Besshouka (別称化)

A local-first Japanese PII anonymization engine. Besshouka detects personally identifiable information (PII), payment card data (PCI), and protected health information (PHI) in Japanese text and transforms it using configurable rules — all without sending data to any external service.

Note: Besshouka is in early development (alpha). It is not yet recommended for production use. Contributions to improve accuracy, coverage, and robustness are welcome — see CONTRIBUTING.md.

Why Besshouka?

• Japanese-native — built specifically for Japanese data patterns: マイナンバー, Japanese phone formats, postal codes, full-width character handling, and GiNZA-powered NER for names, organizations, and locations.
• Local-first — everything runs on your machine. No cloud APIs, no data leaves the device.
• Pluggable — add custom regex recognizers via YAML, write your own operators in Python, or plug in any importable function as a custom operator. No forking required.
• Auditable — every anonymization operation is logged in an audit trail with the original text, the operator used, and the new indices.

Quick Start

pip install besshouka

Anonymize text

besshouka anonymize "田中太郎の電話番号は090-1234-5678です"
# Output: <氏名>の電話番号は090-1234-****です

Analyze (detect only)

besshouka analyze --explain "田中太郎の電話番号は090-1234-5678です"

Adjust confidence threshold

Both commands support --threshold / -t to filter by confidence score:

# Anonymize: only anonymize detections with confidence >= threshold (default: 0.5)
besshouka anonymize --threshold 0.3 "番号は123456789018です"

# Analyze: only display detections with confidence >= threshold (default: 0.0)
besshouka analyze --threshold 0.5 --explain "マイナンバーは123456789018です"

Detections below the threshold are still detected internally but excluded from output. For example, a 12-digit number matching the My Number check digit but lacking context keywords scores 0.4 and is left untouched at the default anonymization threshold.

Use custom rules

besshouka anonymize \
  --recognizers my_patterns.yaml \
  --rules my_operators.yaml \
  --input document.txt \
  --output anonymized.txt

Programmatic Usage

from besshouka.config.loader import load_recognizer_config, load_operator_config
from besshouka.orchestrator.pipeline import run

rec_config = load_recognizer_config("path/to/recognizers.yaml")
op_config = load_operator_config("path/to/operators.yaml")

ctx = run("田中太郎の電話番号は090-1234-5678です", rec_config, op_config,
          score_threshold=0.5)

print(ctx.engine_result.text)   # anonymized text
print(ctx.engine_result.items)  # audit trail

Architecture

Text In → [Analyzer] → [Anonymizer] → Text Out

Each module has its own README with extension guides. See the besshouka/ directory.

Built-in Recognizers

Built-in Operators

Extending Besshouka

Add a regex recognizer (no code)

Add an entry to your recognizers YAML:

recognizers:
  - name: employee_id
    entity_type: EMPLOYEE_ID
    pattern: 'EMP-[A-Z]{2}\d{6}'
    score: 1.0
    source: custom

Add a custom operator (no subclassing)

Write a function anywhere importable:

def my_transform(text: str, params: dict) -> str:
    return text[::-1]  # reverse it, or whatever you need

Reference it in your operators YAML:

operators:
  EMPLOYEE_ID:
    method: custom
    function: "my_module.my_transform"

Development

git clone https://github.com/akhi/besshouka.git
cd besshouka
pip install -e ".[dev]"

Running Tests

# All tests (excluding slow GiNZA model tests)
pytest tests/ -m "not slow"

# All tests including GiNZA
pytest tests/

# With coverage
pytest tests/ --cov=besshouka --cov-report=term-missing

Requirements

• Python >=3.11, <3.14 — Python 3.14 is not yet supported due to PyO3 compatibility with SudachiPy (GiNZA's tokenizer). Python 3.13 is recommended.
• GiNZA / spaCy (for NER)
• See requirements.txt for full list

License

MIT