An LLM-based Multi-Agent System for Bilingual Legal Term Extractor

English | 中文 | 日本語

A powerful AI-driven tool for extracting, normalizing, and standardizing bilingual terminology from parallel texts, designed for legal, technical, and professional documents.

✨ Key Features

• 🤖 AI-Powered Extraction: Intelligently identifies bilingual term pairs using advanced LLMs (GPT-4, Claude, DeepSeek, etc.)
• 🔍 Quality Control: Automatically evaluates term alignment quality and filters low-quality results
• 📝 Intelligent Normalization:
  • Chinese: Traditional/Simplified unification, structural markers (第XX条)
  • English: Singular/plural normalization, verb tense unification, structural markers (Article XX)
  • Japanese: Notation unification, okurigana standardization, structural markers
• 🎯 Deduplication & Standardization: Intelligently merges synonym variants and selects best translations
• ⚡ High Performance: Supports concurrent processing for large-scale documents
• 🌐 Multilingual Support: Chinese, English, Japanese, and more

📦 Installation

Requirements

• Python 3.8+
• OpenAI API key (or OpenAI-compatible API)

Quick Install

# Clone the repository
git clone https://github.com/wang-h/bilingual_term_extractor.git
cd bilingual_term_extractor

# Install dependencies
pip install -r requirements.txt

# Configure environment variables
cp .env.example .env
# Edit .env and set your API keys:
# OPENAI_API_KEY=your-api-key-here
# OPENAI_BASE_URL=https://api.openai.com/v1  # Optional: for OpenAI-compatible APIs
# OPENAI_API_MODEL=gpt-4o-mini  # Optional: default model

🚀 Quick Start

Basic Usage

import asyncio
from src.agents.bilingual_term_extract import BilingualTermExtractAgent
from src.agents.bilingual_term_quality_check import BilingualTermQualityCheckAgent
from src.agents.bilingual_term_normalization import TermNormalizationAgent
from src.agents.bilingual_term_standardization import BilingualTermStandardizationAgent

async def extract_terms():
    # Source text (Chinese)
    source_text = """
    第三条 劳动者享有平等就业和选择职业的权利、取得劳动报酬的权利...
    """
    
    # Target text (English)
    target_text = """
    Article 3: Workers shall have the right to employment on an equal basis...
    """
    
    # Stage 1: Extract terms
    extract_agent = BilingualTermExtractAgent(locale='zh')
    extracted = await extract_agent.run({
        'source_text': source_text,
        'target_text': target_text,
        'src_lang': 'zh',
        'tgt_lang': 'en'
    }, None)
    
    # Stage 2: Quality check
    quality_agent = BilingualTermQualityCheckAgent(locale='zh')
    filtered = await quality_agent.run({
        'terms': [t.__dict__ for t in extracted],
        'source_text': source_text,
        'target_text': target_text,
        'src_lang': 'zh',
        'tgt_lang': 'en'
    }, None)
    
    # Stage 3: Normalize
    norm_agent = TermNormalizationAgent(locale='zh')
    normalized = await norm_agent.run({
        'terms': [t.__dict__ for t in filtered],
        'src_lang': 'zh',
        'tgt_lang': 'en'
    }, None)
    
    # Stage 4: Standardize
    std_agent = BilingualTermStandardizationAgent(locale='zh')
    final_terms = await std_agent.execute({
        'terms': [t.__dict__ for t in normalized]
    }, None)
    
    return final_terms

# Run
asyncio.run(extract_terms())

Run Example

python term_extract.py test_data/sample_zh_en_100.json -o outputs --checkpoint outputs/checkpoint.json

📊 Processing Pipeline

Raw Bilingual Texts
    ↓
[Stage 1] Term Extraction (BilingualTermExtractAgent)
    ├─ AI identifies term pairs
    ├─ Extracts context information
    └─ Confidence scoring
    ↓
[Stage 2] Quality Check (BilingualTermQualityCheckAgent)
    ├─ Semantic consistency validation
    ├─ Term accuracy evaluation
    └─ Filters low-quality results
    ↓
[Stage 3] Term Normalization (TermNormalizationAgent)
    ├─ Format standardization
    │   ├─ Chinese: Traditional/Simplified, "第36条"→"第XX条"
    │   ├─ English: Singular/plural, "Article 36"→"Article XX"
    │   └─ Japanese: Notation, "第36条"→"第XX条"
    ├─ Tense unification (English)
    └─ Abbreviation standardization
    ↓
[Stage 4] Deduplication & Standardization (BilingualTermStandardizationAgent)
    ├─ Deduplicate by normalized forms
    ├─ Merge synonym variants
    └─ Select best translations
    ↓
Final Standardized Terminology

