Ontology
In computer science and artificial intelligence, an ontology is an explicit, formal specification of the concepts, relationships, and constraints that exist in a domain, written so that people and software can share a common vocabulary and so that machines can draw consistent inferences from it. Where a database schema serves one application, an ontology aims to capture the meaning of the terms independently of any single program, so that different systems can agree on what "gene," "patient," or "invoice" denotes. The word is borrowed from philosophy, but the engineering usage is deliberately modest: it is not a claim about what ultimately exists, only a record of what a particular community has agreed to talk about.
The canonical definition comes from Thomas Gruber's 1993 paper "A translation approach to portable ontology specifications," which called an ontology an explicit specification of a conceptualization. A conceptualization is the set of objects, concepts, and relations that a modeler assumes to exist in a domain, and the specification is the written vocabulary and axioms that pin those terms down. Later authors sharpened the phrase: Willem Borst's 1997 thesis added that the conceptualization is shared, and Rudi Studer, Richard Benjamins, and Dieter Fensel's 1998 paper settled on a "formal, explicit specification of a shared conceptualization." The three adjectives each do work. Formal means machine-processable, explicit means nothing is left implicit in a modeler's head, and shared means the point is agreement across people and tools.
The philosophical sense is older and different: ontology in Aristotle's Metaphysics and in W. V. O. Quine's 1948 essay "On what there is" asks what kinds of things exist and what it means for something to exist. The AI sense borrows the discipline's tools, such as the distinction between a thing and its properties or between a process and an object, but swaps the aim. A philosopher argues about whether numbers or events are real; a knowledge engineer decides whether the model should treat a "meeting" as an event with participants and a time. Upper ontologies like the Basic Formal Ontology and DOLCE sit in between, offering carefully reasoned top-level categories that domain ontologies can reuse so that independently built vocabularies line up.
The Web Ontology Language, OWL, became the standard way to publish ontologies after the W3C made it a recommendation in 2004, with OWL 2 following in 2009. It builds on description logics so that classes, properties, and individuals can be given axioms a reasoner can check, and it serializes to RDF so that ontologies are themselves web data. Real examples show the purpose. The Gene Ontology, started around 2000, supplies shared terms for gene products' molecular functions, biological processes, and cellular components so that findings in different organisms can be compared. SNOMED CT does the same for clinical terminology, and schema.org, launched in 2011 by Google, Microsoft, Yahoo, and Yandex, applies a lightweight ontology to web page markup. The Protégé editor from Stanford made building such models accessible to non-programmers.
Several pitfalls recur in practice. An ontology is often confused with a taxonomy, but a taxonomy is only a hierarchy of broader and narrower terms, while an ontology also carries relations and logical constraints. It is also not a statement of truth: it encodes one community's modeling decisions, and two good ontologies of the same domain can disagree. Perhaps the most common failure is over-engineering, where an elaborate model is built with no application consuming the inferences. Open-world reasoning surprises newcomers too, because OWL does not assume that something not stated is false, which is the opposite of how a database treats missing rows.
Ontologies matter again in the era of large language models, because a model that generates fluent text still lacks a controlled vocabulary against which its claims can be checked. Ontology-guided retrieval, schema-constrained extraction of entities from documents, and knowledge graphs layered over enterprise data all depend on stable, agreed definitions of terms. The tradeoff is the same as ever: shared formal meaning requires human effort to build and maintain, and the payoff appears only when many parties actually adopt it.
Read ontology as a negotiated, machine-readable dictionary of what a community means by its words: the Gruber definition makes agreement the point, and the contrast with philosophy is that the engineer is cataloguing a chosen conceptualization to make systems interoperate, not ruling on what the world is made of.