Lexical meaning is lower dimensional in psychosis

Area of Science:

• Computational linguistics
• Psycholinguistics
• Neuroscience

Background:

• Language models (LMs), including large language models (LLMs), analyze semantic organization in speech.
• Individuals with psychosis show heightened semantic similarity, suggesting a compressed semantic space.
• This compressed space may indicate altered underlying geometric properties of semantic associations.

Purpose of the Study:

• To investigate if the dimensionality of semantic vector spaces derived from speech is more reducible in psychosis.
• To explore the geometric underpinnings of semantic organization in speech related to mental disorders.

Main Methods:

• Utilized principal component analysis (PCA) on speech embeddings from LMs to assess semantic space reducibility.
• Calculated metrics like components for 90% variance and variance explained by the first two components.
• Estimated intrinsic dimensionality (ID) for further analysis.

Main Results:

• Confirmed significantly higher reducibility of the semantic space in psychosis across datasets in three languages.
• Demonstrated that the semantic space in psychosis requires fewer dimensions to capture the same amount of variance.
• Results were consistent across different datasets and languages.

Conclusions:

• The findings support the hypothesis of higher semantic space reducibility in psychosis.
• Suggests an underlying intrinsic geometry of semantic associations in speech that is altered in psychosis.
• Offers a novel geometric approach to understanding speech patterns in mental disorders.