An ontology in computer science is an explicit specification of the categories of entities in a domain, the relationships among them, and the constraints governing their interpretation. It supplies a structured vocabulary together with statements about what that vocabulary means. For example, an ontology of scientific research might distinguish people, organizations, instruments, and experiments; describe which people conduct which experiments; and specify that an instrument is a physical object rather than an organization. Ontologies range from relatively simple classifications to formal theories expressed in machine-processable logic.
Ontologies are important in knowledge representation and reasoning because they make assumptions explicit and reusable across applications. They support interoperability, data integration, semantic search, and automated inference. A shared ontology can help two systems recognize that differently named fields concern the same kind of entity, or detect that their apparently similar categories have incompatible meanings. An ontology normally describes a domain’s conceptual structure rather than merely storing particular records, although many ontology languages can represent both general definitions and individual facts.
The term comes from philosophy, where ontology concerns being and the categories of existence. Its computational use is narrower and more operational: the central question is how to specify a domain so that a community and its software can communicate reliably. Ontologies consequently combine conceptual analysis, formal semantics, and practical engineering. Their usefulness depends not only on logical precision but also on the quality of their definitions, their relationship to data, and the institutions that maintain them.
Intellectual origins and historical development
The philosophical background of computational ontology includes long-standing distinctions among objects, properties, relations, and events. Aristotle developed influential accounts of categories and predication, but modern computational ontologies are not simply implementations of Aristotelian classification. They also draw on mathematical logic, database modeling, terminology management, and attempts to represent everyday knowledge in artificial intelligence. The practical problem is to identify distinctions that are sufficiently explicit for computation while remaining useful to specialists who understand the domain.
During the 1960s and 1970s, semantic networks and frame-based systems represented knowledge through connected concepts, attributes, and inheritance. These approaches made it possible to organize information around concepts such as a person, an event, or a vehicle rather than around isolated symbols. They also exposed difficulties: inheritance could have exceptions, different systems interpreted links differently, and informal network diagrams did not necessarily determine precise consequences. Work on description logics, including the KL-ONE system described by Ronald Brachman and James Schmolze in 1985, helped provide a logical foundation for structured concept descriptions.
The rise of expert systems reinforced the need to separate reusable domain knowledge from application-specific procedures. A medical system and an equipment-maintenance system might use similar reasoning methods while relying on entirely different concepts. Conversely, several applications within one field might benefit from the same conceptual vocabulary. Projects such as Cyc, begun in 1984 under Douglas Lenat, explored the construction of large bodies of explicitly represented commonsense knowledge. Their ambition highlighted both the potential value of shared conceptual foundations and the considerable effort required to build them.
In the early 1990s, ontology became a central term in knowledge-sharing research. Thomas Gruber’s 1993 formulation, “an explicit specification of a conceptualization,” became particularly influential. A conceptualization is an abstract account of the entities and relationships relevant to a purpose; its specification makes that account communicable through defined terms and axioms. Later formulations emphasized that an ontology is formal and shared, drawing attention to machine interpretation and community agreement. These definitions describe an engineering role rather than requiring a single philosophical position about what ultimately exists.
The Semantic Web initiative extended these ideas to distributed information on the World Wide Web. World Wide Web Consortium standards provided common mechanisms for naming entities, publishing statements, and defining vocabularies. RDF became a W3C Recommendation in 1999; the first Web Ontology Language recommendations followed in 2004, and OWL 2 in 2009. This development connected ontology engineering to web identifiers and interoperable data exchange, while ontologies continued to be used independently in scientific databases, enterprise integration, manufacturing, and other specialized settings.
Components and conceptual organization
An ontology’s basic components usually include classes, individuals, properties, and axioms. A class describes a category, such as Researcher or Laboratory. An individual denotes a particular entity, such as a specific laboratory. A property represents a relationship or an attribute: worksAt might connect a researcher to an organization, while a measurementValue property might connect a measurement to a number. Terminology varies between languages; description logics commonly use concept for class and role for a binary relationship. Axioms state general conditions, such as that every laboratory is an organization or that no organization is a physical measurement.
Subclass relationships organize classes into hierarchies. If every spectrometer is an instrument, Spectrometer is a subclass of Instrument, and anything classified as a spectrometer must also be an instrument. Such a hierarchy need not be a tree: a class may have several superclasses. A research hospital, for example, might be both a healthcare organization and a research organization. Subclassing must be distinguished from other relationships. An engine is part of a vehicle, not a kind of vehicle; a scientist may belong to an institution without being a subclass of that institution. Confusing these relations produces misleading inferences.
