{
  "id": "computational-ontology",
  "title": "Computational Ontology",
  "lang": "en",
  "source_lang": "en",
  "summary": "A computational ontology formally specifies the entities, relations, and constraints of a domain to support shared interpretation and automated reasoning.",
  "category": "computer-science",
  "subcategory": "artificial-intelligence",
  "type": "concept",
  "aliases": [
    "Computer ontology",
    "machine-processable ontology"
  ],
  "generated_at": "2026-11-04",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/computational-ontology/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/computational-ontology/",
      "json": "https://aiwiki.page/zh/computational-ontology.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "A computational ontology is an explicit, machine-processable account of the categories of entities in a domain, the relationships among them, and the constraints governing their interpretation. It enables software systems to represent not only data but also some of the meaning assigned to those data. In computer science, the expression largely overlaps with ontology, emphasizing formal representation, computational use, and the consequences that can be derived from a model.\n\nComputational ontologies are central to knowledge representation and reasoning, semantic data integration, and many forms of artificial intelligence. An ontology for manufacturing, for example, might distinguish machines, components, production processes, and maintenance events; define how these categories relate; and permit a reasoning system to identify equipment that satisfies a maintenance classification without that classification having been explicitly recorded.\n\nAn ontology is neither a complete description of reality nor necessarily a universal agreement about its structure. It is a representation developed for particular purposes under stated assumptions. Its effectiveness depends on the relationship between its intended meaning, its formal semantics, the quality of the information supplied to it, and the computational services used to interpret it."
    },
    {
      "heading": "Intellectual origins and historical development",
      "text": "The term ontology originated in philosophy, where it concerns what exists and the categories into which existing things fall. Philosophical distinctions involving substances, properties, relations, events, and universals influenced later computational modeling. However, a computational ontology has a practical additional requirement: its distinctions must be represented in a form that software can store, exchange, inspect, or use in inference. A philosophical account can remain partly discursive, whereas an executable knowledge model must make at least some of its commitments explicit.\n\nThe immediate technical background lies in twentieth-century symbolic artificial intelligence. Semantic networks represented concepts and relationships as nodes and edges, while frame systems organized knowledge around structured descriptions of objects and situations. Marvin Minsky presented frames as a framework for representing knowledge in 1975. The KL-ONE system, described by Ronald Brachman and James Schmolze in 1985, helped establish the idea that a representation language should have a precisely defined meaning and support automatic classification. These developments contributed to the emergence of description logics.\n\nDuring the 1980s and 1990s, knowledge-sharing research increasingly separated reusable domain models from the procedures of individual programs. Expert systems often embedded specialized knowledge in application-specific rules, making that knowledge difficult to transfer. Ontology engineering sought to identify the underlying vocabulary and conceptual structure so that different systems could interpret information consistently. The Cyc project, begun in 1984 under Douglas Lenat, pursued an extensive representation of common-sense knowledge. Other projects focused on narrower scientific or engineering domains.\n\nThomas Gruber’s 1993 formulation of an ontology as an “explicit specification of a conceptualization” became influential. Later formulations emphasized formality and sharing. A conceptualization in this context is an abstract account of a domain, not simply a list of words. Specification means that the relevant concepts, relationships, and assumptions are expressed sufficiently clearly to be examined and implemented. Sharing does not imply agreement by every possible user; it means that participating systems or communities have an agreed basis for interpreting the representation.\n\nThe development of the Semantic Web connected ontology research to the infrastructure of the World Wide Web. Work associated with Tim Berners-Lee promoted globally identified resources and machine-interpretable descriptions. The World Wide Web Consortium standardized RDF in 1999, the Web Ontology Language in 2004, and OWL 2 in 2009. These standards provided a common technical foundation without requiring every domain to adopt one universal ontology. Ontologies subsequently became components of scientific repositories, enterprise information systems, and interconnected knowledge services."
