GPT-3 reveals selective insensitivity to globalvs.local linguistic context in speech produced by treatment-naïve patients with positive thought disorder

Abstract

AbstractBackgroundEarly psychopathologists proposed that certain features of positive thought disorder, the disorganized language output produced by some people with schizophrenia, suggest an insensitivity to global, relative to local, discourse context. This idea has received support from carefully controlled psycholinguistic studies in language comprehension. In language production, researchers have so far remained reliant on subjective qualitative rating scales to assess and understand speech disorganization. Now, however, recent advances in large language models mean that it is possible to quantify sensitivity to global and local context objectively by probing lexical probability (the predictability of a word given its preceding context) during natural language production.MethodsFor each word in speech produced by 60 first-episode psychosis patients and 35 healthy, demographically-matched controls, we extracted lexical probabilities from GPT-3 based on contexts that ranged from very local— a single preceding word: P(Wn | Wn-1)—to global— up to 50 preceding words: P(Wn|Wn-50, Wn-49, …, Wn-1).ResultsWe show, for the first time, that disorganized speech is characterized by disproportionate insensitivity to global, versus local, linguistic context. Critically, this global-versus-local insensitivity selectively predicted clinical ratings of positive thought disorder, above and beyond overall symptom severity. There was no evidence of a relationship with negative thought disorder (impoverishment).ConclusionsWe provide an automated, interpretable measure that can potentially be used to quantify speech disorganization in schizophrenia. Our findings directly link the clinical phenomenology of thought disorder to neurocognitive constructs that are grounded in psycholinguistic theory and neurobiology.