Computational Linguistic Alignment in Psychosis from Naturalistic Clinical Interviews

This study demonstrates that computational analysis of naturalistic clinical interviews reveals a dyadic failure of semantic coordination in psychosis, characterized by increased lexical-semantic divergence and reduced synchrony between patients and clinicians, which not only persists over time but also measurably disrupts the clinician's own conversational coherence.

The Gist

The Big Idea: The "Broken Dance" of Conversation

Imagine a conversation as a dance between two people. Usually, when two people talk, they subconsciously match each other's rhythm, vocabulary, and energy. They step in time, use similar words, and stay on the same topic. This is called alignment.

This study looked at what happens when a person with psychosis (like schizophrenia) tries to dance with a clinician. The researchers wanted to know: Does the dance break down? And if it does, does it affect the clinician too, or just the patient?

To find out, they didn't just listen to the patient; they analyzed the entire "dance" between the patient and the interviewer using computer tools.

How They Measured the "Dance"

The researchers recorded interviews and used a computer pipeline to break the conversation down into three specific "moves":

1. The Vocabulary Mismatch (Register Divergence):

   • The Analogy: Imagine one dancer is wearing a tuxedo and speaking in high-level poetry, while the other is wearing a t-shirt and speaking in very simple, concrete words about physical objects. They are trying to dance, but they are using completely different "languages" of movement.
   • The Finding: Patients used words that were much more concrete and less varied than the interviewer's. The "distance" between their vocabularies was much larger than in healthy conversations.

2. The Rhythm Break (Synchrony):

   • The Analogy: In a good dance, when one person spins, the other knows exactly when to step in. They move together. In this study, the researchers measured how well the meaning of the patient's sentences matched the flow of the interviewer's sentences.
   • The Finding: The "rhythm" was off. The patient and interviewer weren't moving through the conversation's meaning in sync. When the patient's thoughts got disorganized, this rhythmic connection broke down even more.

3. The Dancer's Own Balance (Within-Speaker Coherence):

   • The Analogy: This measures how steady a single dancer is on their own. Are they stumbling over their own steps, or do they stay on track?
   • The Finding: The patients themselves were actually quite steady on their own steps (their own sentences made sense to them). However, the interviewer started to stumble when talking to the patient. The interviewer's own speech became less consistent and coherent. The "dance" was so disrupted that it threw off the professional dancer (the clinician) too.

What the Computer Found

The study compared 109 patients with 60 healthy people. Here is what the data showed:

• The Gap is Real: Patients and interviewers were speaking from different "worlds" (different vocabularies) much more than healthy people and interviewers were.
• The Rhythm is Lost: The back-and-forth flow of meaning was weaker in patient interviews.
• It's Not Just the Patient: The most surprising finding was that the interviewer became less coherent when talking to a patient. It's as if the patient's difficulty in connecting "infects" the conversation, making the clinician's own speech less smooth.
• It Stays the Same: These differences didn't go away after 12 months. It seems to be a stable part of how these conversations happen, not just a temporary symptom of a bad day.
• The "First Impression" Effect: The mismatch was biggest at the very start of the conversation (when there are no rules or pictures to guide them). As the interview got more structured (like looking at pictures), the gap got a little smaller, but the problem remained.

Connecting the Dots to Symptoms

The researchers found a specific link between the type of "dance break" and the type of symptom:

• "Poverty of Speech" (Not saying much): When a patient had very little to say or used very simple words, the vocabulary mismatch (Register Divergence) was the biggest problem.
• "Disorganized Thinking" (Jumping topics): When a patient's thoughts were jumbled, the rhythmic break (Synchrony) was the biggest problem.

The Bottom Line

The paper suggests that the "Praecox Feeling"—that gut instinct clinicians have that something feels "off" with a patient—isn't just in their head. It is a measurable breakdown in the dyadic (two-person) connection.

The study shows that psychosis isn't just a problem inside the patient's brain; it is a problem in the space between the patient and the listener. The conversation itself becomes unstable, causing the interviewer to lose their own footing. By measuring these digital "dance steps," we can quantify how well two people are connecting, even without needing special brain scans or expensive equipment.

Technical Summary

Technical Summary: Computational Linguistic Alignment in Psychosis from Naturalistic Clinical Interviews

Problem Statement

Clinical interactions with individuals experiencing psychosis are often characterized by a subjective sense of disconnection or "feeling off" (historically termed the Praecox Feeling) prior to formal assessment completion. While prior research has quantified formal thought disorder (FTD) and language anomalies in patient speech in isolation, these approaches overlook the dyadic nature of clinical interviews. Existing literature largely treats patient transcripts as independent readouts, failing to measure how communicative disruption emerges through the dynamic interplay between the interviewer and the patient. Specifically, there is a lack of quantitative, time-resolved measures of how patients and clinicians co-construct meaning, track each other's semantic registers, and maintain conversational continuity in naturalistic settings.

Methodology

Participants and Data

The study analyzed naturalistic clinical interviews from 109 patients with psychotic disorders (schizophrenia spectrum and other psychoses) and 60 healthy controls (HC). Data were collected at baseline and a 12-month follow-up (n=115 for change analyses). Interviews were conducted using the DISCOURSE semi-structured protocol, comprising seven tasks with varying cognitive and communicative demands (e.g., unstructured conversation, personal narrative, picture description, story retelling). Audio recordings were transcribed and verified, yielding 273 transcripts.

