A neuro-symbolic knowledge graph system with clade-inspired hierarchy and embedding clustering to control ontology growth and mitigate subclassing explosion.
In large ontologies, subclassing explosion is pervasive. Every time a new concept is added, the path of least resistance is to create a new subclass of something nearby. The result is ontologies thousands of nodes deep with near-duplicate classes, poor generalisation, and brittle inference. Existing approaches — OWL reasoners, manual curation — do not scale.
Veloclade approaches this differently: rather than preventing new nodes, it controls where they go by combining biological cladistics with semantic embedding clustering.
New concept proposal
│
▼
┌───────────────────────┐
│ Embedding encoder │ sentence-transformers
└──────────┬────────────┘
│
▼
┌───────────────────────┐
│ Clade membership │ Find nearest existing clade
│ classifier │ by embedding similarity
└──────────┬────────────┘
│
▼
┌───────────────────────┐
│ Growth policy engine │ Merge / place in clade / create new clade
│ │ based on configurable thresholds
└──────────┬────────────┘
│
▼
┌───────────────────────┐
│ Knowledge graph │ RDFLib / custom graph store
│ update │
└───────────────────────┘
Three outcomes are possible for any proposed concept:
| Decision | Condition | Effect |
|---|---|---|
| Merge | High similarity to existing node (> threshold) | Concept unified with existing node; no new node created |
| Place | Moderate similarity to a clade | Concept added within existing clade |
| New clade | Low similarity to all clades | New monophyletic clade created |
pip install -r requirements.txt
python -m veloclade.demoThis runs a demonstration of controlled ontology growth on a sample biological classification dataset, showing subclassing explosion in an unconstrained graph versus controlled growth under Veloclade's policy engine.
- Python 3.10+
- sentence-transformers
- RDFLib
- scikit-learn
- Kulmanov et al. (2019) — ELEmbeddings: Geometric construction of models for the Description Logic EL++
- Chen et al. (2021) — Ontology-Enhanced Pre-training for Bio-Medical NLP
- koios — ontology-grounded transformer for knowledge-augmented reasoning
- kalmanorix — specialist embedding routing and retrieval evaluation
- ananke — ontology-driven world-building system (a practical application of controlled ontology growth)
Research prototype. Core clade membership classifier and growth policy engine are implemented. Evaluation against standard ontology benchmarks is planned.
MIT