IM Engineer Study Notes
← Back to list
AI & Data
#온톨로지#시맨틱웹#OWL#지식표현#추론
Last updated · 2026-10-18

Ontology

1. Overview

An Ontology is an explicit, formal specification that defines the concepts (Concepts) that exist in a particular domain, the relationships among those concepts, and the constraints and rules they must obey, so that humans and machines can share them. T. Gruber defined it as "an explicit specification of a conceptualization," meaning a common perspective on some world committed to in a machine-processable form without ambiguity.

As information systems began to be linked across the boundaries of organizations and institutions, the problem of using the same term with different meanings (homonymy) or different terms with the same meaning (synonymy) emerged as a fundamental obstacle to data integration. For example, people had to interpret and map, one by one, whether a 'customer' in one system and a 'client' in another referred to the same entity, a process that was costly and error-prone. Because a relational schema defines the structure of data but cannot carry its semantics, the demand for machines to understand and reason over meaning accumulated. The ontology emerged precisely from this gap, from the need to pin down "what the data means" in formal logic so that machines can interpret, reason over, and verify it.

The direct catalyst was the Semantic Web vision proposed by Tim Berners-Lee in the early 2000s. Within the idea of making web documents 'understandable' not only to humans but also to machines, so as to realize automatic querying, integration, and reasoning, the ontology took its place as the core backbone of that meaning layer. Recently, as it has been revisited as a Neuro-Symbolic approach that suppresses the hallucinations of large language models (LLMs) and provides grounding for reasoning, and as the schema of knowledge graphs, the ontology is drawing attention again as foundational knowledge for explainable AI, beyond a mere classification scheme.

Desirable characteristics an ontology should have are commonly cited as Clarity (removing ambiguity in term definitions), Coherence (absence of contradiction among axioms), Extensibility (adding concepts without breaking existing definitions), and Minimal Ontological Commitment (not asserting more about the world than necessary). These are design principles put forward by Gruber, conditions for an ontology to become a knowledge asset that multiple systems can share and reuse without being bound to a particular application. In other words, the purpose of an ontology is, beyond one system's data model, to establish a 'common language' of meaning agreed upon across the whole organization and domain.

2. Components and Overall Structure of an Ontology

An ontology consists broadly of classes (Class, concepts), instances (Instance, individuals), properties (Property, relationships and characteristics), and the axioms (Axiom) that impose logical constraints on them. A class is a concept abstracting a set of individuals with the same properties, such as 'person' or 'general'; an instance is a concrete individual that actually belongs to that class, such as 'Yi Sun-sin'. A property is a relationship linking an individual to another individual (ObjectProperty, e.g., commands) or a data characteristic that assigns a value to an individual (DatatypeProperty, e.g., year of birth); an axiom declares constraints that must necessarily be true, such as "every general is a person" or "a person has only one birth year," serving as the basis for inference and consistency checking.

flowchart TB
  O["Ontology"] --> C["Class/Concept(Class)"]
  O --> P["Property"]
  O --> A["Axiom/Constraint(Axiom)"]
  O --> I["Instance"]
  C -->|subClassOf| C2["Super-/sub-concept hierarchy(Taxonomy)"]
  P --> OP["Object property(ObjectProperty)"]
  P --> DP["Data property(DatatypeProperty)"]
  A --> R["Rules/constraints(domain/range, cardinality)"]
  I -->|rdf:type| C

In the structure above, the class hierarchy (Taxonomy) is a backbone that organizes concepts into super- and sub-levels through the 'is-a' relationship, but unlike a classification scheme (Taxonomy) or a controlled vocabulary (Thesaurus), an ontology is intrinsically far more expressive in that it also carries arbitrary semantic relationships and formal-logic axioms. It is precisely this axiom layer that determines "whether facts not stored can still be logically derived," elevating the ontology from a mere term dictionary to a reasoning-capable knowledge model.