Computational Alignment Pipeline

The authors applied a custom Python pipeline to decompose conversational alignment into three distinct levels, utilizing two parallel feature systems: psycholinguistic norm composites and sentence embeddings (paraphrase-multilingual-MiniLM-L12-v2).

1. Register Divergence ($dAUC_{norm}$):

   • Mechanism: Measures the turn-by-turn lexical distance between the interviewer and participant.
   • Features: Content words were scored on five psycholinguistic dimensions: concreteness, semantic diversity, semantic neighbourhood density, visual perceptual strength, and auditory perceptual strength.
   • Computation: The normalized trapezoidal area under the curve (AUC) of the absolute difference between the interviewer's and participant's norm composites across paired turns. Higher values indicate greater divergence in semantic register.

2. Within-Speaker Coherence:

   • Mechanism: Indexes the internal semantic consistency of a speaker's own utterance sequence.
   • Features: Sentence embeddings of full turns (merged consecutive utterances).
   • Computation: Mean cosine similarity between consecutive turn embeddings for each speaker separately. Higher values indicate greater narrative continuity and referential tracking within a speaker's contributions.

3. Embedding-Based Synchrony ($r_{embed}$):

   • Mechanism: Captures the turn-by-turn co-variation in meaning trajectories between speakers.
   • Features: Sequences of sentence embeddings for both interviewer and participant.
   • Computation: Pearson correlation between the two speakers' within-speaker continuity streams (the series of cosine similarities between their own consecutive turns). Higher values indicate stronger rhythmic coupling of semantic trajectories.

Statistical Analysis

Group comparisons and clinical associations were tested using Linear Mixed-Effects Models (LMER) with participant ID as a random intercept to account for repeated measures (baseline and 12-month). Analyses were adjusted for timepoint and corrected for multiple comparisons using the Benjamini-Hochberg False Discovery Rate (FDR).

Key Results

1. Group Differences in Dyadic Alignment

• Register Divergence: Patients exhibited significantly greater register divergence from interviewers compared to healthy controls ($d = 0.48, p < .001$). This effect was most pronounced during unstructured, early interview segments and was driven primarily by concreteness (patients used more concrete, sensory-grounded vocabulary, diverging from the interviewer's varied register).
• Synchrony: Patients showed significantly reduced embedding-based synchrony ($d = -0.59, p < .001$), indicating a failure to couple semantic trajectories with the interviewer in real-time.
• Interviewer Coherence: Crucially, the interviewer's within-speaker coherence was significantly reduced when conversing with patients ($d = -0.33, p = .016$) compared to healthy controls. This suggests that the communicative disruption extends beyond the patient to destabilize the interviewer's own semantic continuity.
• Patient Coherence: Participant-level within-speaker coherence did not significantly differ between groups, suggesting that patients may maintain local narrative continuity even while failing to align dyadically.

2. Clinical Correlations (Double Dissociation)

Two associations survived FDR correction, revealing a double dissociation between alignment metrics and clinical symptoms:

• Register Divergence was positively associated with TLI Impoverishment (poverty of speech), indicating that restricted, concrete vocabulary drives lexical distance.
• Embedding Synchrony was negatively associated with TLI Disorganization (looseness of association), indicating that disorganized thinking disrupts the temporal coordination of meaning.
• These relationships were stable across the 12-month interval, suggesting a partially stable trait-like profile rather than a purely state-dependent marker.

3. Stability and Specificity

• Group differences persisted at the 12-month follow-up with no significant Group $\times$ Timepoint interactions.
• Patients with schizophrenia showed significantly greater register divergence than those with other psychotic diagnoses (e.g., schizoaffective disorder), suggesting the metric is sensitive to specific diagnostic subgroups.
• Sensitivity analyses confirmed that findings were not driven by differences in turn length or structure.

Significance and Claims

The paper claims to provide the first quantitative, dyadic decomposition of conversational alignment in psychosis using naturalistic clinical interviews. Its primary contributions are:

1. Objective Measurement of the "Praecox Feeling": The study offers a computational index for the phenomenological disturbance described by clinicians, demonstrating that the "feeling off" may be measurable as a reduction in the interviewer's own semantic coherence and a failure of semantic synchrony.
2. Dyadic Failure: It challenges the view of thought disorder as a purely intrapersonal deficit, showing that psychosis involves a dyadic failure of semantic coordination that affects the entire interaction system.
3. Scalable Biomarkers: The derived metrics ($dAUC_{norm}$ and $r_{embed}$) offer a scalable, low-burden approach to quantifying interpersonal communicative function from routine clinical transcripts without specialized equipment.
4. Distinct Mechanisms: The findings support the existence of distinct mechanisms for different thought disorder symptoms: impoverishment manifests as register divergence (lexical restriction), while disorganization manifests as reduced synchrony (temporal misalignment).

The authors conclude that these transcript-derived metrics bridge the gap between phenomenological observation and computational analysis, providing a foundation for future studies on interpersonal communication in psychosis, though they note limitations regarding generalizability to chronic populations and other languages.