ducut91/korean-court-judgments

📚 Korean Court Judgments Dataset (Enhanced Edition)

This dataset provides a preprocessed and enriched collection of 163,546 court rulings from Korean judicial bodies, scraped from the National Legal Information Sharing Service. It is suitable for legal NLP, judgment summarization, legal information retrieval, and generative AI applications.

🚀 What's New:

• AI-generated judgment summaries for missing entries (using GPT-4o-mini)
• Cleaned and normalized reference statutes (regex-based parsing)
• Citation count (number of times cited by other cases)
• Token count estimation for each case
• Court hierarchy scoring (Supreme Court: 10, High Court: 7, District Court: 3)

---

📦 Dataset Summary

• Total Records: 163,546 court rulings
• Language: Korean (ko)
• Fields: 16 columns with English names
• Source: open.law.go.kr
• Version: Enhanced Edition (v2.0)
• Contact: ducut91@gmail.com
• License: This dataset is based on public legal data provided by the National Law Information Center (law.go.kr) and is freely available for anyone to use. You may use, modify, and redistribute the data for research, commercial, or educational purposes without restriction.

Enhancements Applied

1. AI-Generated Summaries: For cases missing judgment_summary, summaries were automatically generated using GPT-4o-mini and gemini2.5-flash-lite based on the full case_content
2. Reference Statute Normalization: Cleaned and formatted statute references using regex patterns for better readability
3. Citation Network: Added reference_case_count - the number of times each case is cited by other cases in the dataset
4. Token Count: Estimated token count for each case (useful for LLM context window planning)
5. Court Hierarchy Score: Added court_level_score for ranking:
   • Supreme Court (대법원): 10
   • High Court (고등법원): 7
   • District Court (지방법원, 특허법원, etc.): 3

---

🧾 Column Descriptions

Column Name	Type	Description
case_id	string	Unique identifier of the judgment
case_name	string	Name or title of the legal case
case_number	string	Official case number assigned by the court (e.g., "2013누26042")
decision_date	string	Date of the judgment (format: YYYYMMDD)
court_name	string	Name of the court that issued the ruling (e.g., "서울고등법원", "대법원")
case_type	string	Type/category of the case: "민사", "형사", "세무", etc.
issues	string	Legal issues or principles considered (may be empty)
judgment_summary	string	Summary of the court's ruling (original or AI-generated) ⭐
judged_statute	string	Constitutional provisions under review (for Constitutional Court cases only)
reference_statute	string	Relevant statutes cited (cleaned and normalized) ⭐
reference_case	string	Previous precedents cited
case_content	string	Full text of the ruling
case_detail_link	string	Link to the full case document on the official government website
token_count	integer	Estimated token count (using tiktoken) ⭐
reference_case_count	integer	Number of times this case is cited by other cases in the dataset ⭐
court_level_score	integer	Court hierarchy score: Supreme (10), High (7), District (3) ⭐

⭐ = Enhanced/Added fields in this version

---

📊 Data Quality Statistics

Missing Value Statistics

Column Name	Missing Count	Missing Ratio (%)	Notes
judged_statute	163,546	100.00%	Only for Constitutional Court (currently excluded)
issues	88,081	53.86%	Many cases don't extract issues
reference_case	79,224	48.44%	Many cases cite no precedents
reference_statute	55,909	34.19%	Cleaned and normalized
judgment_summary	0	0.00%	✅ AI-generated for missing values
case_name	10	0.01%	Nearly complete
Other core columns	0	0.00%	Complete

Enhanced Field Statistics

Field	Min	Max	Mean	Median	95th %tile	Description
token_count	51	558,488	2,936	2,145	7,613	Token count (tiktoken cl100k_base)
reference_case_count	0	808	1.16	0	8	Citation count (78.8% cited 0 times)
court_level_score	5	10	-	-	-	Supreme: 10 (44.3%), High: 7 (21.3%), District: 5 (34.3%)

Token Count Distribution:

• 25th percentile: 1,152 tokens
• 75th percentile: 3,618 tokens
• 99th percentile: 14,644 tokens
• Cases with >50K tokens: ~0.1% (extreme outliers)

Citation Count Distribution:

• Cited 0 times: 128,881 (78.8%)
• Cited 1-5 times: 26,044 (15.9%)
• Cited 6-10 times: 5,049 (3.1%)
• Cited 11-50 times: 3,268 (2.0%)
• Cited 50+ times: 304 (0.2%) - highly influential cases

⚠️ Data Quality Note: All missing judgment_summary values were filled using GPT-4o-mini based on case_content. These AI-generated summaries are marked internally for transparency.

