrina Jones, James W. Pennebaker, Ian H. Gotlib
Finding:
Stress interviews with 204 youths, mean age 11.38 and range 9 to 13, 58% female, were analysed with a suite of natural language processing techniques to predict internalizing psychopathology up to six years later. The interviews averaged about 30 minutes and covered lifetime exposure to traumatic events, recorded when none of the children had yet been diagnosed with a mental health disorder. Follow-up assessments came at four or six years. Linguistic features explained more than twice the variance of traditional, human-rated risk factors.The comparison condition was demographic information plus expert-produced stress severity scores derived from the same interviews. In many cases the language models outperformed those expert ratings, which are the gold-standard measurement in this context.Style carried more signal than content. The strongest predictors included elevated narrative complexity, marked by heavier use of function words such as prepositions and interpreted as a possible marker of ruminative thinking; rigid language expressing absolute certainty; and self-referential speech indexed by first-person singular pronouns. Semantic models added a thematic layer: narratives about routine engagement with hobbies, structured activities and access to healthcare appeared protective, while narratives about acute physical violence and social exclusion moved risk in the opposite direction.
Relevance:
Co-authored by James W. Pennebaker of the University of Texas at Austin, co-founder and Chief Science Officer at Receptiviti, and the researcher whose work established the psycholinguistic framework Receptiviti's dimensions are built on.
Function words, absolutist certainty, and first-person singular pronouns are the closed-class, low-content categories that speakers do not consciously select and that a reader would not register. They are also precisely the categories psycholinguistic measurement extracts and that content-level or sentiment-level analysis discards as noise. That these outperformed expert clinical judgment of the same interviews is a direct demonstration that the psychological signal in language sits substantially in style rather than in what is said.