project / ML · NLP
Bias-Aware Mental Health Sentiment Analysis
- Python
- BERT
- HuggingFace
- Pandas
- Scikit-learn
- Reddit API
Problem
Sentiment models misread mental-health posts written in non-standard dialects far more often than standard English.
Outcome
Fine-tuned BERT at 84.0% accuracy, with dialect bias (DPD) cut from 0.327 to 0.066.
Overview
Master's thesis (FairBERT): fine-tuned BERT with adversarial debiasing and counterfactual augmentation on RMH-Bias-10K, a 10,000-post Reddit mental-health dataset built for this research. Achieved 84.0% accuracy, macro-F1 0.688, and ROC-AUC 0.843 — outperforming BiLSTM and vanilla BERT baselines — while reducing mean demographic parity difference by 35.4% and dialect DPD by 79.8% (0.327 → 0.066). Written as an IEEE-format thesis at NCIT, Pokhara University.