Lexical meaning is lower-dimensional in psychosis: the intrinsic geometry of the semantic space

Diverse language models (LMs), including large language models (LLMs) based on deep neural networks have come to provide an unprecedented opportunity for mapping out the semantic spaces navigated in speech and their distortions in mental disorders. Recent evidence has pointed to higher mean semantic similarities between words in psychosis, conceptualized as a ‘shrunk’ (more compressed) semantic space. We hypothesized that the high dimensionality of the vector spaces defined by the embeddings of speech samples through LMs would also be easier to reduce in psychosis. To test this, we used principal component analysis (PCA) to calculate different metrics serving as proxies for reducibility, including the number of components needed to reach 90% of variance, and the cumulative variance explained by the first two components. For further exploration, intrinsic dimensionality (ID) was also estimated. Results confirmed significantly higher reducibility of the semantic space in psychosis across all measures and three languages. This result points to the existence of an underlying intrinsic geometry of semantic associations during speech, which may underlie more surface-level measurements such as semantic similarity and illustrates a new foundational approach to speech in mental disorders.