A. Classes and Hierarchy (Taxonomy)

A class is a set abstracting a domain's concepts, forming super- and sub-level hierarchies through the subClassOf (subsumption) relationship. If one defines a hierarchy such as 'person ⊃ soldier ⊃ general', then from the mere fact that 'Yi Sun-sin' is an instance of general, the fact that he is both a soldier and a person is automatically inherited and inferred. This is a core device that reduces redundant storage and guarantees data consistency.

The reason hierarchy design matters is that the inheritance structure is itself the backbone of inference. The more refined the class hierarchy, the more a query about a sub-concept automatically leverages the knowledge of its super-concept; for example, the query "retrieve the affiliation of all soldiers" covers at once even the instances of sub-classes such as generals and privates. Conversely, if the hierarchy is designed wrongly, inheritance instead propagates errors, so the principle is to examine strictly whether the 'is-a' relationship is true ('is a general necessarily a soldier?') and not to abuse it.

However, since hierarchy design reflects the worldview of domain experts, an excessively deep hierarchy makes maintenance difficult while an excessively shallow one lowers the utility of inference. In practice, one reuses upper-level concepts (Upper Ontology, e.g., BFO·DOLCE) to standardize the top of the hierarchy and newly designs only the lower domains, thereby securing both design quality and interoperability. A representative case is medical ontologies adopting BFO as their top to raise consistency among different medical ontologies.

B. Properties (Relationships and Characteristics) and Their Constraints

A property corresponds to the 'verb' that connects an individual to the world. An object property links individual to individual (commands), and a data property links individual to value (year of birth = 1545). Properties can be given domain and range constraints to restrict their use, as in "the subject of 'commands' is a person, the object is a warship," and logical characteristics such as symmetry, transitivity, and inverse relations can be declared.

The logical characteristics of a property are not mere metadata but inference rules that themselves create new facts. For example, if 'ancestor' is declared a transitive property, then when 'A's ancestor is B and B's ancestor is C', the fact 'A's ancestor is C' is automatically derived by the Reasoner. Declaring an inverse relation (inverseOf) automatically generates 'child' from 'parent', and the symmetric property 'spouse', entered in only one direction, implies the opposite direction. This ability to fill in facts not explicitly stored by means of rules is the principle by which an ontology maintains complete knowledge while saving storage space.

Thus, declaring the logical characteristics of properties precisely determines the inferential power of an ontology, and if constraints are weak, erroneous data is not filtered out and the reliability of knowledge collapses. Conversely, excessively strong constraints judge even legitimate data to be inconsistent and make operation difficult, so the balanced sense to examine a domain's actual rules and exceptions closely and tune the strength of axioms is a core competency of the knowledge engineer.

C. Axioms and Consistency

An axiom makes explicit, in formal logic, a constraint that must necessarily be true in the domain. Axioms such as "male and female are disjoint" and "every order has exactly one customer (cardinality)" become the basis of the Consistency Checking that detects contradictions upon data entry. If some individual is described as belonging to the male class and the female class at the same time, the reasoner reports, on the basis of the disjointness axiom, that the graph is inconsistent.

Another role of axioms is to classify individuals by inferring their membership. OWL can define classes with necessary and sufficient conditions (equivalentClass), so that, for example, with the axiom "a person who has a child is a parent," the reasoner automatically classifies 'an individual with a child relationship entered' into the parent class. Without the designer specifying the type one by one, individuals are assigned to the appropriate concept by rules alone, which becomes a powerful foundation for rule-based automatic tagging and categorization.

This is semantic-level verification far more expressive than the constraints of a relational DB (CHECK·FK), a means of guaranteeing data quality declaratively. However, as axioms multiply, both the cost of inference and the difficulty of design grow together, so the restraint of promoting only "rules that machines must reason over and verify" to axioms and leaving the rest as documentation governs maintainability.

