Semi-supervised Regulatory NLP
Semi-supervised text classification for regulatory documents using IndoBERT with contrastive learning and pseudo-labeling, where labeled data is scarce.
- NLP
- Semi-supervised Learning
- Contrastive Learning
- IndoBERT
An exploration of learning from unlabeled Indonesian regulatory text when gold labels are expensive.
Problem
Regulatory documents in Indonesian need classification, but annotation is slow and expert-dependent. A practical classifier must get by with very few labels.
Research Question
Can semi-supervised learning — pseudo-labeling plus contrastive pretraining on unlabeled text — match supervised performance with a fraction of the labeled data?
System
Unlabeled regulatory text ──► contrastive pretraining (IndoBERT)
│
Small labeled set ──► fine-tune ──► pseudo-label the rest
│
retrain / iterate
│
▼
Evaluation (held-out)
- IndoBERT as the backbone encoder.
- Contrastive learning over unlabeled documents to learn useful representations.
- Pseudo-labeling loop to expand the effective training set.
My Contribution
- Preprocessing pipeline for Indonesian regulatory text.
- Contrastive training setup and data augmentation for positive pairs.
- Pseudo-labeling loop with confidence thresholds and iteration.
- Evaluation across label-set sizes to show the data-efficiency curve.
Experiments
Trained with 1%, 5%, 10%, and 25% of the labeled data, comparing supervised vs semi-supervised regimes.
Results
Semi-supervised training recovered most of the supervised performance at low label budgets; contrastive pretraining helped most in the extreme low-label setting.
Results are placeholder text — replace with actual numbers.
What I Learned
- Label efficiency is a systems decision: augmentation, confidence thresholding, and iteration matter.
- Contrastive objectives are a strong fit when unlabeled data is abundant.
- Pseudo-labeling is sensitive to threshold calibration.