Ontologies and formal models of a domain

What an ontology adds beyond a taxonomy

A taxonomy arranges labels into a hierarchy and stops there. It says that one term is broader than another and nothing more. It cannot tell a machine that a concept has a required property, that two concepts exclude each other, or that a relation carries a direction that matters.

An ontology states those things with a small vocabulary: classes for kinds of thing, properties for relations between things and between things and values, and individuals for the specific things described. What makes it an ontology rather than a schema is that its statements are axioms — they constrain what can be true, and a reasoner can compute their consequences. A declared domain and range pushes anything asserted with a property into the right class; disjointness means nothing can be in both. A taxonomy classifies; an ontology constrains.

Classes, properties, individuals and entailment

The description-logic view is that an ontology describes individuals and the classes they belong to, and that every statement has logical consequences. Entailment is the name for those consequences: what must be true given the axioms and assertions, whether or not anyone wrote it down. Subclass relations compose, so an individual belongs to every class above its own; transitive, symmetric and inverse properties propagate assertions along the graph.

Two consequences matter for agentic systems. A reasoner can detect inconsistency: an individual asserted into two classes declared disjoint is a contradiction, and the model is unsound until it is resolved — a validation service prose documentation cannot provide. And ontologies are open-world by default, so a missing statement means unknown, not false. Agents and databases are usually built closed-world, and the difference catches teams out, because "there is no record, therefore the fact is false" is not available under the standard semantics without saying so explicitly.

RDF Schema is the lighter layer of the same idea: classes, subclass and subproperty relations, and domain and range, which is enough for typed graphs and simple inference without a full description logic. It is where most teams start, and RDFS covers it in detail.

OWL, its profiles, and the decidability trade-off

OWL is the Web Ontology Language, and OWL 2 is the version in current use. Its overview describes three ways of using it. OWL 2 DL is the description-logic version and is decidable, so a reasoner can always answer whether a conclusion follows. OWL 2 Full keeps RDF's flexibility but is undecidable, so a reasoner may not terminate on some queries. The OWL 2 profiles restrict the language to keep reasoning tractable for particular purposes.

The profiles are the practical answer to the trade-off: each is designed for a different balance of expressive power and reasoning cost, giving up some of what full OWL can say in exchange for guarantees about how expensive reasoning will be. The engineering consequence is plain — the more an ontology can say, the harder it is to answer questions over it, so choosing a profile is choosing how much inference you will pay for. See OWL for the standard and its tooling.

When an ontology is worth its cost

It is worth building when the domain has rules a machine must check or apply, when several sources must be merged under one set of meanings, or when questions involve combination rather than lookup. Regulated domains, engineering models, scientific data and cross-organisation exchange are the familiar cases.

It is not worth building when the work is retrieval. If the requirement is to find the right document or record, a taxonomy with disciplined naming, a stable identifier per concept and a complete alias set does more per unit of effort, and it creates no modelling obligation to maintain for years. Formal models decay faster than vocabularies when nobody owns them, because a wrong axiom is worse than a missing label.

Entity resolution is usually what forces the decision. Two sources name the same thing differently, or one name covers two things. An ontology gives the resolution somewhere to live — one identifier, with the other names kept as labels and mapping statements — but the resolution itself is a judgement a person makes and the model records. Matching tools propose candidates; they cannot decide that two records describe the same real-world thing. That decision, made once and written down, is most of what an ontology contributes to merging.