D. Instances and the Knowledge Base (A-Box/T-Box)

An ontology is divided into a schema layer (T-Box, Terminology) that specifies concepts, relationships, and axioms, and a fact layer (A-Box, Assertion) that holds actual individual facts. The T-Box holds conceptual knowledge such as "a general is a person," and the A-Box holds individual facts such as "Yi Sun-sin is a general." This separation is important in design because the schema (T-Box) is reused relatively stably, whereas facts (A-Box) are constantly added and updated.

This distinction also carries meaning in that the reasoning services of description logic handle the two layers separately. T-Box reasoning checks subsumption, equivalence, and unsatisfiability among concepts to verify the soundness of the schema itself, and A-Box reasoning judges the membership of individual entities and the consistency of facts. In large systems, since the A-Box swells to billions of triples, it is common to design so as to perform heavy T-Box reasoning at design and deployment time and to process only lightweight query answering during operation, thereby distributing the cost of inference along the time axis.

A knowledge graph is often the result of accumulating the A-Box at large scale, and it can be understood as a relationship in which the ontology (T-Box) supplies meaning and inference rules on top of it. Therefore the task of "building a knowledge graph" must be approached by separating the work of gathering facts (A-Box) from the work of defining the rules to interpret and verify those facts (T-Box); neglecting the latter leaves one with a 'meaningless graph' in which data piles up but inference and verification are impossible.

3. The Semantic Web Stack and Ontology Representation Languages

For machines to process ontologies, a standardized representation language is needed. The W3C standardized the Semantic Web stack that layers data identification, structure, and meaning, and on top of it an ontology is described with RDFS and OWL.

flowchart TB
  subgraph STACK["Semantic Web layer structure"]
    direction TB
    L1["URI/IRI · characters(identification/encoding)"]
    L2["RDF(triple: subject-predicate-object)"]
    L3["RDFS(class/property hierarchy)"]
    L4["OWL(formal logic/constraints/reasoning)"]
    L5["SPARQL(query) · Reasoner(reasoning)"]
    L1 --> L2 --> L3 --> L4 --> L5
  end

RDF (Resource Description Framework) is a data model that reduces every fact to a subject-predicate-object triple and uniquely refers to individuals by global identifiers (URIs). RDFS (RDF Schema) is a lightweight vocabulary layer on top of it that defines class and property hierarchies (subClassOf·subPropertyOf) and domain·range. OWL (Web Ontology Language) is a full-fledged ontology language that, based on Description Logic, can express rich axioms such as disjointness, cardinality constraints, property transitivity, and logical union/intersection; established in 2004, it was revised to OWL 2 in 2009. OWL 2 provides profiles (EL·QL·RL) that differ in reasoning complexity and expressiveness, allowing one to choose computational characteristics suited to the use, such as large-scale medical terminologies (EL) or high-volume queries (QL).

Category RDFS OWL(OWL 2) Comparison perspective
Theoretical basis Simple schema Description logic(DL) Expressiveness/reasoning capability
Expressible constraints Hierarchy·domain/range Disjointness·cardinality·transitivity·equivalence Semantic precision
Reasoning depth Shallow(hierarchy inheritance) Deep(contradiction detection·classification) Level of knowledge verification
Computational complexity Low High(adjusted by profile) Performance trade-off

As the table shows, RDFS is light and fast but cannot express precise constraints such as "disjoint" or "exactly one," so deep reasoning is impossible. Conversely, OWL supports rich logic but at a greater reasoning cost, so a compromise through profiles between expressiveness and performance becomes the core of design. That is, the practical criterion of choice is not "what to express" but "what level of reasoning to afford at what cost."

