Agentic data science with the Data Agent Kit
cloud.google.com·6h ago
AI Research
arXiv CS.AI·Aug 3·5 min read
An Ontology-Guided, Deduplication-Aware Extraction Layer for Knowledge Graph Construction from Heterogeneous Documents
TL;DR
A new extraction layer for knowledge graph construction was developed, utilizing ontology-guided methods to enhance entity and relationship extraction from diverse document types. This system significantly improved search recall from 70% to 95% while eliminating false merges.
✦ Why It Matters
Implement ontology-guided extraction in your document processing pipeline to enhance data accuracy and retrieval.
Key Takeaways
- Integrate real-time ontology retrieval to reduce catalog overhead.
- Utilize Qwen3.5-9B for improved entity extraction.
- Implement five-stage refinement for data accuracy.
- Adopt deduplication algorithms to prevent entity conflation.
How It Works
The extraction layer processes documents by first routing them through format-specific handlers. It then extracts entities and relationships in two passes using the Qwen3.5-9B model, which is enhanced by live retrieval of relevant ontology slices from a graph database.
This approach minimizes the need for static domain slices, significantly reducing catalog overhead.