In computer science and knowledge representation, an ontology is a formal, explicit specification of a shared conceptualization of a domain. It defines the types of entities (classes), their attributes (properties), and the relationships between them, enabling machines to reason about and exchange information with a common understanding. Ontologies are foundational to the Semantic Web, artificial intelligence, and data integration, providing a structured vocabulary that can be reused and extended across applications.

1.1 Formal semantics

An ontology is equipped with a formal semantics, meaning that all terms and relationships are defined in a mathematically precise way. This allows automated reasoners to derive new knowledge that is logically entailed by the ontology’s axioms. Formal semantics typically rely on description logics or first-order logic, ensuring that interpretations are unambiguous and that inferences are sound and complete relative to the chosen logic.

1.2 Components of an ontology

An ontology consists of several core building blocks that collectively define the structure of a domain.

1.2.1 Classes (concepts)

Classes represent sets of entities that share common characteristics. For example, in a biology ontology, "Mammal" and "Bird" would be classes. Classes are often organized in a hierarchy (taxonomy) using subclass relationships, e.g., "Dog" is a subclass of "Mammal."

1.2.2 Individuals (instances)

Individuals (also called instances) are the concrete objects or entities that belong to a class. For instance, "Fido" could be an individual of the class "Dog." Individuals are the ground-level elements about which facts are stated in the ontology.

1.2.3 Properties (roles, attributes)

Properties describe attributes of classes and individuals or relationships between them. Object properties link two individuals (e.g., "hasParent"), while datatype properties link an individual to a data value (e.g., "hasAge"). Properties can have domain and range restrictions.

1.2.4 Relations and axioms

Relations define how classes and properties interact beyond simple hierarchies. Axioms are formal statements that constrain the interpretation of the ontology. Examples include domain and range axioms, disjointness between classes, and cardinality restrictions. Axioms enable reasoning by imposing logical conditions.

1.3 Types of ontologies

Ontologies can be classified by their scope and purpose.

1.3.1 Upper (foundational) ontologies

Upper ontologies (e.g., SUMO, DOLCE, BFO) provide a high-level, domain-independent framework of general concepts such as time, space, objects, and events. They serve as a top-level structure that can be specialized by domain ontologies, promoting interoperability across different fields.

1.3.2 Domain ontologies

Domain ontologies cover a specific subject area, such as medicine (e.g., SNOMED CT), genomics (Gene Ontology), or geography. They reuse terms from upper ontologies where possible and add domain-specific classes, properties, and axioms.

1.3.3 Task ontologies

Task ontologies describe concepts and relationships relevant to a particular type of problem-solving or procedure, such as diagnosis, planning, or scheduling. They focus on the activities and roles involved in a task rather than static domain entities.

1.3.4 Application ontologies

Application ontologies are tailored to a specific software system or user requirement. They often extend or specialize domain and task ontologies to meet the needs of a particular application, sometimes sacrificing generality for efficiency or usability.

2.1 Philosophical roots in metaphysics

The term “ontology” originates from philosophy, where it refers to the study of being and existence. Metaphysicians such as Aristotle and later medieval scholastics classified categories of entities and their relations. These philosophical inquiries laid the groundwork for later computational approaches by emphasizing formal definitions and taxonomies.

2.2 Rise in artificial intelligence

In the mid-20th century, AI researchers recognized the need for formal representations of domain knowledge to enable intelligent reasoning.

2.2.1 Early AI knowledge bases

Early knowledge-based systems, such as MYCIN for medical diagnosis and the DENDRAL project in chemistry, used rule-based representations. These systems foreshadowed ontologies by encoding domain knowledge in structured forms, though they lacked explicit formal semantics and shared vocabularies.

2.2.2 Cyc and WordNet

Cyc (started in 1984) aimed to encode commonsense knowledge in a massive formal ontology. WordNet (launched in 1985) provided a lexical database organized by semantic relations. Both projects demonstrated the feasibility and value of large-scale ontological resources, influencing later standards.

2.3 Ontologies in the Semantic Web

The Semantic Web vision, articulated by Tim Berners-Lee in the late 1990s, placed ontologies at its core.