---

🛠️ Preprocessing Pipeline

1. Data Collection

• Source: National Legal Information Sharing Service
• Method: Web scraping via official API/portal
• Period: All available historical data up to crawl date

2. Data Cleaning & Normalization

• Column Standardization: Korean column names → English names
• Text Normalization: Whitespace trimming, encoding fixes
• Date Formatting: Standardized to YYYYMMDD format

3. AI-Based Enhancement

• Summary Generation:
  • Used GPT-4o-mini to generate judgment_summary for cases missing both issues and original summary
  • Prompt: "Summarize the key legal points of this court ruling in 2-3 sentences"
  • Applied to ~X,XXX cases (exact count available in logs)

4. Reference Processing

• Statute Normalization:
  • Regex-based parsing of reference_statute field
  • Extracted law names and article numbers
  • Standardized format: "법명 제X조, 법명 제Y조"

5. Network Analysis

• Citation Network Construction:
  • Parsed reference_case field to extract cited case numbers
  • Built case-to-case citation graph
  • Calculated reference_case_count (in-degree) for each case
  • Used for ranking and importance scoring

6. Feature Engineering

• Token Count: Estimated using tiktoken library (cl100k_base encoding)
  • Range: 51 to 558,488 tokens
  • Median: 2,145 tokens (suitable for most LLM context windows)
  • 95% of cases: <7,613 tokens
• Court Level Score: Rule-based assignment:

  if "대법원" in court_name: score = 10  # Supreme Court: 44.3%
  elif "고등법원" in court_name: score = 7  # High Court: 21.3%
  else: score = 5  # District/Special courts: 34.3%
  

---

💡 Suggested Use Cases

Legal AI Applications

• 🧠 Legal Document Summarization – Train/evaluate models to summarize court rulings
• 💬 Legal Q&A Systems – Use in retrieval-augmented generation (RAG) pipelines
• 🔍 Semantic Search – Build hybrid search systems (BM25 + dense embeddings)
• ⚖️ Legal Judgment Prediction – Predict case outcomes based on facts and precedents

Research Applications

• 📊 Citation Network Analysis – Study precedent influence using reference_case_count
• 🏛️ Legal Reasoning Analysis – Analyze how courts reference statutes and prior cases
• 📈 Court Hierarchy Studies – Use court_level_score to study decision propagation
• 🔬 Legal NLP Benchmarking – Evaluate Korean legal language models

Engineering Applications

• 🗂️ Knowledge Graph Construction – Build legal knowledge graphs from citations
• 🔧 Search Engine Development – Train ranking models using citation and court hierarchy signals
• 📦 LLM Context Planning – Use token_count for efficient context window management
• ⚡ Elasticsearch Indexing – Ready for hybrid search with computed features

---

📜 License & Purpose

This dataset is based entirely on open government data provided by the Republic of Korea. Under Korean law, public legal information is not subject to copyright and may be freely used, modified, and redistributed for research, commercial, or educational purposes without restriction.

No attribution is legally required, but citing the open.law.go.kr source is recommended.

---

🎯 Why We Open-Sourced This

We believe that global large language models (LLMs) should better understand and reflect the legal system of South Korea, just as they do for other major jurisdictions.

By releasing this dataset publicly, we hope to:

• Promote inclusion of Korean law in multilingual legal AI systems
• Encourage international research in Korean legal language processing
• Lower the barrier to entry for developers working on Korean legal services

🇰🇷 A legally informed model is a better model — for everyone.

---

📅 Changelog

Date	Version	Changes Made
2025-10-07	v2.0	Enhanced Edition: Added AI summaries, citation counts, token counts, court scores, normalized statutes
2025-07-18	v1.0	Initial release: Column headers translated to English

---

🔧 Technical Details

Preprocessing Scripts

• Script: scripts/preprocess_data.py
• Dependencies: pandas, tiktoken, openai, regex
• Runtime: ~X hours on standard hardware (batch processing with rate limiting)

Quality Assurance

• ✅ All rows validated for required fields (case_id, case_number, decision_date, court_name)
• ✅ Token counts verified against sample manual counts (±5% accuracy)
• ✅ Citation network integrity checked (no dangling references to non-existent cases)
• ✅ Court level scores manually reviewed for 100 random samples

Known Limitations

1. AI-Generated Summaries: May not perfectly capture all nuances of the original ruling
2. Citation Count: Only counts citations within this dataset (external citations not tracked)
   • 78.8% of cases have 0 citations (may be less influential or too recent)
3. Token Count: Estimated using tiktoken; actual LLM token usage may vary by ±5%
   • Extreme outliers exist (max: 558K tokens = ~A4 200+ pages)
4. Court Score: Simplified 3-tier system (10/7/5); doesn't distinguish specialized courts' authority in their domains
5. Constitutional Court Cases: judged_statute field is 100% empty as Constitutional Court cases were excluded from this dataset

---

📚 Citation

If you use this dataset in your research, please cite:

@dataset{korean_court_judgments_enhanced_2025,
  title={Korean Court Judgments Dataset (Enhanced Edition)},
  author={[Your Name/Institution]},
  year={2025},
  version={2.0},
  url={https://github.com/[your-repo]},
  note={Enhanced with AI summaries, citation analysis, and ranking features}
}