banglamentalBERT-mBERT (DAPT + TAPT)
Project: Detecting Mental Health and Suicidal Tendencies on Social Media using Multimodal NLP
Institution: BRAC University, Department of CSE
Model Description
Fine-tuned bert-base-multilingual-cased (mBERT) for 4-class depression severity detection in Bangla text.
This model achieved the largest relative gain (+0.88%) over its baseline in the project by using DAPT on 250k mental health posts followed by TAPT on 17k synthetic posts (Experiment: fix_dapt Exp A).
Performance (Test Set)
| Metric | Value |
|---|---|
| Accuracy | 85.05% |
| Weighted F1 | 85.05% |
| Label 3 (Moderate) F1 | 0.705 |
Pipeline
mBERT -> DAPT (250k posts) -> TAPT (17k synthetic texts) -> Fine-tune (3,426 real texts)