Study Guide · Knowledge Representation & Reasoning · 7 min read
How Knowledge Representation Works Under the Hood
From ontologies to embeddings to graph databases, representation systems share one goal: make what is known explicit enough to query and consistent enough to trust.
Ontologies: the agreed vocabulary
An ontology defines the types of things that exist in your domain — Patient, Drug, Diagnosis — the relationships between them, and their rules. It is a schema for knowledge, agreed before facts are stored.
Shared vocabularies prevent the classic enterprise disease of five names for the same customer. With an ontology, 'client', 'account holder', and 'buyer' can formally mean the same entity.
Graphs, triples, and queries
Under the hood, most graphs reduce to triples: subject–predicate–object. Billions of them live in graph databases that traverse connections at speeds relational joins cannot match for multi-hop questions.
Multi-hop matters: 'which suppliers employ workers affected by this new regulation?' chains several relationships — exactly where graphs outperform keyword search.
Embeddings meet symbols
Knowledge graph embeddings compress entities into vectors, letting ML systems reason over structured facts statistically — link prediction suggests missing edges worth adding.
The frontier combines both worlds: language models read unstructured text, extract candidates, and check them against the symbolic store; retrieval from verified graphs then grounds model answers. Each covers the other's weaknesses.
Key Points
- Ontologies define shared vocabulary so integration stops being guesswork.
- Triples (subject–predicate–object) are the atoms of machine-processable knowledge.
- Graph databases excel at multi-hop questions across chained relationships.
- Neural extraction + symbolic verification is today's most promising hybrid pattern.
All study guides for this term: Knowledge Representation, Explained Simply · How Knowledge Representation Works Under the Hood · Knowledge Representation in the Real World