Automated detection of referential features in schizophrenic speech using large language models.

Area of Science:

• Psycholinguistics
• Computational Linguistics
• Clinical Neuroscience

Background:

• Cross-linguistic studies reveal distinct noun phrase (NP) distributions in the spontaneous speech of individuals with schizophrenia, particularly those with formal thought disorder (FTD).
• Extracting referential NP features, crucial for understanding meaning organization, traditionally necessitates manual linguistic annotation.

Purpose of the Study:

• To investigate the feasibility of using state-of-the-art large language models (LLMs) for the automatic extraction of referential NP features in spontaneous speech.
• To validate LLM-based findings against manual annotations and assess their utility in identifying linguistic markers associated with schizophrenia and FTD.

Main Methods:

• Application of LLMs to a manually annotated dataset of comic strip descriptions from individuals with schizophrenia (SZ) (with and without FTD) and neurotypical controls (NC).
• Utilized in-context (few-shot) learning with LLMs to enhance feature extraction accuracy.
• Comparative analysis of LLM-derived NP features with results from manual annotation.

Main Results:

• LLM-based analyses successfully replicated findings from manual annotations, confirming underuse of definite NPs in the schizophrenia with FTD (SZ+FTD) group.
• Definite NPs, indicative of discourse-based reference, grammatical complexity, and narrative coherence, were significantly underutilized by the SZ+FTD group.
• LLMs, particularly with few-shot learning, demonstrated effectiveness in automatically extracting these clinically relevant referential features.

Conclusions:

• LLMs provide a promising and automated approach for extracting referential NP features in spontaneous speech.
• This automated assessment using Natural Language Processing (NLP) can identify clinically significant linguistic deviations validated across languages.
• LLM-driven analysis offers a scalable method for assessing linguistic patterns relevant to schizophrenia and FTD.