The Divergence Hypothesis: Unmasking Lexical Interference and Label Bias in Mental Health NLP

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Computational Mental Health · Depth: Expert, quick

Summary

Moustafa Hassan introduces the Triple-Stream Stress probe (TSS), a multi-channel diagnostic framework designed to unmask lexical interference and label bias in computational mental health (CMH) classifiers. TSS decomposes text into lexical character n-grams, a morpho-syntactic channel, and a 154-feature psycholinguistic style channel. Across four English datasets totaling 12,906 instances, TSS revealed a lexical interference effect where adding lexical features to the style channel reduced Macro-F1 by a mean of 0.072 (p < 10??) on human-labeled data, but not on auto-labeled data. The study proposes Degree of Divergence (DoD), an econometrics-adapted statistic for label-source auditing, with a headline estimate of DoD(BC?A) = 0.0374 (95% CI [0.0097, 0.0651], p = 0.0032). Interventional masking further showed Channel C's performance is ~95?99% retained after content word destruction, indicating its independence from lexical surface form. TSS functions as a diagnostic audit tool to flag shortcut learning.

Key takeaway

For Machine Learning Engineers developing computational mental health (CMH) classifiers, you should integrate diagnostic frameworks like TSS into your model development pipeline. This helps identify and mitigate label-source-specific shortcut learning and lexical interference effects before deployment. By auditing your datasets with Degree of Divergence (DoD), you can ensure your models generalize robustly to human-labeled data, preventing performance degradation under distribution shift.

Key insights

CMH classifiers suffer from label bias and lexical interference, which TSS can diagnose.

Principles

• Human and auto-labeling reward different linguistic signals.
• Psycholinguistic style features are largely independent of lexical content.

Method

TSS decomposes text into lexical n-grams, morpho-syntax, and psycholinguistic style features to identify label-source-specific shortcut learning. DoD quantifies divergence.

In practice

• Use TSS to audit label sources for bias.
• Evaluate CMH models for lexical interference.

Topics

• Computational Mental Health
• Natural Language Processing
• Label Bias
• Lexical Interference
• Diagnostic Frameworks
• Distribution Shift

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Multi-View Attention Multiple-Instance Learning Enhanced by LLM Reasoning for Cognitive Distortion Detection — Combining LLMs with MIL and ELB components improves cognitive distortion detection by enhancing interpretability and reasoning.
• MentalMARBERT: Domain-Adaptive Pre-training and Two-Stage Fine-Tuning for Arabic Mental Health Disorders Detection — Domain-adaptive pre-training and hierarchical classification significantly improve Arabic mental health disorder detection.
• Fully Automated Identification of Lexical Alignment and Preference-Stage Shifts in Large Language Models — Automated metrics can identify and quantify lexical overuse and preference learning shifts in LLMs without manual curation.
• Team MKC at CLPsych 2026: Capturing and Characterizing Mental Health Changes through Social Media Timeline Dynamics — An LLM-based pipeline analyzes social media timelines for early mental health change detection and continuous monitoring, integrating post and user-level assessments.
• Probing Semantic Alignment, Lexical Invariance, and Syntactic Influence in LLM Metaphor Processing — LLMs struggle with deep metaphor comprehension, often misinterpreting concepts and relying on lexical associations over context or full syntactic understanding.
• Mitigating Data Scarcity in Psychological Defense Classification with Context-Aware Synthetic Augmentation — Context-aware synthetic data augmentation with theory-driven prompts significantly improves psychological defense mechanism classification in low-resource settings.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.