This project provides a fully modular system for the automated extraction, representation, and exploration of legal knowledge from EU legislative texts such as the:
- AI Act
- GDPR
- Data Act
- Data Governance Act
Unlike cloud-based solutions, this pipeline runs entirely locally using Ollama and Zephyr 7B, ensuring privacy and full control over legal data.
It consists of three distinct tools:
- Legislative Text Parser – Parses a legal PDF into structured JSON and extracts key legal entities.
- Property Graph Creator – Uses Zephyr 7B to extract relationships between entities and outputs a filtered, normalized GraphML.
- Entity Checklist Generator – Visualizes the graph and provides Markdown reports for compliance and audit purposes.
- Converts PDF legal documents into structured JSON.
- Breaks text into Chapters, Articles, Paragraphs, and Points.
- Extracts definitions and legal entities.
- Visual UI allows navigation of legal text structure.
- Output: JSON (structure + entities)
- Uses Zephyr 7B via Ollama to extract legal relationships.
- Builds a GraphML property graph using NetworkX.
- Executes custom post-processing:
- Node normalization
- Entity merging
- Edge filtering
- Output:
graph.graphml(raw),subgraph.graphml(cleaned)
- Loads final GraphML and JSON to:
- Display all entities and their relations
- Link them to their original legal context
- Exports a Markdown report checklist
- Enables interactive exploration for legal analysts
| Category | Tools / Frameworks |
|---|---|
| Programming Languages | Java (19+), Python (3.10+) |
| UI Frameworks | Java Swing (Legislative Text Parser), JavaFX (Checklist UI) |
| LLM Execution | Ollama + Zephyr-7B |
| NLP Libraries | NLTK, Stanford CoreNLP |
| Graph Libraries | NetworkX (Python), Custom Java Graph Model + GraphML |
| Visualization | yEd-compatible GraphML output |
| Parsing & I/O | lxml, JSON, Markdown generation |
| Build Tools | Maven (Java), pip (Python) |
| Benchmarking | Precision/Recall analysis, MT-Bench (model eval) |
- OS: Windows 10+ / Linux / macOS
- RAM: 16 GB recommended
- Java: Version 19+
- Python: Version 3.10+
- CPU: AVX2 support required for Ollama
# Download and install from:
https://ollama.com
# Run the Zephyr model locally
ollama run zephyrmvn clean packagepip install nltk networkx matplotlib graphviz lxml- Support for newer LLMs (e.g., GPT-4 via API)
- Parallel LLM chunk processing
- Unified GUI for all tools
- Improved model prompting & benchmarking
Bouzampalidis Vasileios Ioannis
Department of Computer Science & Engineering, University of Ioannina
Supervisor: Prof. Panagiotis Vassiliadis