2.3.1 W3C standards and vision

The World Wide Web Consortium (W3C) developed standards like RDF, RDFS, and OWL to formalize ontologies on the web. The goal was to make data machine-readable and interoperable across the web, enabling intelligent agents to discover, combine, and reason over information.

2.3.2 Linked data movement

Linked data principles encourage publishing structured data on the web using URIs and RDF, often drawing on ontologies to define vocabularies. Initiatives such as DBpedia and Linked Open Data (LOD) cloud grew from this movement, demonstrating large-scale use of ontologies for interlinking datasets.

3.1 Resource Description Framework (RDF) and RDF Schema

3.1.1 Triple-based data model

RDF represents information as triples: subject, predicate, object. Each triple states a fact about a resource. This graph-based model is flexible and can be combined with any ontology that defines the resources and predicates.

3.1.2 RDFS for vocabulary description

RDF Schema (RDFS) extends RDF with primitives for defining classes, subclasses, properties, and domain/range constraints. RDFS provides a basic ontological vocabulary but limited reasoning capabilities.

3.2 Web Ontology Language (OWL)

OWL is a more expressive ontology language built on RDF. It adds constructs such as equivalence, disjointness, cardinality restrictions, and property characteristics (e.g., transitivity, symmetry).

3.2.1 OWL profiles (EL, RL, QL)

To balance expressivity and computational complexity, OWL 2 defines three profiles:

  • OWL EL for ontologies with large class hierarchies (e.g., in biomedicine).
  • OWL RL for ontologies that can be efficiently reasoned about using rule-based engines.
  • OWL QL for query answering over large data volumes using relational database technology.

3.2.2 Sub-language expressivity

OWL is divided into OWL DL (Description Logic) for maximal expressivity while remaining decidable, OWL Full for unlimited expressivity but no computational guarantees, and the OWL 2 profiles. Users choose a sub-language based on their reasoning and scalability needs.

3.3 Other languages (DAML+OIL, KIF, Common Logic)

DAML+OIL was a precursor to OWL, combining earlier efforts from the DARPA Agent Markup Language and the Ontology Inference Layer. Knowledge Interchange Format (KIF) is a language for exchanging ontologies in first-order logic. Common Logic (ISO 24707) provides a framework for sharing logical theories across different syntaxes. These languages influenced the development of W3C standards.

4.1 Design principles

4.1.1 Clarity and coherence

An ontology should clearly define each term using natural language and formal axioms. The definitions must be logically coherent, meaning they do not lead to contradictions. Consistent naming conventions and documentation support clarity.

4.1.2 Extensibility and minimal encoding bias

A well-designed ontology anticipates future extensions without requiring restructuring. Minimal encoding bias means the ontology avoids forcing a particular implementation or data structure, allowing reuse across different systems.

4.2 Development life cycle

4.2.1 Specification and knowledge acquisition

The first phase identifies the ontology’s purpose, scope, and intended users. Knowledge acquisition involves gathering domain information from experts, literature, and existing datasets.

4.2.2 Conceptualization and formalization

Domain concepts are organized into a class hierarchy, properties are defined, and relationships are axiomatized. The conceptual model is then translated into a formal ontology language (e.g., OWL), with precise logical axioms.

4.2.3 Implementation and maintenance

The ontology is coded in a serialization format (e.g., RDF/XML, Turtle) and made available. Maintenance includes versioning, updating concepts as the domain evolves, and fixing errors. Documentation and community feedback are essential for long-term usability.

4.3 Tools and environments

4.3.1 Protégé

Protégé is a free, open-source ontology editor developed by Stanford University. It supports OWL, RDFS, and various reasoners. Its graphical interface, plugins, and API make it the most widely used ontology development tool.

4.3.2 TopBraid and others

TopBraid Composer is a commercial IDE for ontology development with strong integration for enterprise data and rule systems. Other tools include WebProtégé (online editor), OntoEdit, and Neon Toolkit. Each offers features for editing, reasoning, and collaboration.

4.4 Evaluation and quality assurance

4.4.1 Correctness and completeness

Correctness checks whether the ontology accurately represents the domain according to expert knowledge. Completeness assesses whether all necessary concepts, relationships, and axioms are included, though complete coverage is often unattainable.

4.4.2 Consistency and scalability checks

Consistency verification uses a reasoner to detect logical contradictions, unsatisfiable classes, or redundant axioms. Scalability tests measure how the ontology performs with large instance data, especially for query answering and reasoning tasks.

