- https://arxiv.org/pdf/1805.02408.pdf
- Source: arxiv
TL;DR
Injecting entity type constraints and non-negativity constraints into existing KG embedding models consistently improves link prediction accuracy on FB15k-237 and WN18RR with minimal added complexity.
Summary
Knowledge graph embedding models like TransE and DistMult suffer from scoring invalid entity candidates during link prediction. This ACL 2018 paper injects two simple ontological constraints: entity type constraints that filter candidates incompatible with a relation’s domain/range, and non-negativity constraints on entity embeddings that improve interpretability by enforcing a parts-based representation. Evaluated on FB15k-237 and WN18RR, constraint-enhanced models consistently outperform unconstrained counterparts, demonstrating that structural ontological knowledge is a low-cost, high-impact signal for KG completion.
Key contributions
- Introduces entity type constraints that eliminate invalid candidates during link prediction scoring using relation domain/range metadata.
- Proposes non-negativity constraints on embeddings that encourage interpretable parts-based representations.
- Demonstrates consistent improvement on FB15k-237 and WN18RR by applying constraints to existing models (TransE, DistMult) without architectural changes.
- Establishes that ontological metadata already in KGs is a zero-cost source of inductive bias for embedding models.
When to cite
- When adding type constraints to knowledge graph embeddings as a model-agnostic quality improvement.
- When arguing that KG completion models routinely underuse structured ontological metadata available in the graph.
- When discussing non-negativity regularization as a technique that simultaneously improves accuracy and interpretability.
- When motivating the use of domain/range schema in KG embedding to reduce the effective candidate search space.