MH-EABERT: an adaptive fusion approach with emotion knowledge-enhanced BERT for explainable mental health sentiment analysis from social media

While deep learning models achieve high accuracy in mental health screening, their inherent “black box” nature precludes clinical trust, and pure rule-based systems fail to dynamically grasp complex linguistic nuances such as sarcasm. To rigorously bridge this critical research gap, this study introduces MH-EABERT (Mental Health Emotion-Augmented BERT), a novel neuro-symbolic framework that seamlessly integrates domain-specific clinical knowledge with deep semantic learning to heavily enhance both classification accuracy and explicit explainability. We construct a Mental Health Emotion Lexicon (MHEL) that maps 13,872 emotion terms and 181 clinical phrases to six cognitive dimensions. The architecture operates across five distinct computational modules, utilizing an Emotion Knowledge Extraction (EKE) scorer to apply 10 clinical logic rules that distinguish pathological states from general distress. To prevent semantic distortion, a Self-Adaptive Fusion (SAF) algorithm dynamically governs knowledge injection based on continuous confidence calibration. Using mathematical “hard positioning,” we embed clinical context directly into the sentence structure, forcing BERT’s attention mechanism to process it as integral semantics. Comprehensive evaluation utilizing 5-fold cross-validation repeated across 5 random seeds on three diverse datasets—Reddit Communities (151,288 posts), Comprehensive MH (53,043 statements), and Synthetic Posts (1000 samples)—demonstrated statistically significant superiority ( \(p < 0.0001\) ) over state-of-the-art baselines including RoBERTa, XLNet, DistilBERT, MentalBERT, and ClinicalBERT. Crucially, all evaluated models were rigorously trained and tested under the exact same experimental setup. MH-EABERT achieved a peak accuracy of 87.97% on Reddit data and crucially reduced the False Negative Rate for suicidal detection from 14.67 to 8.55%. Extensive ablation studies and SHAP analyses confirmed that predictions are grounded in transparent, auditable clinical features. MH-EABERT represents a significant advancement toward clinically grounded, highly interpretable AI systems capable of reliable, automated triage from unstructured social media text. To ensure absolute reproducibility, the code, lexicon, and framework weights are publicly available at https://doi.org/10.5281/zenodo.23285908 .