Ontology Learning and Knowledge Graph Construction
Citation: da Cruz, T., Tavares, B., & Belo, F. (2025). Ontology Learning and Knowledge Graph Construction: A Comparison of Approaches and Their Impact on RAG Performance. arXiv:2511.05991.
Core Question
How do different knowledge graph construction strategies influence RAG performance, specifically comparing vector-based retrieval, GraphRAG, and ontology-guided approaches derived from either text corpora or relational databases?
Key Findings
Performance results (20-question evaluation):
- GraphRAG: 90% correct (18/20)
- Text Ontology KG with chunks: 90% correct
- RDB Ontology KG with chunks: 90% correct
- Vector RAG baseline: 60% correct (12/20)
- Ontology KGs without chunks: 15-20% correct
Critical insight: Chunk integration is essential. Adding textual chunks to KG nodes dramatically improves performance. The symbolic structure alone isn’t enough.
Practical finding: Database-derived ontologies perform as well as text-derived ones, with two advantages:
- One-time ontology learning (schemas are stable)
- No complex ontology merging required
Methodology
Compared six configurations on a real grant application corpus:
- Vector RAG (FAISS + cosine similarity)
- Microsoft GraphRAG (default local search)
- Text Ontology KG (with and without chunks)
- RDB Ontology KG (with and without chunks)
Custom retriever using Prize-Collecting Steiner Tree optimization for subgraph extraction.
Extracted Content
→ 06-atom—chunk-integration-critical → 04-atom—schema-stability-advantage → 06-atom—ontology-merging-complexity → 07-molecule—vectors-vs-graphs → 06-atom—ontology-guided-kg-construction