    },
    {
      "heading": "Conceptual structure and ontological commitments",
      "text": "A computational ontology usually contains classes, relations, and axioms, and may also contain individuals. A class denotes a category, such as Pump, ResearchArticle, or ChemicalReaction. An individual denotes a particular entity, such as a specific pump or publication. A relation connects entities, while a data property can connect an entity to a literal value, such as a temperature measurement or publication date. Axioms state conditions that must hold under the ontology’s formal interpretation.\n\nSubclass relations organize categories by inclusion. If every centrifugal pump is a pump, then CentrifugalPump is a subclass of Pump. This differs from an instance assertion: a particular installed pump is an instance of CentrifugalPump, not a subclass of it. Part–whole relations are different again. An impeller can be part of a pump without being a kind of pump. Confusing subclass, instance, and part–whole relationships produces models that may look plausible in a diagram but support invalid inferences.\n\nAn ontology’s ontological commitments are the distinctions and assumptions it adopts. A transport model may treat a vehicle as a persistent object and a journey as an event unfolding through time. A financial model may distinguish a legal organization from its trading name, registration record, and physical offices. These distinctions affect whether two records refer to the same entity, how changes are represented, and which relationships are permitted. They are not resolved merely by selecting a file format or drawing a hierarchy.\n\nOntologies vary in scope and depth. A domain ontology describes a field such as materials engineering. A task ontology describes activities such as planning or fault diagnosis. An application ontology combines selected domain and task concepts for a particular system. An upper or foundational ontology supplies broad categories intended for reuse across domains. Basic Formal Ontology, for example, distinguishes continuants, which persist through time, from occurrents, which unfold through time. DOLCE provides another foundational approach, with attention to categories underlying human conceptualization and language. Such frameworks offer disciplined modeling choices, but adopting one does not automatically resolve domain-specific disagreements.\n\nThe boundary between an ontology and neighboring artifacts is gradual. A glossary defines terms; a taxonomy organizes categories; a thesaurus records lexical and associative relationships; a database schema specifies the organization of stored data. An ontology typically adds explicit interpretations and logical constraints that support shared meaning or inference. A knowledge graph can contain instance-level facts organized by one or more ontologies, although some knowledge graphs have only a lightweight schema. Neither a graph structure nor an elaborate vocabulary alone guarantees a coherent ontological model."
    },
    {
      "heading": "Formal semantics and the meaning of axioms",
      "text": "Formal semantics explains what the expressions of an ontology mean independently of a particular software implementation. A common approach is model-theoretic: an interpretation assigns a domain of entities to the ontology and meanings to its names. Class names are interpreted as sets, individual names as domain elements, and binary relation names as sets of ordered pairs. An interpretation is a model of the ontology when it satisfies all the ontology’s axioms.\n\nFor example, the axiom that every surgeon is a physician can be written as\n\n\\[\\mathrm{Surgeon}\\sqsubseteq \\mathrm{Physician}.\\]\n\nIts meaning is that the set assigned to Surgeon must be included in the set assigned to Physician. If an individual named Lee is asserted to be a surgeon, every model of these statements also makes Lee a physician. The ontology therefore entails that classification. Entailment concerns what follows in every model, not what appears likely or what happens to be recorded in a particular data file.\n\nMany ontology languages draw on first-order logic or carefully restricted fragments of it. Description logics provide constructors for describing classes and relationships while controlling the computational properties of reasoning. A class of pumps having at least one motor as a part can be expressed as\n\n\\[\\mathrm{MotorizedPump}\\equiv \\mathrm{Pump}\\sqcap\\exists\\,\\mathrm{hasPart}.\\mathrm{Motor}.\\]\n\nThis equivalence gives both necessary and sufficient conditions: every motorized pump satisfies the description, and every pump satisfying it belongs to the defined class. A one-way subclass axiom would provide only a necessary condition.\n\nA crucial distinction is the open-world assumption used in OWL and related formalisms. Information not present in the ontology is not automatically false. If a pump is known to have some motor as a part, the system need not know a name for that motor. The existential requirement may be satisfied by an unnamed entity in a model. Likewise, two different individual names are not automatically interpreted as different entities. Explicit inequality or other axioms may be needed to establish that distinction.\n\nThese semantics differ from familiar database expectations. A cardinality restriction stating that a component has at most one manufacturer may cause two named manufacturers to be inferred identical, rather than immediately producing a data-entry error. A property’s domain and range generally generate classifications; they are not simply instructions to reject an input of the wrong type. Classical ontology reasoning is also normally monotonic: adding premises does not retract previous entailments. Defaults and exceptions require different treatments, including nonmonotonic reasoning. These distinctions explain why ontology axioms, database constraints, and operational business rules cannot always be substituted for one another."
