Abstract
This study explores the intersection of Natural Language Processing (NLP) and clinical psychology by measuring "emotional coherence"—the alignment between a client's verbal expressions and their internal feelings. Using Transformer-based models (AlephBERT) on a large-scale Hebrew psychotherapy dataset (BIU-872), the researchers demonstrate that automated emotion recognition can track emotional processes. Key findings indicate that higher coherence in positive emotions significantly correlates with patient well-being, suggesting a new automated diagnostic tool for clinicians to monitor therapeutic progress.
Introduction
In psychotherapy, a primary goal is helping clients align their internal emotional experiences with the words they use to describe them. This "emotional coherence" is considered a hallmark of mental health and effective regulation. Historically, studying this phenomenon has been difficult, relying on small-scale laboratory tasks or requiring human experts to manually code thousands of hours of therapy transcripts—a process that is slow, expensive, and hard to scale.
Furthermore, it wasn't empirically clear if coherence measured during actual therapy sessions (rather than lab tasks) actually correlated with a client's overall well-being. This paper addresses these challenges by determining if NLP can automate this measurement and provide a scalable diagnostic tool for clinicians.
Methodology
The researchers developed an end-to-end pipeline to measure coherence using the following steps:
- Emotion Recognition (ER) Model: They adapted AlephBERT, a Transformer-based model pre-trained on a large Hebrew corpus, for utterance-level emotion classification.
- Classification: The model was trained to categorize client utterances into four labels: Positive, Negative, Neutral, or Mixed.
- Normalization: To account for varying session lengths, they normalized the frequency of predicted emotions per session.
- Coherence Calculation: They calculated the Pearson correlation between the model’s predicted emotions (verbal expression) and the client’s self-reported moods (subjective experience) collected via the Profile of Mood States (POMS) survey after each session.
- Well-being Association: Finally, they correlated these coherence scores with the Outcome Rating Scale (ORS), a standard measure of client well-being.
Data and Experimental Setup
The study utilized the BIU-872 dataset, which consists of 872 psychotherapy sessions conducted in Modern Hebrew. The data was split into two primary subsets:
- BIU-872_Gold: A subset of 196 sessions annotated by clinical experts at the utterance level to train and validate the ER model.
- BIU-872_Silver: The remaining un-annotated sessions used to test the model's application in a real-world, "in-the-wild" scenario.
Psychological metrics, including post-session self-reports (POMS) and pre-session well-being scales (ORS), provided the ground truth for internal states and clinical outcomes.
Results and Analysis
The AlephBERT model achieved a moderate F1 micro score of 0.66 for emotion recognition in the complex domain of psychotherapy dialogue. The primary findings regarding emotional coherence and its clinical implications are detailed in the tables below.
In the "Silver" (unlabeled) dataset, the correlation between predicted and self-reported positive emotions was 0.27, and 0.21 for negative emotions (both statistically significant). Crucially, higher coherence in positive emotions was strongly associated with better client well-being, showing a 0.67 correlation (p < 0.05) with the ORS. Interestingly, coherence in negative emotions did not show a significant positive relationship with well-being in this study.
Discussion and Future Work
The significant correlation between the model's predictions and the clients' private self-reports validates that NLP can capture genuine internal emotional processes from dialogue, not just surface-level word usage. A major insight was the specificity of positive coherence; while coherence was observed for both positive and negative emotions, only positive emotion coherence significantly correlated with well-being. The authors suggest that while negative emotions might be more "salient," the ability to recognize and verbally express positive states is a stronger driver of clinical improvement.
This approach moves psychotherapy research from "snapshots" to "movies," allowing for longitudinal tracking across months of real therapy. Future work could explore more nuanced emotion categories (e.g., shame vs. guilt) and the development of real-time "coherence alerts" to help therapists identify sessions where a client is highly dissociated from their feelings.
Summary updated