Review Paper of Multi-Task DistilBERT for Joint Sentiment and Depression Severity Detection

Depression and related mental health conditions impose a profound global burden, affecting hundreds of millions of people and straining healthcare systems worldwide. Automated text-based screening tools capable of inferring both emotional tone and clinical severity from written language offer a promising complement to traditional assessment pathways. This review paper presents and critically evaluates a multi-task learning framework built upon DistilBERT for the simultaneous prediction of three-class sentiment (negative, neutral, positive) and four-class depression severity (minimum, mild, moderate, severe) from mental health text. The system integrates a cross-task gating mechanism that routes sentiment context into the severity branch, CORAL-based ordinal regression with learnable per-boundary thresholds, and Kendall homoscedastic uncertainty weighting augmented by an ordinal smoothing penalty. A heterogeneous corpus is assembled from six publicly available datasets spanning social media posts, counseling transcripts, and suicide-risk forums. Comprehensive evaluation against Logistic Regression, Support Vector Machine, and single-task DistilBERT baselines demonstrates consistent superiority across accuracy, macro F1, mean absolute error, and quadratic weighted kappa (QWK). Statistical significance is verified through McNemar, Wilcoxon, and bootstrap testing. The review addresses architectural rationale, data engineering choices, evaluation methodology, limitations, and clinical implications.