Tracking the dynamic breakdown of contextual coherence in schizophrenia using language models

Area of Science:

• Computational Linguistics
• Psychiatry
• Cognitive Neuroscience

Background:

• Disorganized speech is a key symptom of schizophrenia, linked to impaired contextual integration.
• Objective quantification of speech dynamics is needed to understand schizophrenia's psychopathology.

Purpose of the Study:

• To investigate the dynamic temporal trajectory of contextual coherence breakdown in schizophrenia speech.
• To quantify disruptions in lexical predictability using sequential surprisal from language models.

Main Methods:

• Analysis of speech transcripts from 249 schizophrenia patients and 159 controls across eight tasks.
• Calculation of token-by-token surprisal using Korean autoregressive language models (Polyglot-ko-5.8b, Kanana-1.5-8b-base).
• Generalized additive mixed models used to analyze temporal divergence of surprisal trajectories from token 11 to 100.

Main Results:

• Patients exhibited progressively divergent surprisal trajectories, indicating a breakdown in contextual predictability.
• Deviations intensified between tokens 50-70, showing rapid deterioration in maintaining global context.
• Higher overall mean surprisal and pronounced temporal divergence were observed in unstructured narrative tasks.

Conclusions:

• Language anomalies in schizophrenia are dynamic, unfolding rapidly within short discourses.
• Temporal natural language processing metrics effectively capture contextual disruption, aiding in understanding schizophrenia psychopathology.
• Findings highlight the utility of sequential surprisal in objectively measuring speech disorganization.