The terminology problem in knowledge graphs
The knowledge graph field has a terminology problem. Not a small one.
When a client says “we need reasoning over our knowledge graph,” I’ve learned to pause before answering. They might mean: - OWL/RDFS entailment (subclass propagation, inverse properties) - a Datalog or SPARQL-based rules engine (RDFox) - a constraint validator (SHACL, ShEx) - an LLM generating inferences over graph context - a simulation stepping through state transitions (computational graphs) - a probabilistic reasoner (Markov logic, belief propagation, Bayesian networks).
These are not variations on the same thing. They have different computational complexity classes, different failure modes, and different integration costs.
It doesn’t stop there.
“Ontology” can mean: a formal OWL class hierarchy with axioms, a lightweight schema in your LPG, a business glossary someone put in Confluence, or a controlled vocabulary that was once a thesaurus. The word carries radically different engineering implications depending on which camp your client just walked in from.
“Semantic layer” might refer to: provenance and lineage tracking, a federated query interface across heterogeneous sources, a virtual knowledge graph layer (Ontopic, Stardog), an RDF named-graph partitioning strategy, or (increasingly) a GraphRAG retrieval layer someone branded “semantic” because it felt right.
The word “Graph” itself: property graph, RDF graph, hypergraph, computational graph, factor graph, scene graph, knowledge graph used synonymously with “database with edges.”
“Embedding” in a graph context: node2vec, graph transformer representations, RDF2Vec, knowledge graph embeddings (TransE, RotatE), or just the vector stored on a node property.
And then there’s “knowledge graph” itself, used to describe everything from a Wikidata-scale linked data store to a Neo4j schema with three node types and a handful of relationships.
The terminology hasn’t stabilized because the field is genuinely in chaos: endless platforms, storage solutions, big claims, and LLM-native architectures all pulling on the same vocabulary from different directions.
And yet the response from most teams is to move faster, not slower. Find a platform with easy ontology ingestion, wire up a pipeline, ship something. The thinking gets deferred because coding feels like progress and whiteboarding feels like delay.
But knowledge graphs are unforgiving of underspecified concepts. The graph structure you commit too early propagates through every query, every downstream integration. When “reasoning” meant five different things at the start of the project, you don’t discover that until later when untangling it costs ten times more than the conversation you didn’t have.
Artificial intelligence can accelerate a lot of this work. It cannot substitute for the human patience required to ask: what do we actually mean by that word?