EarlyModernNER

Named Entity Recognition for Early Modern English documents (1500-1800).

Overview

EarlyModernNER extracts four types of entities from historical texts:

Entity Type	Description	Examples
TOPONYM	Place names	London, Jamaica, West Indies
PERSON	Individual people	Oliver Cromwell, Governor Modyford
ORGANIZATION	Institutions	East India Company, Parliament
COMMODITY	Trade goods & materials	sugar, tobacco, silk

Performance

Evaluated on 100 gold-standard annotated documents:

Entity Type	Precision	Recall	F1
TOPONYM	0.93	0.82	0.87
PERSON	0.93	0.69	0.80
ORGANIZATION	0.93	0.46	0.62
COMMODITY	0.85	0.80	0.83
Overall	0.89	0.77	0.83

Quick Start

Installation

pip install earlymodernner

Or install from source:

git clone https://github.com/polayj/earlymodernner.git
cd earlymodernner
pip install -e .

Model adapters (~680MB total) are automatically downloaded from Hugging Face Hub on first use.

Usage

# Process a single file
python -m earlymodernner --input document.txt --output results.jsonl

# Process a directory
python -m earlymodernner --input /path/to/docs/ --output results.jsonl

# Output as CSV
python -m earlymodernner --input docs/ --output results.csv --csv

# Pre-download adapters (optional, for offline use)
python -m earlymodernner --download

Output Format

JSONL (default):

{
  "doc_id": "document_name",
  "text": "The sugar trade between Jamaica and Bristol...",
  "entities": [
    {"text": "Jamaica", "type": "TOPONYM"},
    {"text": "Bristol", "type": "TOPONYM"},
    {"text": "sugar", "type": "COMMODITY"}
  ]
}

CSV (with --csv):

doc_id,entity_text,entity_type
document_name,Jamaica,TOPONYM
document_name,Bristol,TOPONYM
document_name,sugar,COMMODITY

Requirements

• Python 3.9+
• CUDA-compatible GPU with 8GB+ VRAM
• See requirements.txt for dependencies

Project Structure

earlymodernner/
├── earlymodernner/          # Main package
│   ├── __main__.py          # CLI entry point
│   ├── pipeline.py          # Inference pipeline
│   ├── constants.py         # Entity types & prompts
│   └── adapters/            # Trained LoRA adapters
├── dev/                     # Training & development tools
│   ├── train_lora.py        # Training script
│   ├── evaluate.py          # Evaluation script
│   ├── training.md          # Training documentation
│   └── config/              # Training configurations
├── docs/                    # Documentation
│   ├── usage.md             # Detailed usage guide
│   └── corpus.md            # Training corpus details
└── results/                 # Default output directory

Documentation

• Usage Guide - Detailed usage instructions, input/output formats
• Training Corpus - Data sources and annotation process
• Training Guide - How to train your own adapters

How It Works

EarlyModernNER uses an ensemble approach with four specialized models:

1. Each entity type has its own fine-tuned LoRA adapter
2. Documents are processed by all four adapters
3. Results are merged using priority-based cascade (TOPONYM → COMMODITY → PERSON → ORGANIZATION)
4. Overlapping entities are resolved by giving priority to higher-performing models

Technical details:

• Base model: Qwen3-4B-Instruct
• Fine-tuning: QLoRA (4-bit quantization)
• Training: Silver-standard annotations + synthetic hard negatives

Citation

@software{earlymodernner,
  title = {EarlyModernNER: Named Entity Recognition for Early Modern English},
  author = {Polay, Jacob},
  year = {2026},
  url = {https://github.com/polayj/earlymodernner}
}

License

MIT License

Author

Jacob Polay, MA Student, University of Saskatchewan

Acknowledgments

• Built on Qwen models
• Uses PEFT for efficient fine-tuning
• Training data from Old Bailey Online, PCEEC2, Royal Society Corpus, EEBO, and Archive.org