4. Ontology Construction Methodology and Reasoning

An ontology is not completed in one pass but is built through an iterative lifecycle. Representatively, METHONTOLOGY presents the stages of specification→conceptualization→formalization→implementation→maintenance, and Noy and McGuinness's "Ontology Development 101" presents the practical procedure of determining domain and scope→reviewing reuse of existing ontologies→enumerating key terms→defining the class hierarchy→defining properties and constraints→creating instances. The common principles are reuse first and scope-setting based on Competency Questions. A competency question is to first define "what questions this ontology should be able to answer"; for example, the question "in what year did the naval battle in which a particular general participated occur?" becomes a touchstone for verifying the completeness of the design.

flowchart LR
  Q["Define competency questions(required scope)"] --> M["Review reuse of existing ontologies"]
  M --> T["Design classes/properties(T-Box)"]
  T --> AX["Declare axioms/constraints"]
  AX --> INST["Load instances(A-Box)"]
  INST --> RSN["Run the reasoner(Reasoner)"]
  RSN -->|inconsistency found| T
  RSN -->|classification/inference result| USE["Query/use(SPARQL)"]

Reasoning is the stage that realizes the value of an ontology. Running reasoners such as HermiT·Pellet·ELK in a tool like Protégé automatically classifies super-/sub-relationships not stated, detects facts that violate axioms as inconsistent, and returns answers including even entailed facts for a query. One premise to beware of is that an ontology follows the Open World Assumption (OWA). That is, since "a fact not stated is not false but merely not yet known," a designer accustomed to the closed-world assumption of a relational DB (if absent, false), expecting "it is a violation because it is not in the data," may find the reasoning result at odds with intuition. Failing to understand this difference, axiom design will not work as intended.

5. Comparison with Similar Concepts and the Relationship with Knowledge Graphs

An ontology is often confused with a classification scheme, a thesaurus, or a data model, but is clearly distinguished in expressiveness and purpose. A classification scheme (Taxonomy) handles only the 'is-a' hierarchy, and a thesaurus (Thesaurus) only relationships at the level of synonyms and related terms, whereas an ontology carries arbitrary semantic relationships and even formal-logic axioms to support machine reasoning. The reason this difference matters is that a classification scheme without axioms is used only for "query expansion" or so, but an ontology with axioms provides qualitatively different capabilities, namely "contradiction detection" and "entailment reasoning."

Category Classification scheme Thesaurus Ontology Key difference
Kinds of relations is-a synonym·related arbitrary semantic relations Expressive scope
Axioms/constraints None Weak Formal logic Reasoning capability
Machine reasoning Not possible Limited Possible Verification·entailment
Representative use Classification·navigation Query expansion Integration·reasoning·verification Depth of use

The relationship with the Knowledge Graph is especially often examined. If a knowledge graph is the data (A-Box) that accumulates individuals and relationships at large scale, the ontology corresponds to the semantic rules and schema (T-Box) that the data must follow. By analogy, the ontology is the 'grammar' and the knowledge graph is the 'set of sentences' written in that grammar. Therefore the two are not in competition but complementary, and Google, Amazon, and others back the quality and inferential power of their large-scale knowledge graphs with ontologies. In practice, only when an ontology is combined with a knowledge graph do "inference of facts not stored" and "automatic detection of contradictory data" become possible.

However, in practice the degree of their combination is adjusted because of the trade-off between expressiveness and scale. An industrial knowledge graph of billions of triples finds it hard to perform full OWL reasoning constantly, so it takes the compromise of keeping only core axioms in a lightweight ontology (RDFS+) and running heavy verification in batch. Conversely, domains where precision is vital, such as medicine and law, apply rich OWL axioms even at the sacrifice of scale. Thus "in what ratio to combine data (knowledge graph) and rules (ontology)" is a practical issue of architecture design, and the right answer varies with the domain's precision requirements and data scale.

6. In Depth: Latest Trends and Industry Application Cases

The latest trend in ontology stands out most in its combination with LLMs. LLMs flexibly generate vast knowledge but have the weakness of fabricating facts (hallucination), while ontologies are precise but hard to expand automatically. Hence a Neuro-Symbolic approach combining the two is rising: when an LLM extracts triple candidates from unstructured text, they are verified and refined with ontology axioms and loaded into the knowledge graph, and conversely, at query time, the ontology-based reasoning results are injected as the LLM's grounding (Context) to suppress hallucination. GraphRAG is a representative application, where the concept hierarchy provided by the ontology semantically expands and constrains the search scope, raising the factuality and explainability of answers.