Properties also require careful distinctions. An object property links individuals to individuals, whereas a datatype property links an individual to a literal value, such as a date or decimal number. A property may be transitive, symmetric, asymmetric, or related to another property by an inverse axiom, depending on the language. For example, an ancestor relation can be modeled as transitive, while a direct-parent relation ordinarily should not be. A property’s domain and range constrain the types of entities participating in it. In common ontology languages, these declarations support inference; they are not necessarily input-validation rules.
Definitions can state necessary conditions, sufficient conditions, or both. Saying that every laboratory employs at least one researcher gives a necessary condition for laboratory membership. Defining a research institution as exactly an organization that conducts research supplies a sufficient condition as well, provided the formal expression captures the intended meaning. This distinction matters for automated classification: a reasoner cannot normally infer membership from conditions that were stated only as necessary. Labels, comments, examples, and editorial notes supplement the formal axioms but do not automatically carry the same logical force.
Ontologies are often differentiated by scope. A domain ontology concerns a field such as chemistry or transport; a task ontology describes activities such as diagnosis or planning; an application ontology specializes concepts for a particular system. An upper ontology supplies highly general categories such as object, process, quality, and role. These distinctions concern intended use rather than mutually exclusive file types. A project may import an upper ontology, reuse several domain modules, and add application-specific terms. Its metadata should identify authorship, scope, version, licensing, and the intended meaning of its components.
Formal semantics and ontological commitments
A formal ontology acquires precise meaning through a semantics that specifies how its expressions are interpreted. In a typical model-theoretic account, an interpretation provides a nonempty domain of entities and assigns meanings to names. A class is interpreted as a set of domain elements, an individual name as a domain element, and a binary property as a set of ordered pairs. An interpretation is a model of the ontology when it satisfies all its axioms. This makes it possible to distinguish what the ontology explicitly states from what follows logically from those statements.
For example, consider the axioms
and
The first states that the class of spectrometers is included in the class of instruments; the second states that the entity named is a spectrometer. Together they entail . More generally, an ontology entails a statement , written , when every model of satisfies . Entailment therefore concerns all interpretations compatible with the axioms, not merely a particular stored graph or a programmer’s intended example.
Many web ontology systems use the open-world assumption: information absent from the knowledge base is not thereby false. If an ontology records one author of a paper, it does not follow that the paper has only one author. This differs from many conventional database queries and rule systems, which often treat missing facts as unavailable or false for the purpose of an operation. Ontology reasoning also commonly avoids a unique-name assumption. Two different identifiers may denote the same entity unless their difference is asserted or otherwise follows from the axioms. Accordingly, a maximum-cardinality constraint can sometimes entail identity rather than report a duplicate-record error.
Ontologies also express ontological commitments: decisions about the categories used to represent a domain. A system may treat a university as an organization distinct from its buildings, distinguish a person from the roles that person occupies, and model a repair as an event rather than a property of a machine. These choices influence how facts can be combined. Modeling employment as a simple binary relation is convenient, but a separate employment entity may be needed to record employer, employee, contract, location, and time. Such reification increases expressiveness at the cost of a more elaborate structure.
The treatment of change, context, and exceptions presents further choices. Classical logical entailment is monotonic: adding axioms does not retract conclusions already entailed, although it may expose inconsistency. Default claims such as “birds normally fly” therefore cannot be treated as universal subclass axioms without excluding exceptions. Nonmonotonic reasoning, temporal representations, and contextual models address related issues through additional machinery. An ontology’s formal precision is always relative to its language and modeling assumptions; it does not establish that its categories are uniquely correct descriptions of reality.
Representation languages and web standards
Ontology representation languages differ in expressive power, computational properties, and intended environment. Some support little more than named categories and subclass links, while others express quantified relationships, equality, disjointness, and complex definitions. A useful distinction separates an abstract language from its concrete serialization. The same logical ontology can be written in several notations, and superficially different files may express equivalent axioms. Conversely, files that use the same syntax may have very different semantics if their vocabularies and interpretation rules differ.
The Resource Description Framework represents information as triples consisting of a subject, predicate, and object. A triple might state that a particular instrument has a particular manufacturer. Resources are commonly named using an internationalized identifier based on the same naming architecture as a uniform resource identifier, while literals represent values such as strings and numbers. RDF Schema adds vocabulary for classes, subclasses, properties, domains, and ranges. These mechanisms support basic semantic organization but do not by themselves provide the full range of restrictions available in more expressive ontology languages.
