Ontario Reg 153/04 Chemical Name Normalization System

A self-training chemical name matcher that normalizes lab analyte names to canonical Ontario Regulation 153/04 identities.

Architecture

Three-Layer Learning System:

• Layer 1 (Primary): Continuous synonym accretion - every validation adds knowledge immediately
• Layer 2 (Calibration): Monthly statistical threshold tuning
• Layer 3 (Neural): Rare model retraining (6-12 months, only when signal plateaus)

Key Innovation: System matures through vocabulary growth, not perpetual AI retraining. After 12 months, becomes an Ontario environmental language registry with 80%+ exact match rate.

Project Structure

reg153-chemical-matcher/
├── data/
│   ├── raw/                      # Original source files
│   │   ├── reg153/               # Regulation PDFs
│   │   ├── api_harvest/          # API response dumps
│   │   └── lab_edds/             # Lab reports
│   ├── processed/
│   │   ├── canonical/            # reg153_master.csv (ground truth)
│   │   └── synonyms/             # Merged synonym corpus
│   ├── training/                 # Labeled ground truth
│   ├── embeddings/               # Vectors + FAISS indices
│   └── snapshots/                # Versioned releases
├── src/
│   ├── bootstrap/                # API harvesters (one-time use)
│   ├── extraction/               # PDF parsing + validation
│   ├── normalization/            # Text cleaning (analyte-aware)
│   ├── matching/                 # Resolution engine
│   ├── learning/                 # Incremental learning system
│   └── database/                 # SQLite models
├── scripts/                      # Numbered workflow scripts
├── tests/                        # Comprehensive test suite
└── docs/                         # Architecture + lineage docs

Quick Start

# Install dependencies
pip install -r requirements.txt

# Setup database
python scripts/setup_database.py

# Load synthetic demo data (or connect your own lab data)
python scripts/generate_synthetic_edds.py

# Bootstrap API synonyms (PubChem + NCI - no auth needed)
python scripts/04_harvest_api_synonyms.py

# Generate embeddings
python scripts/09_generate_embeddings.py

# Match chemicals from lab report
python scripts/22_batch_ingest.py --input lab_data.xlsx

APIs Used (Bootstrap Phase Only)

• PubChem - No authentication required, 5 req/sec
• NCI Chemical Identifier Resolver - Open service for structure conversion
• NPRI (Canada) - Open government data
• CompTox (EPA) - Environmental chemical database
• CAS Common Chemistry - Free registration required

APIs used ONCE for synonym harvesting, then system runs locally.

Maturity Trajectory

Metric	Month 1	Month 6	Month 12
Exact match rate	40%	71%	82%+
Semantic reliance	45%	24%	<20%
Unknown rate	15%	5%	<2%
Validation effort	15 hrs/wk	8 hrs/wk	2-3 hrs/wk

License

Internal tool for environmental consulting - Ontario-specific implementation.

Version

v0.1.0 - Initial implementation (2026-02-12)