Lexical versus referential meaning in spontaneous speech in psychosis: computational explorations of the semantic space

Disorders of thought involve alterations in linguistic meaning, which can now be approximated using computational models. This thesis targets two dimensions of meaning separately, lexical-conceptual and referential meaning, both of which are affected in schizophrenia spectrum disorder (SSD). The overall aim is to supplement recent computational methods with a more hypothesis-driven approach, and to use this to improve generalizability of findings and deepen our theoretical and methodological understanding in this field. The first study operationalizes referential recurrences in speech as network graphs, with entities manually identified by noun phrases as nodes. This allows referentiality to be studied as a topological problem and reveals distinct distributional patterns in a chronic SSD sample as compared to healthy controls. The second study, conducted within a transdiagnostic sample comprising SDD and major depression, expands the scope of current semantic similarity measures into a dynamic direction suggesting alterations in the geometry of the semantic space. The third study pursues the theme of a ‘shrinking’ semantic space in a new direction, showing through a principal component analysis (PCA) of high-dimensional word embeddings that the semantic space is more reducible in psychosis. In the final study, different measures of semantic similarity are combined to calculate a single overall composite index of how words are associated in speech, which is then demonstrated to track group differences and diagnostic changes over time. Together, these studies advance our understanding of linguistic meaning in psychosis, highlighting both the potential and limitations of analyses based on current computational language models.