Knowledge graph embedding is the technique of representing the entities and relations of a knowledge graph as continuous low-dimensional vectors that preserve the graph’s structural and semantic regularities. Scoring functions over these vectors model the plausibility of triples, enabling tasks such as link prediction, entity resolution and similarity-based retrieval through algebraic operations rather than symbolic traversal. It bridges symbolic knowledge representation with vector-based machine learning.
Overview
- Each entity and relation is assigned a vector, and a scoring function measures the plausibility of a candidate triple (head, relation, tail).
- Translational models treat relations as translations in vector space, while bilinear and neural models capture richer interactions.
- Graph neural networks aggregate neighbourhood structure to produce context-aware entity embeddings.
- The learned space supports algebraic inference and approximate nearest-neighbour retrieval over knowledge.
Mechanisms
- A scoring function ranks true triples above corrupted ones during training.
- Negative sampling supplies contrastive examples for optimisation.
- Regularisation and normalisation constrain the geometry of the space.
- Inductive variants generalise to entities unseen during training.
Applications
- Link prediction and knowledge-base completion.
- Entity resolution and deduplication.
- Semantic search and recommendation over linked data.
- Feature provision for downstream reasoning and relation extraction.