The Web Ontology Language, usually called OWL, provides richer class expressions and property axioms. OWL 2 supports intersections and unions of classes, complements, existential and universal restrictions, cardinalities, inverse properties, and selected property-chain axioms. Under its Direct Semantics, OWL 2 DL is based on the description logic SROIQ, subject to structural restrictions needed for decidability. OWL 2 also has an RDF-Based Semantics that applies more broadly to RDF graphs. The two semantic frameworks are related, but arbitrary RDF graphs do not automatically meet the restrictions required for OWL 2 DL reasoning.
OWL 2 defines three profiles for particular computational needs. OWL 2 EL supports efficient reasoning over large class structures and is useful for ontologies containing many existential definitions. OWL 2 QL is designed to support ontology-mediated querying through rewriting over relational data. OWL 2 RL enables rule-oriented implementations suitable for many scalable data-processing tasks. Each profile restricts the permitted constructions; none is simply a smaller file format. Selection depends on the required inferences, data volume, and available reasoning infrastructure.
Other standards serve complementary purposes. SKOS, standardized in 2009, represents thesauri, subject headings, and concept schemes using relations such as broader, narrower, and related. A SKOS concept is not automatically equivalent to an OWL class. The Shapes Constraint Language, or SHACL, became a W3C Recommendation in 2017 and describes conditions for validating RDF data graphs. RDF can be serialized using Turtle, RDF/XML, or JSON-LD; these are representations rather than substitutes for ontology semantics. Outside the web stack, Common Logic provides a framework for first-order-style knowledge exchange, and other formalisms offer different balances between expressiveness, portability, and implementability.
Automated reasoning and computational behavior
Automated reasoning determines consequences of an ontology according to its formal semantics. Common tasks include checking consistency, testing whether a class is satisfiable, finding subclass relationships, and determining which individuals belong to which classes. These are related but distinct operations. An ontology can be consistent while containing an unsatisfiable class: a class defined as both a Person and an Organization is impossible if those categories are disjoint, but no contradiction arises unless the ontology also requires an instance of that class. If an individual is asserted to belong to the impossible class, the ontology becomes inconsistent.
Classification computes a class hierarchy, including relationships that were not explicitly asserted. Suppose a ResearchLaboratory is defined as a Laboratory that conducts some ResearchActivity, while every SpectroscopyLaboratory is a Laboratory and conducts some SpectroscopyExperiment. If SpectroscopyExperiment is a subclass of ResearchActivity, the reasoner can infer that SpectroscopyLaboratory is a subclass of ResearchLaboratory. The inference combines several definitions rather than matching a single label. This ability makes formal ontologies different from purely editorial hierarchies, in which the placement of each term must be maintained manually.
Different algorithms implement these tasks. Tableau and hypertableau methods test satisfiability by constructing a possible model, systematically expanding logical conditions and checking for clashes. Completion procedures efficiently derive consequences for restricted description logics such as those underlying OWL 2 EL. Rule-based systems may materialize inferred statements in advance, while query-rewriting systems transform a user query into a form answerable over existing data. The choice between materialization, rewriting, and reasoning during query evaluation depends on update frequency, query workload, and the expressiveness of the ontology.
Expressiveness has substantial computational costs. For unrestricted ontologies, reasoning in the description logic underlying OWL 2 DL has doubly exponential nondeterministic worst-case time complexity for standard satisfiability problems. This is a worst-case statement about combined input size, not a prediction that every practical ontology is intractable. Restricted profiles admit substantially better guarantees for relevant reasoning tasks. In practice, performance also depends on modeling patterns, numbers of individuals, property interactions, and incremental update strategies. Computational complexity therefore informs language design but does not replace empirical benchmarking.
Reasoning results must also be interpreted correctly. In classical logic, an inconsistent theory entails every statement, so useful systems commonly report inconsistency instead of returning arbitrary consequences. Explanation services can identify a minimal set of axioms sufficient for an entailment or contradiction, helping engineers locate a mistaken disjointness declaration or unintended class equivalence. Automated theorem proving establishes consequences of the supplied formalization; it does not independently verify that a domain statement is scientifically true. Likewise, existential entailment may require an unnamed entity without producing a persistent identifier or a complete record for it.
Ontology engineering, testing, and governance
Ontology engineering begins with purpose and scope rather than with a demand to model everything in a field. A team identifies users, source materials, intended applications, and questions the ontology must help answer. These are often expressed as competency questions, such as which instruments can perform a specified measurement or which datasets concern a particular process. Such questions guide decisions about granularity and formalization. An ontology intended for catalog navigation may require fewer logical commitments than one intended to detect incompatible component configurations.