Ontologies enhance search engines by enabling semantic query expansion, concept-based indexing, and faceted search. They help disambiguate terms (e.g., “bank” as river bank vs. financial institution) and retrieve documents that match the user’s conceptual intent rather than exact keywords.

5.2 Natural language processing

In NLP, ontologies provide lexical and semantic resources. WordNet, for instance, is used for word sense disambiguation, named entity recognition, and relation extraction. Domain ontologies improve machine translation and text classification by adding world knowledge.

5.3 Biomedical and life sciences (e.g., Gene Ontology, SNOMED CT)

Biomedicine is a leading adopter of ontologies. The Gene Ontology (GO) provides a controlled vocabulary for gene functions, cellular components, and biological processes. SNOMED CT is a comprehensive clinical terminology. These ontologies enable data integration across experiments, hospitals, and databases, supporting diagnosis and research.

5.4 E-commerce and product catalogs

E-commerce platforms use ontologies to standardize product descriptions, attributes, and categories. Taxonomies like the UNSPSC (United Nations Standard Products and Services Code) and company-specific ontologies allow cross-catalog searching, recommendation, and supply chain interoperability.

5.5 Robotics and autonomous systems

Robotics ontologies, such as the IEEE Ontologies for Robotics and Automation (ORA), define basic concepts like actions, objects, and environments. They enable robots to share knowledge about task models, sensor data, and interactions, facilitating collaboration and adaptability.

6.1 Ontology matching and alignment

When different ontologies cover overlapping domains, matching algorithms find correspondences between their classes, properties, and individuals. Alignment is necessary for data integration, ontology merging, and agent communication. State-of-the-art techniques use string similarity, structure graphs, and machine learning.

6.2 Ontology evolution and versioning

Ontologies change as domains evolve. Evolution must manage versioning, deprecation of terms, and backward compatibility. Tools and protocols (e.g., OWL 2 versioning) help track changes and ensure that dependent systems can adapt without breaking.

6.3 Scalability and performance

Large ontologies (millions of axioms) and massive instance datasets pose challenges for reasoning and query execution. Partitioning, approximate reasoning, and database-backed storage are used to achieve practical performance while maintaining correctness when possible.

6.4 Reasoning and inference

Automated reasoning derives implicit knowledge from explicit axioms. Description logic reasoners (e.g., Pellet, HermiT, RDFox) support classification, consistency checking, and realization. However, expressivity must be balanced against computational feasibility; some OWL profiles are prioritized for tractable reasoning.

6.5 Lightweight and community-driven ontologies (e.g., schema.org)

Schema.org provides a lightweight, shared vocabulary for structured data on the web. It is maintained by a community of major search engines and web developers. Unlike formal OWL ontologies, schema.org emphasizes simplicity, extensibility, and adoption, often with less strict logical axiomatization.

6.6 Ontology learning from text and data

Ontology learning aims to automatically extract ontological structures (classes, relations) from natural language corpora or structured datasets. Techniques include clustering, pattern-based extraction, and deep learning. While fully automatic construction remains challenging, learned ontologies can bootstrap manual engineering.

7.1 Integration with deep learning and large language models

Deep learning models, particularly large language models (LLMs) like GPT and BERT, can complement ontologies. Neural networks excel at pattern recognition but lack explicit symbolic knowledge. Hybrid approaches use ontologies to constrain or interpret neural outputs, or employ LLMs to populate and refine ontologies from unstructured text.

7.2 Ontologies for the Internet of Things

The Internet of Things (IoT) generates vast heterogeneous data from sensors and devices. Ontologies provide a semantic layer for representing devices, their capabilities, observations, and contexts. Standardized IoT ontologies (e.g., SOSA/SSN) enable interoperability across platforms and facilitate automated reasoning for smart cities, industrial automation, and environmental monitoring.

7.3 Ethics and bias in ontological representation

Ontologies encode choices about what entities matter and how they are categorized. These choices can reflect or reinforce biases (e.g., gender or cultural assumptions). Future work will emphasize transparent design, participatory modeling, and auditing tools to detect and mitigate unintended bias. Ethical guidelines and inclusive representation will become integral to ontology engineering.