    },
    {
      "heading": "Representation languages and web standards",
      "text": "The Resource Description Framework represents information as triples consisting of a subject, predicate, and object. A triple can assert that a particular component has a particular manufacturer or belongs to a class. Resources are commonly named with a uniform resource identifier, allowing identifiers to be shared across documents and systems. RDF supplies a graph data model rather than a complete ontology methodology. Its use does not by itself establish detailed definitions, sound modeling choices, or agreement between data publishers.\n\nRDF Schema adds basic vocabulary for classes, subclass relationships, properties, domains, and ranges. The Web Ontology Language, usually called OWL, adds richer constructs, including class equivalence, disjointness, property characteristics, cardinality restrictions, and property chains. OWL 2 has multiple concrete syntaxes, including Functional-Style Syntax and Manchester Syntax, and ontologies can also be serialized through RDF formats. Syntax determines how an ontology is written; semantics determines what its expressions mean. Two files with different surface forms can therefore express the same axioms.\n\nOWL 2 DL imposes structural restrictions that retain decidable reasoning under its Direct Semantics. Those restrictions matter particularly when expressive property axioms interact with cardinality constraints. OWL 2 also defines three profiles designed for different computational needs. EL supports large class hierarchies with extensive use of existential restrictions. QL is designed to support efficient query answering over large datasets through query rewriting. RL supports rule-oriented implementations. These are formally defined language fragments, not merely informal recommendations to write simpler ontologies.\n\nOntology use on the web also depends on surrounding standards. SPARQL provides query languages and protocols for RDF data; its results depend on the dataset and any entailment regime used by the system. SHACL expresses conditions against which RDF graphs can be validated, such as required values or permitted value counts. SKOS represents concept schemes, including thesauri and classification systems, with relationships such as broader, narrower, and preferred label. A SKOS concept hierarchy does not automatically have the same semantics as a hierarchy of OWL classes.\n\nOther representation choices serve different requirements. Common Logic, standardized as ISO/IEC 24707, supports a broader first-order logical framework. Logic programming and rule languages can express operational implications, but their handling of variables, negation, and recursion may differ from OWL. Serialization technologies such as XML and JSON-LD address representation and exchange rather than replacing the ontology’s conceptual content. Selecting a language therefore involves several independent questions: what must be stated, what must be inferred, how data will be exchanged, and which computational guarantees are required."