Knowledge acquisition draws on domain experts, standards, technical documentation, datasets, and existing vocabularies. Reuse can reduce duplicated effort and improve interoperability, but importing an ontology also imports its assumptions and maintenance dependencies. Engineers must inspect definitions rather than rely on matching labels. A term named “sample” might denote a material object in one resource, a sampling event in another, and a statistical subset in a third. Tools such as Protégé provide editors for classes, properties, annotations, and logical expressions, often integrating reasoners that reveal unintended consequences during development.
Testing combines logical and nonlogical checks. A reasoner can identify inconsistency, unsatisfiable classes, and inferred hierarchies that differ from expectations. Test queries can verify competency questions against representative data. Structural checks can detect missing labels, broken references, undeclared entities, and departures from naming conventions. Human review remains necessary for whether a definition reflects accepted domain usage, whether examples are representative, and whether distinctions are useful. An ontology may be logically consistent while systematically misrepresenting its subject.
Publication requires stable identifiers and accessible documentation. Labels can change without changing the identity of a concept, whereas replacing an identifier may break downstream data and queries. Deprecation mechanisms allow obsolete terms to remain resolvable while directing users toward replacements. Version control records editorial changes, but semantic change requires additional care: altering a class from “vehicle designed for rail travel” to “vehicle operating on rails” can change membership even if its identifier and label remain unchanged. Release notes, version identifiers, and regression tests help dependent systems assess such changes.
Governance determines who can propose terms, approve definitions, resolve disputes, and authorize releases. Community ontologies often use editorial boards, issue trackers, contribution policies, and designated maintainers. Licensing affects whether software vendors and public institutions can redistribute or modify a vocabulary. Sustainability also depends on funding and succession: an extensively used ontology may become difficult to maintain when its original project ends. Ontology quality is therefore partly institutional. Formal elegance has limited practical value if identifiers disappear, definitions are undocumented, or revisions are introduced without coordination with users and data providers.
Data integration, schemas, and ontology alignment
An ontology and a database schema both describe structure, but they usually emphasize different concerns. A schema specifies how a particular data system organizes and constrains records, including tables, columns, keys, and permissible values. An ontology describes categories and relationships that may apply across several systems. The distinction is not absolute: schemas can encode substantial domain meaning, and ontologies can impose detailed structural restrictions. The important question is whether a representation primarily governs storage and validation, supports shared interpretation, or serves both functions.
In ontology-based data access, a conceptual vocabulary is placed above existing data sources. Mapping rules connect database queries or record patterns to ontology classes and properties. A user can then ask for research organizations without knowing which source stores universities in one table and laboratories in another. The system rewrites or otherwise evaluates the conceptual query against the sources. In a relational database, a join over institution and activity tables might supply instances of a formally defined research-organization class without requiring all source records to be copied into a separate ontology store.
Ontology alignment identifies correspondences between independently developed ontologies. These correspondences may state equivalence, subsumption, disjointness, or a more complex relationship. Lexical similarity, structural similarity, shared instances, and statistical methods can suggest candidate matches. Nevertheless, identical labels do not prove equivalence. A transport vocabulary may define a “station” as a physical site, while an administrative vocabulary uses the same word for an operating organization. Treating the classes as equivalent could propagate category errors throughout an integrated system.
Identity alignment is particularly consequential. In OWL, asserting that two identifiers denote the same individual permits properties of either to apply to the same entity. That is stronger than saying that two records concern related entities or describe the same topic. A city, its municipal government, its administrative territory, and its statistical reporting unit may need separate identifiers. Mappings must also preserve distinctions among units, time periods, and measurement procedures. Aligning two fields called “mass” is insufficient if one records kilograms, another grams, and one refers to dry material while the other includes water.
Ontologies do not eliminate the practical work of data quality management. Source records may contain misspellings, incomplete dates, duplicate identifiers, and contradictory measurements. Data transformation pipelines must normalize representations and preserve data lineage so that an integrated statement can be traced to its origin. Validation languages can test whether required properties are present, while ontology reasoning determines consequences under declared semantics. Keeping these functions distinct is especially useful in distributed settings, where incomplete information may be acceptable logically but unacceptable for a particular transaction, report, or regulatory submission.
Scientific, industrial, and information-retrieval applications
Life-science research provides prominent examples of large, collaboratively maintained ontologies. The Gene Ontology project began in 1998 and was described in a widely influential publication in 2000. It organizes concepts concerning molecular function, biological process, and cellular component. Annotations connect gene products with ontology terms and include evidence information. The ontology enables comparison across organisms and databases despite differences in naming. Its graph contains several carefully distinguished relationship types, so an ontology edge should not be interpreted as an interchangeable generic association.
