A Neuro-Symbolic framework for automated conflict detection in ODRL data usage policies. Uses LLMs to bridge semantic gaps and deterministic N3 logic (EYE Reasoner) for verified conflict detection.
.png)
Data ecosystems involve multiple stakeholders (Data Providers, Governance Authorities) with potentially conflicting policies. This pipeline automatically detects semantic conflicts that keyword matching would miss.
| Component | Folder | Description |
|---|---|---|
| Data Policies (𝒢ₚ) | pipeline/offline/input/policy_kg/ |
Input ODRL policies in Turtle format |
| LLM-Agent Abstraction | pipeline/offline/prompts/abstraction/ |
Prompt for generating abstraction rules (𝑅ₐᵦₛ) |
| LLM-Agent Interaction | pipeline/offline/prompts/interaction/ |
Prompt for generating interaction rules (𝑅ᵢₙₜ) |
| Rulebook (𝑅ᵦₒₒₖ) | pipeline/offline/output/ |
Generated N3 rules combining 𝑅ₐᵦₛ ∪ 𝑅ᵢₙₜ |
| Semantic Reasoner | pipeline/online/ |
EYE reasoner applies rulebook to detect conflicts |
| Discovered Conflicts | pipeline/online/output/ |
Analysis results with conflict explanations |
# Install dependencies
npm install -g eye-js
pip install -r requirements.txt
# Set your Google API key (required for Gemini LLM)
export GOOGLE_API_KEY="your-api-key-here"
# Or create a .env file with: GOOGLE_API_KEY=your-api-key-here
# Run full pipeline
python3 pipeline/run_pipeline.py├── pipeline/
│ ├── offline/ # Phase 1: Rule Generation (LLM)
│ │ ├── input/policy_kg/ # ODRL policies (30 policies)
│ │ ├── prompts/ # Abstraction & Interaction prompts
│ │ └── output/ # Generated rulebook.n3
│ └── online/ # Phase 2: Conflict Detection (EYE)
│ └── output/ # Conflict analysis results
├── example-policies/ # Individual policy files (30 TTL files)
├── experiments/ # Evaluation scripts
│ └── generate_stats.py # Pipeline statistics generator
└── img/ # Documentation images
| Metric | Value |
|---|---|
| Policies | 30 (22 original + 8 synthetic) |
| Semantic Concepts | 15 |
| Conflict Pairs Detected | 243 |
| Conflict Types | 6 (Access, Geographic, Performance, Temporal, Integrity, Bandwidth) |