AI Tracks Broken Reference Chains in Speech to Map How Meaning Unravels in Schizophrenia and Depression

Language has long been recognized as a window into the mind, but psychiatric research is now moving beyond counting words or measuring vocabulary to interrogate something far more subtle: how speakers keep track of who or what they are talking about as discourse unfolds. A new open-access study published on 30 September 2026 in the journal Schizophrenia reports that an automatic coreference resolution model, a class of natural language processing tool designed to determine when different expressions in a text point to the same entity, performs measurably worse on the speech of people with schizophrenia spectrum disorder and major depressive disorder than on the speech of healthy controls. The finding, from a team led by Claudio Palominos and Wolfram Hinzen at Universitat Pompeu Fabra in Barcelona together with collaborators in Marburg, Groningen, Zurich and the German multicenter FOR2107 consortium, suggests that the referential architecture of meaning, the system that binds pronouns, names and descriptions to their targets, is disrupted in both conditions, but in ways that may reflect partially distinct underlying mechanisms.

Most previous computational studies of psychiatric speech have concentrated on lexical-conceptual meaning, using word embeddings and semantic similarity measures to quantify how ideas relate to one another across a stretch of discourse. That work has produced striking results, including demonstrations that semantic coherence degrades in psychosis in ways that defy simple similarity accounts. Yet meaning has at least two dimensions. Lexical-conceptual meaning concerns what words and concepts denote and how they cluster semantically. Referential meaning, by contrast, concerns how a speaker anchors those concepts to particular individuals and situations, maintaining a stable web of entities across sentences. When someone says that a neighbor complained and then adds that she would not stop, the listener must resolve that she refers back to the neighbor. Failures of this binding process, long noted clinically in formal thought disorder, had never been systematically quantified with modern language models across diagnostic groups until now.

The research team tested a specific hypothesis: that coreference would be harder for the automatic model to resolve in the speech of clinical participants than in controls. They assembled three groups of participants who produced spoken narratives, comprising 36 healthy controls, 41 individuals with schizophrenia spectrum disorder and 42 with major depressive disorder. The coreference resolution model was applied to transcripts of this speech, and its performance, essentially how accurately it could link referring expressions to the correct antecedents, was compared across groups. The results confirmed the prediction. Model performance was significantly lower in both the schizophrenia and depression groups relative to healthy controls, indicating that the referential threads running through clinical discourse are genuinely harder to track, even for a machine system trained on vast amounts of typical language.

What elevates the study beyond a simple group difference is its topological approach. The researchers constructed speech graphs, network representations in which linguistic units become nodes and the relations between them become edges, separately for referential structure and for lexical-semantic structure. Graph theory then allowed them to characterize the topology of each kind of meaning and, crucially, to compare the two. In healthy speech, the referential graph and the semantic graph diverge in characteristic ways: the web of entities a speaker tracks and the web of concepts they deploy are related but structurally distinguishable systems. In the schizophrenia group, the researchers found a reduced structural divergence between these two graphs, suggesting that the normal differentiation between referential and lexical-semantic organization is blunted. Meaning, in effect, loses some of its articulated internal architecture in schizophrenia spectrum disorder.

The correlational analyses sharpened the clinical picture considerably. Within the schizophrenia group, poorer model performance on coreference was negatively associated with positive symptoms, meaning that participants with more severe positive symptoms, the hallucinations, delusions and disorganized thinking that historically define psychosis, showed the greatest referential disruption. This aligns with a long clinical tradition linking formal thought disorder to breakdowns in the cohesion of discourse, but it now provides a quantitative, model-based handle on the phenomenon. Within the depression group, by contrast, coreference performance was associated with verbal memory measures rather than with symptom severity in the positive-symptom domain. This dissociation is theoretically important. It implies that although both disorders show referential difficulty at the group level, the cognitive machinery underlying that difficulty may differ: in schizophrenia it tracks the psychotic process itself, while in depression it may reflect the well-documented memory and executive impairments that accompany the illness.

