Examining coreference Resolution in schizophrenia with a language model: detection gap, model performance and clinical correlations

Abstract

The occurrence of referential anomalies in spontaneous speech in psychosis, and more specifically in patients with formal thought disorder, has been documented since the late 1970s, and recent linguistic studies have shown in several typologically different languages that the quantitative distribution and quality of use of different types of noun phrases differs in people with psychosis, specifically in anaphoric noun phrases that pick up a previously mentioned entity. To identify changes in such spontaneous speech patterns, automated natural language processing (NLP) technologies are now frequently used with the intention that this line of research may eventually help with clinical goals of diagnosis and prognosis. Here we are taking a coreference resolution model applicable to German and test the prediction that it will fare worse when processing coreferential expressions from schizophrenic speech. Comparisons between patient and control groups, as well as correlations with clinical variables, revealed statistically non-significant results, suggesting that the referential coherence ofschizophrenic patients' speech is comparable to that of healthy controls, from the viewpoint of this model. A post-hoc analysis of averaged semantic similarity between consecutive words yielded similarly non-significant group differences. Together, these results suggest semantic structure at both the referential and lexical-conceptual level to be intact in schizophrenia, questioning the generalizability of current evidence to the contrary.