Scientific use frequently combines a domain ontology with statistical analysis. In gene-set enrichment analysis, researchers test whether a selected set of genes is associated with particular ontology categories more than expected under a specified background model. The ontology supplies the organized categories and annotation relationships; the statistical procedure supplies the test. Results depend on annotation coverage, the selected background, dependencies among categories, and multiple-testing procedures. The existence of an ontology term or an inferred annotation is not, by itself, experimental evidence for a biological mechanism.
In manufacturing and engineering, ontologies can represent components, materials, processes, capabilities, and requirements. A production-planning system might distinguish what a machine is from which operations it can perform and under which operating conditions. This allows data from equipment catalogs, enterprise systems, and maintenance records to be interpreted together. Building and infrastructure information can similarly distinguish a physical asset, its digital description, its location, and events in its service history. Such distinctions matter when a component is replaced: the location may remain while the individual component and its maintenance record change.
Cultural heritage applications connect objects, people, places, activities, and documentary sources. The CIDOC Conceptual Reference Model, standardized as ISO 21127, uses an event-centered approach that can represent the production, acquisition, movement, and modification of objects. This supports integration across collections whose catalog structures differ. A painting’s creator, previous owner, current custodian, and depicted person are separate relationships, even when a local catalog places them in similarly labeled fields. Ontology-based modeling helps preserve those distinctions rather than flattening them into a single association.
Ontologies also support information retrieval and knowledge graphs. A search system may expand a request for instruments to include recognized subclasses or distinguish identical words with different domain meanings. A knowledge graph can combine individual entities and relationships with an ontology that specifies their types and expected interpretation, although not every graph uses a rich formal ontology. Wikidata, for example, uses item-and-property statements, qualifiers, and references within its own data model. Ontology-guided retrieval can improve access to structured information, but its effectiveness depends on mappings between user language, ontology terms, and the actual coverage of indexed data.
Limitations, contested choices, and connections to machine learning
No ontology is neutral in the sense of being independent of purpose, perspective, and abstraction. Choosing whether a river is a persistent object, a changing body of water, or a spatial feature affects how observations are connected across time. Social and administrative categories can be even more dependent on legal systems and institutional practice. Ontology design therefore requires explicit scope statements and attention to whose definitions are being represented. Formal consistency cannot resolve disagreements that originate in different purposes, evidence standards, or conceptual traditions.
General-purpose integration faces a trade-off between shared structure and local detail. A very general ontology may be reusable but too weak to answer specialized questions. A highly detailed ontology may fit one community well while imposing assumptions that others reject. Upper ontologies attempt to provide common foundations, but their distinctions can introduce substantial learning and mapping costs. Lightweight vocabularies may achieve broader adoption precisely because they make fewer commitments. The appropriate degree of formalization depends on whether the main objective is navigation, interoperability, validation, explanation, or deductive inference.
Incomplete and changing knowledge complicates maintenance. Scientific classifications evolve, organizations reorganize, and terms acquire new definitions. Classical ontologies often distinguish timeless axioms from time-dependent facts only through additional modeling conventions. Representing that an organization had a particular legal status during one interval requires more than an unqualified type assertion. Uncertainty raises a different issue: ordinary OWL axioms do not attach probabilities or confidence weights to propositions. Probabilistic extensions and hybrid systems can represent uncertainty, but they require semantics and computational methods beyond standard deterministic ontology reasoning.
Machine learning can assist ontology construction by extracting terms, proposing relationships, identifying duplicates, and suggesting alignments. Natural language processing systems can identify candidate entities and relations in documents, while large language models can draft definitions or translate between informal descriptions and structured expressions. These outputs remain candidates rather than validated axioms. A plausible sentence may conflate a necessary condition with a sufficient one, confuse a part with a subtype, or introduce an unsupported relationship. Review and logical testing are particularly important because a single strong axiom can have many downstream consequences.
The relationship also runs in the opposite direction. Ontologies can constrain extraction outputs, provide labels for learning tasks, support explanations, and structure retrieval for retrieval-augmented generation. In neuro-symbolic artificial intelligence, learned representations and symbolic reasoning are combined in several ways, including systems that use logical constraints during learning or reason over facts extracted by neural models. Vector similarity and logical entailment remain different notions: two concepts can have similar linguistic contexts without being equivalent, and an entailed fact need not have a nearby textual embedding. Hybrid systems must specify which component supplies evidence, which enforces constraints, and how errors or contradictions are handled.