Integrating language model embeddings and topic modeling reveals psychosis-associated disruptions to semantic foraging in narrative recollections

Thought disorder as observed through language disturbance is a hallmark symptom ofpsychosis. Disruptions of verbal fluency task behavior (e.g., listing animals) among individualsexperiencing psychosis have been explained as alterations to semantic foraging processes.Semantic foraging views semantic information retrieval as consisting of bouts of exploitationwithin clusters of related information punctuated by strategic exploratory jumps betweenclusters. However, the semantic foraging framework has not yet been applied to naturalisticspeech. Here, we examine whether disruptions to semantic foraging operating at the level ofspoken narrative can explain the disruptions to natural language that are coupled to psychoticsymptom severity (e.g., tangential speech, poverty of content). To explore this question, wecombined the language model BERT with Discourse Atom Topic Modelling to analyze how 93participants with schizophrenia-spectrum disorders forage for topics as they narrate their lifestories. Negative symptom severity was associated with distinct narrative content, less topicalfocus, and smaller cluster sizes. Positive symptom severity was associated with greatersemantic distances between successive clusters. Cognitive and disorganized symptom severitywas associated with distinct narrative content, larger cluster sizes, fewer topics retrieved percluster, and shorter between-cluster semantic distances. These results were largely robust tomodeling choices and transdiagnostic effects of subject demographics, general indices ofpsychopathology, and inferred substance use history. By using language models within asemantic foraging framework to analyze narratives, this study extends observations from verbalfluency tasks into naturalistic speech and offers mechanistic insights into psychosis-relatedspeech disturbances.