🎯 Normalization Rules

Chinese Normalization

1. Traditional/Simplified: 協議 → 协议
2. Abbreviation: 有限公司 → 有限责任公司
3. Structural Markers:
   • 第36条 → 第XX条
   • 第三十六条 → 第XX条
   • 第40条第一项 → 第XX条第XX项
   • 第二章 → 第XX章
   • （一） → （XX）

English Normalization

1. Singular/Plural: contracts → contract/contracts
2. Verb Tense: terminated → terminate
3. Structural Markers:
   • Article 36 → Article XX
   • Section 5 → Section XX
   • Chapter 3 → Chapter XX
   • Paragraph 2 → Paragraph XX

Japanese Normalization

1. Notation: けいやく → 契約
2. Okurigana: Following Cabinet Notice standards
3. Structural Markers:
   • 第36条 → 第XX条
   • 第三十六条 → 第XX条
   • 第2章 → 第XX章

📁 Project Structure

bilingual_term_extractor/
├── src/
│   ├── agents/              # Core Agent modules
│   │   ├── base.py         # Base Agent class
│   │   ├── bilingual_term_extract.py
│   │   ├── bilingual_term_quality_check.py
│   │   ├── bilingual_term_normalization.py
│   │   └── bilingual_term_standardization.py
│   ├── lib/                 # Utility libraries
│   │   └── llm_client.py   # LLM client
│   └── workflows/           # Workflows
│       └── bilingual_term_extract.py
├── examples/                # Example scripts
│   ├── simple_extract.py   # Simple example
│   └── concurrent_bilingual_term_extract_v2.py  # Concurrent processing
├── outputs/                 # Output directory
├── requirements.txt
├── README.md               # English documentation
├── README_zh.md            # Chinese documentation
└── README_ja.md            # Japanese documentation

⚙️ Configuration

LLM Configuration

Supports the following LLM providers:

• OpenAI (GPT-4, GPT-4-turbo, GPT-3.5)
• Anthropic (Claude-3.5-sonnet, Claude-3-opus)
• Other OpenAI-compatible APIs

Term Extraction Configuration

# Quality check batch size
batch_size = 10

# Maximum target terms per source term
max_targets_per_source = 3

# Scoring weights
confidence_weight = 0.4
quality_weight = 0.6

📊 Output Format

Example of standardized term output:

{
  "source_term": "劳动报酬",
  "target_term": "remuneration for work",
  "original_source_term": "劳动报酬",
  "original_target_term": "remuneration for work",
  "category": "Legal Concept",
  "confidence": 0.95,
  "quality_score": 0.92,
  "combined_score": 0.93,
  "law": "Labor Law",
  "domain": "LaborLaw",
  "year": 1995,
  "occurrence_count": 3
}

🔧 Advanced Usage

Concurrent Batch Processing

Use concurrent_bilingual_term_extract_v2.py for large-scale document processing:

python examples/concurrent_bilingual_term_extract_v2.py \
    --input data/parallel_texts.json \
    --output outputs/ \
    --max-workers 5

Custom Normalization Rules

Customize normalization behavior by modifying prompt templates in Agents:

# Add custom rules in TermNormalizationAgent
custom_rules = """
7. **Custom Rules**: Your domain-specific rules
   - Example: Specialized terminology handling
"""

🤝 Contributing

Contributions are welcome! Please:

1. Fork the repository
2. Create a feature branch (git checkout -b feature/AmazingFeature)
3. Commit your changes (git commit -m 'Add some AmazingFeature')
4. Push to the branch (git push origin feature/AmazingFeature)
5. Open a Pull Request

📄 License

This project is licensed under the MIT License - see the LICENSE file for details

📧 Contact

• Project Homepage: https://github.com/wang-h/bilingual_term_extractor
• Issue Tracker: Issues

🙏 Acknowledgments

• OpenAI for GPT models
• Anthropic for Claude models
• All contributors

---

Note: Using this tool requires valid LLM API keys. Please ensure compliance with relevant terms of service.