The technical machinery behind these conclusions deserves attention, because it illustrates how computational linguistics is becoming a precision instrument for psychopathology. Coreference resolution is one of the harder problems in natural language understanding. A model must recognize that a definite description, a pronoun, a name and even an elliptical phrase can all denote the same individual, and it must do so across sentence boundaries where ambiguity abounds. When such a model, calibrated on ordinary language, stumbles more often on clinical speech, the errors are not random noise; they signal that the discourse itself provides weaker or more inconsistent cues about entity identity. Combined with speech graph topology, which quantifies properties such as connectivity and structural divergence between representational layers, the method yields a multidimensional profile of how meaning is organized, or disorganized, in an individual speaker.

The study forms part of a broader research program in which language is treated not merely as a symptom carrier but as a measurable biomarker of brain function. The same Barcelona group and their collaborators have previously shown progressive changes in descriptive discourse in first-episode schizophrenia using computational semantics, and related work has connected reduced linguistic coherence in psychosis to altered large-scale cortical hierarchy. The new findings extend this agenda by profiling two dimensions of meaning against each other rather than in isolation. The authors describe this as linking lexical-conceptual and referential-semantic alterations in a more integrated foundational model of language in major psychopathology. In practical terms, it means future computational assays of speech could in principle distinguish referential breakdown from semantic drift, refining both diagnosis and the tracking of illness course.

The clinical implications are potentially far-reaching. Speech samples are cheap, noninvasive and increasingly easy to collect, and language-based digital biomarkers are already being developed for early detection and monitoring of psychosis risk, including within European projects such as TRUSTING, which supported this work alongside the German Research Foundation. If coreference resolution performance and speech graph topology can be computed reliably from a few minutes of natural speech, they could complement clinical interviews with objective, quantifiable measures. The differential associations reported here, positive symptoms in schizophrenia and verbal memory in depression, hint that such measures might eventually help discriminate between conditions that overlap superficially, or track distinct therapeutic targets within a single patient over time. Much validation remains to be done, including longitudinal studies and replication across languages and clinical settings, but the direction of travel is clear.

The study also carries conceptual weight for theories of language in the mind and brain. Referential capacity, the ability to maintain a coherent cast of entities across discourse, has been proposed by some theorists, including Hinzen, to be a distinctive vulnerability point in schizophrenia, connected to the very architecture of human grammar and thought. The observation that referential and semantic graphs normally diverge structurally but converge in schizophrenia gives that proposal an empirical, topological form. Meanwhile, the depression findings caution against a single unified account of linguistic disruption in psychiatry: referential difficulty is not a psychosis-specific signature when it also appears, with different cognitive correlates, in mood disorder. Disentangling these threads is precisely what the new integrated framework is designed to do.

For now, the study stands as a demonstration that the fine grain of meaning, down to the humble pronoun and its antecedent, can be measured, modeled and related to symptoms and cognition in major mental illness. As language models grow more capable and clinical speech datasets grow larger, the boundary between computational linguistics and psychiatry is dissolving, and with it may come a new generation of tools that read the structure of thought directly from the structure of speech. The participants in the FOR2107 consortium whose narratives made this analysis possible contributed to a result that is as conceptually elegant as it is clinically promising: meaning, in its referential dimension, has a measurable topology, and in psychosis that topology changes.

Subject of Research: Computational analysis of coreference and speech graph topology in schizophrenia and major depressive disorder

Article Title: Co-reference and the topology of meaning in schizophrenia and depression

Article References: Co-reference and the topology of meaning in schizophrenia and depression. (n.d.). https://doi.org/10.1038/s41537-026-00802-6

Image Credits: AI Generated

DOI: 10.1038/s41537-026-00802-6

Keywords: schizophrenia, major depressive disorder, coreference resolution, natural language processing, speech graphs, computational psychopathology, formal thought disorder, positive symptoms, verbal memory, semantic topology, language biomarkers, digital psychiatry