Mature cases in industrial settings are also abundant. In the medical field, SNOMED CT is one of the world's largest ontologies, describing more than 350,000 clinical concepts in OWL EL, semantically integrating the diagnostic terms of different hospital EMRs to realize interoperability. In the life sciences, the Gene Ontology (GO) describes gene functions in standard concepts, enabling cross-analysis of research data worldwide, and in manufacturing and engineering, ontologies are used in industrial data standards (e.g., the asset administration shell, the semantic model of digital twins) to integrate heterogeneous equipment data. In finance, FIBO (Financial Industry Business Ontology) standardizes the meaning of financial products and transactions and is used for regulatory reporting and risk integration. What these cases share is that in a "large-scale, heterogeneous, long-term operation" environment, semantic-level standardization has dramatically reduced integration costs.

Likely exam directions include ① the relationship between ontology and knowledge graph (T-Box/A-Box), ② the Semantic Web stack and OWL reasoning, ③ ontology-based RAG (GraphRAG) for suppressing LLM hallucination, and ④ strategies for securing interoperability in public and medical data integration. When writing an answer, taking "what differs from a classification scheme (axioms·reasoning)" as the axis and clearly revealing the difference in expressiveness is a high-scoring point. In particular, explaining the open-world assumption and the behavior of the reasoner with examples reveals that one has internalized the concept beyond rote memorization, giving it high discriminating power.

7. Considerations and Implications

  • Trade-off between expressiveness and performance: OWL's rich axioms raise inferential power but sharply increase computational complexity. One needs a strategy to analyze the domain's reasoning requirements and appropriately choose OWL 2 profiles (EL·QL·RL) or an RDFS level, and to control cost through partial reasoning at query time in areas that do not require exhaustive reasoning.
  • Construction/maintenance cost and governance: An ontology requires collaboration between domain experts and knowledge engineers, and when an organization's worldview changes, the schema is shaken. One must reuse upper ontologies and open standards (SNOMED·FIBO, etc.) and establish ontology governance with change history, versioning, and approval procedures to secure long-term consistency.
  • Understanding the Open World Assumption (OWA): A designer accustomed to a closed world expects "if not in the data, false," but an ontology is not so. At points where integrity verification is needed, one must reflect the difference in assumptions in the design, such as running shape constraints (Shape Constraint) like SHACL in parallel to complement closed-world verification.
  • Redefining the role in the LLM era and related technologies: In a phase where generative AI partly replaces knowledge generation, the ontology's value is being redefined as a "trust layer that supplies the grounding and constraints of facts." A strategic perspective that positions the ontology as 'semantic infrastructure' in connection with knowledge graphs, GraphRAG, explainable AI, and data spaces (data-sovereignty-based sharing) is required of the professional engineer.
  • Interoperability and standards adoption: A closed ontology aiming only at internal optimization ultimately breeds yet another silo. Taking international domain standards as the baseline and designing URIs and namespaces carefully to secure the possibility of connection (Linked Data) with external knowledge from the outset governs long-term investment efficiency.
  • Quality evaluation and evolution management: An ontology is not a deliverable that ends once built but an asset that evolves with changes in the domain. A system that periodically evaluates quality with quantitative and qualitative indicators such as competency-question satisfaction, logical consistency, and concept coverage, and regression-tests the impact of changes on existing reasoning and queries to guarantee 'evolution without breakage', determines operational success or failure.

References

In one line: An ontology is a shared specification that makes a domain's concepts, relationships, and axioms explicit in machine-processable formal logic; providing description-logic reasoning on top of the RDF·RDFS·OWL Semantic Web stack, it enables data integration, consistency checking, and explainable AI, and is re-emerging as the schema (T-Box) of knowledge graphs and a trust layer that suppresses LLM hallucination.