    },
    {
      "heading": "Automated reasoning and computational behavior",
      "text": "An ontology reasoner computes consequences of axioms according to a specified semantics. Important reasoning tasks include checking whether an ontology is consistent, determining whether a class can have instances, testing whether one class is subsumed by another, and classifying individuals. Classification can also mean computing the complete subclass hierarchy implied by the ontology, including relationships not explicitly asserted. These services make an ontology more than a static diagram: a formally defined category can acquire subclasses and members through inference.\n\nConsistency and class satisfiability are related but distinct. An ontology can contain a class that cannot have any instances without being globally inconsistent. For example, suppose Sensor and SoftwareProcess are declared disjoint, while HybridSensor is made a subclass of both. HybridSensor must then be empty. The ontology becomes inconsistent if it also asserts that some individual belongs to HybridSensor. Reasoners can reveal such conflicts, but they cannot determine whether the fault lies in the disjointness statement, a subclass assertion, or the intended domain interpretation.\n\nTableau and related procedures reason by attempting to construct models, expanding the implications of class descriptions while detecting contradictions. Completion-based methods are particularly effective for some less expressive languages, including EL-family logics. Rule-oriented approaches can derive consequences by repeated application of inference rules. Query-rewriting approaches transform a semantic query into a form that can be evaluated by a relational database. Different methods exploit different features of the representation and workload; there is no single implementation strategy optimal for every ontology.\n\nExpressivity has direct consequences for computational complexity. Highly expressive description logics can have very high worst-case reasoning complexity, even though many practical ontologies are manageable. Complexity can be measured with the ontology and data both variable, or with the ontology fixed and only the data growing. This distinction helps explain why an approach can support large datasets yet be expensive when the schema itself is complex. Structural restrictions, indexing, modularization, caching, and specialized algorithms all affect real performance.\n\nReasoning results also depend on implementation scope. A tool that materializes selected rules may not compute every consequence licensed by an expressive ontology language. Soundness means that reported consequences are valid under the intended semantics; completeness means that all relevant consequences can, in principle, be obtained. Resource limits or unsupported constructs can undermine expectations of complete reasoning. An ontology does not provide formal verification of the surrounding software merely because a reasoner reports it consistent. Correctness still depends on the translation of domain requirements, the supplied facts, the reasoning service’s guarantees, and the way an application uses its output."
    },
    {
      "heading": "Ontology engineering, testing, and maintenance",
      "text": "Ontology engineering begins with scope rather than with a maximal vocabulary. A project identifies the intended users, the kinds of entities to represent, the information available, and the questions the system must answer. Competency questions make these requirements explicit. A maintenance ontology might need to answer which machines contain components supplied by a particular manufacturer, or which inspections concern equipment operating during a specified interval. Questions of this kind expose missing relationships and clarify whether temporal, spatial, or provenance information is required.\n\nDevelopment usually combines domain expertise with formal modeling expertise. Domain specialists explain distinctions that may be absent from existing datasets; ontology engineers identify the logical consequences of encoding them. Terms require definitions, examples, counterexamples, and decisions about borderline cases. Existing ontologies may be reused through imports, selected modules, or carefully documented mappings. Reuse can improve compatibility, but an imported ontology also brings assumptions and dependencies. An ontology about physical components cannot be adopted safely merely because its labels resemble a local parts catalog.\n\nRecurring modeling problems can be addressed through ontology design patterns. A direct binary relationship is sufficient for “component installedIn machine,” but an installation event may also have a date, installer, location, and inspection status. Representing that event as an entity allows these participants and attributes to be attached coherently. Measurements present a similar issue: a number alone does not identify the measured property, unit, object, time, or method. Such patterns reduce ambiguity, although their added structure must be justified by actual application requirements.\n\nTesting includes logical and nonlogical checks. A reasoner can detect inconsistency, unsatisfiable classes, and unexpected inferred classifications. Query tests can verify competency questions. Structural checks can identify missing definitions, malformed identifiers, cyclic editorial dependencies, or undocumented terms. Regression tests record expected entailments and expected non-entailments. This last distinction is important under open-world semantics: failure to entail a proposition is not the same as entailing its negation. Logical testing must therefore be supplemented by expert review and validation against representative data.\n\nOntologies also require lifecycle management. Stable identifiers separate entity identity from labels, which can change or be translated. Deprecation allows obsolete terms to remain interpretable without being recommended for new use. Release notes, dependency control, and version control document changes. A seemingly small edit—such as adding a disjointness axiom or changing a property domain—can alter many downstream inferences. Maintenance consequently resembles software engineering, but with an additional concern: compatibility involves the preservation of meaning, not just file formats or callable interfaces."
