NAKG
AI Research
Neural Adaptive Knowledge Graph for medical document understanding. Perfect F1 scores on knowledge graph queries. 62% higher recall on complex queries. Deployed in production.
Machine Learning
Key Features
- Neural Adaptive Knowledge Graph architecture
- Edge Transformer for dynamic relationship learning
- Hierarchical document clustering
- Medical document processing
- Complex relationship detection
- F1 scores of 0.982±0.015 on knowledge graph tasks
- 62.0% higher recall on complex queries
NAKG (Neural Adaptive Knowledge Graphs) is a document understanding system for medical billing. In medical billing, the answer is usually in several related documents, not in one single document. The system does not treat documents as isolated chunks. It learns the connections between them.
Core Components
- Edge Transformer, which learns relationships between documents dynamically
- Neural Gossip Network, which shares information between related documents
- Hierarchical clustering for large document sets
Results
- 1.000 F1 on knowledge graph queries
- 0.978±0.012 F1 on temporal reasoning
- 62.0% higher recall on complex queries
Training
The training has three phases: pre-training on medical documents, relationship learning, then end-to-end fine-tuning.
NAKG runs in production at Backwork, where it processes millions of medical billing documents.