Schizophrenia Speech Transcripts with Surprisal Metrics

A 2026 study by Seunghyong Ryu provides transcripts from 249 patients with schizophrenia and 159 healthy controls across eight speech tasks. The dataset includes token-by-token surprisal metrics derived from Korean language models, analyzing the dynamic breakdown of contextual coherence. The data is hosted on figshare under a CC-BY-4.0 license.

Use Cases

• Modeling the temporal trajectory of speech disorganization based on sequential surprisal metrics
• Comparing contextual coherence between patient and control groups based on surprisal divergence
• Investigating the impact of task structure on language anomalies based on narrative vs. projective task data
• Validating language model-derived metrics against clinical descriptions of disorganized discourse

Strengths

• Includes data from 408 total participants (249 patients, 159 controls), providing a substantial clinical sample
• Analyzes eight diverse speech tasks, allowing for comparison across different contexts
• Employs two specific Korean language models (Polyglot-ko-5.8b and Kanana-1.5-8b-base) for metric generation

Limitations

• Column-level documentation is absent; field semantics must be inferred after download
• Row count is unknown, which may limit suitability assessment
• The dataset is very small (183.3 KB), indicating limited scope, likely containing summary metrics rather than full transcripts

Provenance

Source

Seunghyong Ryu

Collection Method

Analysis of speech transcripts using autoregressive language models to calculate token-by-token surprisal.

Freshness

Last updated 2026-06-01 05:23:18; freshness should be verified

Geography

Likely South Korea, based on the use of Korean language models and tasks.