    },
    {
      "heading": "Data integration, mappings, and identity",
      "text": "One of the principal uses of computational ontologies is semantic interoperability: allowing systems with different local structures to exchange information with a sufficiently shared interpretation. Two databases may both contain a field called “customer” while applying different definitions. One may include organizations that requested a quotation; another may include only entities with completed purchases. Conversely, fields with different names may refer to the same kind of entity. Matching names and data types is therefore insufficient for reliable integration.\n\nAn ontology can provide a conceptual layer between source schemas and consuming applications. Mappings specify how source records correspond to classes, relationships, and values. In ontology-based data access, information can remain in its original databases while queries are formulated using the ontology vocabulary. The system combines ontology axioms with mappings to rewrite those queries into forms executable by the underlying database engines. This approach avoids requiring all information to be copied into one physical repository, but its correctness depends on mapping quality and source behavior.\n\nOntology alignment identifies correspondences between independently developed ontologies. A correspondence may assert equivalence, subsumption, or a more limited relationship. Lexical similarity, structural context, instance data, and domain knowledge can all contribute evidence. However, equivalence is a strong claim. A geographic “region” defined for statistical reporting is not necessarily equivalent to an administrative region with a similar name. Mapping errors can propagate through reasoning, producing incorrect classifications far beyond the originally matched pair.\n\nIdentity is especially consequential. In OWL, asserting that two names denote the same individual permits substitution of one for the other throughout the logical interpretation. This is stronger than saying that two records are similar, describe related entities, or concern the same broad topic. A company, its website, its brand, and the organization record maintained by a registry should not be merged merely because they share a label. Primary keys and local identifiers also carry assumptions about uniqueness that may not hold outside their original systems.\n\nProvenance and context help preserve these distinctions. Integrated information may need to record its source, observation time, transformation history, and authority. Data lineage explains how a result was produced, while named graph organization can separate sets of RDF assertions; neither automatically supplies a complete logic of trust or context. Differences in units, time zones, classification versions, and temporal validity may require explicit transformations. Ontology-mediated integration is therefore not simply schema unification. It is the controlled reconciliation of meanings, identities, and assumptions while preserving enough information to interpret the result."
    },
    {
      "heading": "Scientific, industrial, and information applications",
      "text": "The Gene Ontology, established in 1998, illustrates ontology use in scientific data integration. It organizes descriptions of molecular functions, biological processes, and cellular components. Researchers and curators associate gene products with ontology terms, accompanied by evidence and source information. Shared identifiers and relationships allow results from different organisms and experiments to be compared more systematically than would be possible with uncontrolled terminology. The ontology is not a catalog of genes, and an annotation is not simply an assertion that a gene name is synonymous with a biological process.\n\nScientific analysis can exploit the structure of such ontologies. Enrichment analysis tests whether categories occur more often in a selected set of gene products than expected under a specified statistical model. Relationships between categories affect how annotations are propagated and how results are interpreted. Multiple testing, the choice of background population, evidence quality, and annotation coverage remain statistical and methodological concerns. Formal organization improves consistency, but it does not remove uncertainty from experiments or make every derived scientific interpretation correct.\n\nIn health information systems, SNOMED CT provides a comprehensive clinical terminology with a formal concept model and description-logic-based relationships. Its structure supports the consistent representation and retrieval of clinical meanings across software systems. The relationship between a terminology, a patient record, and a clinical assertion nevertheless requires care. A concept for a disorder is not itself a statement that a particular patient currently has that disorder. Negation, historical context, uncertainty, and the subject of an observation may be represented elsewhere in the information model.\n\nIndustrial uses include product lifecycle management, equipment maintenance, supply-chain integration, and the description of engineering assets. Ontologies can distinguish a product design from an individual manufactured item, a planned operation from an executed event, and a nominal specification from an observed measurement. These distinctions support queries across otherwise separate engineering and operational systems. For example, a system may connect installed equipment, component types, manufacturer documents, and maintenance records while preserving the difference between a class of components and each serialized component in service.\n\nCultural collections and information services offer another important setting. CIDOC CRM, an ontology developed for cultural heritage information, emphasizes entities and events such as production, acquisition, and ownership, allowing institutional records to be connected without requiring identical catalog structures. In information retrieval, ontologies can support query expansion, concept-based navigation, and the reconciliation of synonyms. WordNet provides lexical-semantic relationships useful in language applications, although a lexical database and a domain ontology serve different purposes. Across these settings, benefits arise from explicit distinctions and reusable identifiers rather than from formalization alone."
