Multi-Class Mental Health Discourse Classification from Online Texts: Classical ML vs Transformer/LLM Modeling
TL;DR
It is shown that con-textual transformer/LLM models yield more reliable macro-level performance under imbalance than TF–IDF baselines, particularly for semantically adja- cent classes.
Abstract
Online mental-health discourse contains diverse in-tents including awareness education, wellness advice, support seeking, and symptom narratives associated with anx- iety and depression. These categories are semantically overlapping, short and noisy, and imbalanced in public datasets, making robust multi-class classification difficult. This paper presents an end-to-end framework for six- class mental health discourse classification and provides a publication-oriented SN Computer Science style analysis including mathematical formulation, gradient deriva-tions for cross-entropy learning, computational complexity anal-ysis, ablation design, and statistically grounded compari- son using McNemar’s test. Using a public dataset with six classes (awareness, wellness, anxiety, support, depression, general) and the provided empirical results and figures, we show that con-textual transformer/LLM models yield more reliable macro-level performance under imbalance than TF–IDF baselines, particularly for semantically adja- cent classes. A confusion-matrix driven analysis identifies two dominant confusion clusters (awareness vs wellness and anxiety vs depression) and motivates intent-aware modeling, calibration and uncertainty reporting. We also provide a rigorous reporting protocol (splits, leakage con-trol, seeds), per-class metrics derived from the supplied confusion matrix, and a Q1-ready ablation blueprint.