Perplexity and pseudo-perplexity as objective measures of linguistic deficits in schizophrenia: Evidence from a Korean-Speaking cohort

07/05/2026

AI Summary

Patients with schizophrenia showed significantly elevated perplexity (PPL) and pseudo-perplexity (PPPL) across tasks and model architectures compared with healthy controls.
Both metrics correlated with global cognitive impairment (CGI-SCH Cognition) and impaired abstract thinking (PANSS N5), linking linguistic deviation to symptom severity.
Elevated PPL and PPPL reflect heightened contextual uncertainty, supporting language models as automated, scalable tools to quantify linguistic pathology.

Schizophr Res. 2026 May 5;294:147-154. doi: 10.1016/j.schres.2026.04.030. Online ahead of print.

ABSTRACT

BACKGROUND: Linguistic deficits are core features of schizophrenia. Perplexity (PPL) and pseudo-perplexity (PPPL), derived from language models, serve as probabilistic metrics of contextual uncertainty. This study investigated their utility as objective markers for distinguishing patients with schizophrenia from healthy controls.

METHODS: Speech samples were collected from 249 Korean-speaking patients with schizophrenia and 159 healthy controls across eight tasks, including free, emotional, and projective narratives. PPL and PPPL were calculated using autoregressive and bidirectional models, respectively. Group differences were assessed via Rank ANCOVA, adjusting for age, sex, education, and token count. Additionally, exploratory analyses examined correlations with clinical symptom severity.

RESULTS: Patients exhibited significantly elevated PPL and PPPL values compared to controls across most tasks, a finding consistent across different model architectures. Exploratory correlation analyses indicated that both metrics were associated with global cognitive impairment (CGI-SCH Cognition) and difficulty in abstract thinking (PANSS N5), linking linguistic deviations to psychopathological severity.

CONCLUSIONS: Elevated PPL and PPPL effectively characterize linguistic deviations in schizophrenia, reflecting heightened contextual uncertainty in language modeling associated with the underlying psychopathology. These findings support the utility of language models as automated, scalable tools for objectively quantifying linguistic pathology and elucidating the neurocognitive substrates of schizophrenia.

PMID:42092305 | DOI:10.1016/j.schres.2026.04.030

Document this CPD