    },
    {
      "heading": "Limits, uncertainty, and relations to machine learning",
      "text": "An ontology’s formal precision should not be mistaken for complete or uniquely correct knowledge. Many domains contain contested categories, overlapping purposes, or context-dependent boundaries. An organization’s classification of an employee, contractor, and external collaborator may depend on legal jurisdiction and reporting requirements. A geographic object may have different boundaries for navigation, administration, and ecological analysis. A formally consistent ontology can encode a poor account of such distinctions, while a useful local model may not transfer safely to another setting.\n\nTime and change are persistent difficulties. A machine can move between factories, an organization can merge with another, and a scientific classification can be revised. A relation represented without temporal qualification may incorrectly suggest a permanent fact. Modeling approaches include events, time-indexed assertions, temporal parts, or explicit states, but these introduce additional entities and reasoning requirements. The choice depends on whether the application must answer historical questions, distinguish planned from actual events, or reconstruct what was believed at a particular time.\n\nVagueness and uncertainty present related but different problems. Classical class membership is interpreted as true or false in each model, whereas terms such as “large” can have context-sensitive boundaries. Uncertain evidence concerns confidence in an assertion rather than necessarily vague meaning. Fuzzy and probabilistic extensions have been developed, but their semantics and inference mechanisms are not interchangeable with standard OWL reasoning. Attaching a numerical confidence value as metadata does not automatically cause a classical reasoner to perform probabilistic inference.\n\nMachine learning can assist ontology construction by proposing terms, relations, alignments, or classifications from texts and datasets. Natural language processing supports entity recognition and relation extraction, while a large language model can generate candidate definitions and axioms. These outputs still require checks against domain evidence, modeling requirements, and formal consequences. A syntactically valid axiom may be unsupported, overgeneralized, or inconsistent with the intended model. Conversely, an ontology’s existing vocabulary can constrain extraction and provide structured targets for learned systems.\n\nOntologies also participate in neuro-symbolic artificial intelligence and retrieval-augmented generation. A system may retrieve facts through ontology-aware queries, use class restrictions to assess candidate outputs, or combine learned predictions with symbolic rules. These arrangements can make selected assumptions and inference paths explicit, but they do not automatically eliminate hallucinated claims or guarantee correct interpretation of natural language. Evaluation must separately examine the ontology, the source data, the learned components, and their interfaces. Active technical problems include scalable reasoning over changing datasets, dependable cross-ontology mappings, contextual representation, and methods for preserving formal guarantees when symbolic and statistical components interact."
    }
  ],
  "links": [
    "computer-science",
    "ontology-computer-science",
    "knowledge-representation-and-reasoning",
    "artificial-intelligence",
    "philosophy",
    "symbolic-artificial-intelligence",
    "marvin-minsky",
    "expert-system",
    "world-wide-web",
    "tim-berners-lee",
    "world-wide-web-consortium",
    "database-schema",
    "knowledge-graph",
    "first-order-logic",
    "nonmonotonic-reasoning",
    "resource-description-framework",
    "uniform-resource-identifier",
    "logic-programming",
    "json-ld",
    "relational-database",
    "computational-complexity",
    "formal-verification",
    "version-control",
    "software-engineering",
    "interoperability",
    "primary-key",
    "data-lineage",
    "information-retrieval",
    "wordnet",
    "metadata",
    "machine-learning",
    "natural-language-processing",
    "large-language-model",
    "neuro-symbolic-artificial-intelligence",
    "retrieval-augmented-generation",
    "ai-hallucination"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}