🕸️ Ontology engineering

Concept · 25 key papers · all concepts

Ontology engineering is the discipline that studies how to design, build, evaluate, maintain and reuse ontologies. An ontology here is an explicit, formal model of a domain's concepts, relations and constraints that both people and software can interpret and reason over.

Also: ontological engineering, ontology development, ontology construction, ontology building, ontology design

What it is

An ontology, in the computing sense, is a shared and machine-readable description of what exists in some domain and how those things relate. Gruber (1993) defined it as an explicit specification of a conceptualization. Borst (1997) and Studer et al. (1998) refined this to a formal, explicit specification of a shared conceptualization, which is the usual starting point today. Ontology engineering covers the practical work behind such artefacts. It includes eliciting knowledge from experts, choosing a representation language, writing the model, checking it with reasoners, aligning it with other ontologies, and evolving it over time.

Most ontologies are built from three kinds of ingredient:

• Classes, or concepts, denote sets of things such as Disease, Verb or Stadium. They are arranged in an is-a (subclass) hierarchy.
• Properties, or relations, link individuals to each other or to data values, for example hasSymptom or partOf.
• Axioms are logical statements that constrain how classes and properties may be used. Examples are disjointness, cardinality, domain and range, and definitions by necessary and sufficient conditions.

Individuals, also called instances, populate the classes. A lightweight resource may stop at a taxonomy, while a heavyweight one carries rich axioms. Even carefully built lexical resources can sit near the lightweight end. The Arabic Ontology, for instance, is described by its author as closer to a taxonomy than a richly axiomatized model (Jarrar 2022).

The key idea

The core mechanism is to state domain knowledge as logical axioms. A reasoner can then derive facts that were never written down explicitly and detect contradictions. In description-logic notation, an ontology and a small worked example look like this:

O = ⟨C, P, I, A⟩

Pizza ⊑ Food
VegetarianPizza ≡ Pizza ⊓ ∀hasTopping.(Vegetable ⊔ Cheese)
Vegetable ⊓ Meat ⊑ ⊥
Cheese ⊓ Meat ⊑ ⊥
margherita : Pizza,  tomato : Vegetable,  mozzarella : Cheese
hasTopping(margherita, tomato),  hasTopping(margherita, mozzarella)
margherita : ∀hasTopping.{tomato, mozzarella}

In the first line, O is the ontology. C is its set of classes, P its set of properties, I its set of individuals and A its set of axioms.

The remaining lines use standard description-logic symbols:

• ⊑ means "is a subclass of".
• ≡ means "is defined as exactly".
• ⊓ is intersection ("and") and ⊔ is union ("or").
• ∀hasTopping.X means "every topping, if any, belongs to X".
• ⊥ is the empty class, so the two disjointness axioms say nothing can be both meat and a vegetable, or both meat and cheese.
• "margherita : Pizza" asserts that an individual belongs to a class.
• hasTopping(margherita, tomato) asserts that a relation holds between two individuals.
• {tomato, mozzarella} is the class whose only members are the listed individuals. The last line is therefore a closure axiom: margherita has no toppings besides these two.

Because the toppings are closed, a reasoner can classify margherita as a VegetarianPizza without being told. If someone later wants to add a ham topping, the closure axiom must be revised first. Once ham is typed as Meat, which is disjoint from both Vegetable and Cheese, the definition no longer applies. If the old closure were left in place, a reasoner would have to identify ham with tomato or mozzarella, because OWL makes no unique-name assumption. With ham typed as Meat, that would make the ontology inconsistent. Likewise, if ham were asserted as both Vegetable and Meat, the disjointness axiom would expose the error as an inconsistency.

On the Semantic Web, the same knowledge is serialised as RDF triples (subject, predicate, object), such as "margherita hasTopping tomato". RDF Schema and OWL supply vocabulary such as rdfs:subClassOf and owl:equivalentClass for the axioms. This triple layer lets ontologies be queried with SPARQL and checked by OWL reasoners. Recent work has even had a language model write OWL 2 ontologies for syllogisms and then used the HermiT reasoner to test validity, with limited success (Lopez-Ponce et al. 2026).

History and founding papers

Philosophical ontology is ancient. The engineering discipline, however, grew out of knowledge-based systems research in the 1990s, when Gruber (1993) and Guarino (1998) framed ontologies as reusable components for knowledge sharing.

Structured methodologies followed. METHONTOLOGY (Fernández-López, Gómez-Pérez and Juristo, 1997) treated ontology building as a software-like life cycle. Its activities include specification, knowledge acquisition, conceptualization, integration, implementation, evaluation and maintenance. The table below follows those activities:

Noy and McGuinness (2001) offered a widely taught step-by-step guide. The NeOn methodology (Suárez-Figueroa and colleagues, 2012) later added scenarios for reusing and re-engineering existing resources. One such scenario, converting non-ontological resources, has recently guided the transformation of Brazil's national health terminology tables into an RDF/OWL/SKOS graph (Pires et al. 2026).

Standardisation came through the W3C. RDF was released as a recommendation in 1999, OWL in 2004 and OWL 2 in 2009.

Upper, or foundational, ontologies provide domain-neutral top categories so that domain ontologies can interoperate. BFO (associated with Barry Smith) divides entities into continuants and occurrents. DOLCE (Guarino and colleagues) draws the parallel distinction between endurants and perdurants. One illustration of their value is the cleanup of lexical resources against them. Jarrar aligned the top levels of an Arabic wordnet with BFO, DOLCE and DOLCE+DnS. He argued that speaker intuition alone yields errors such as treating the Morning Star and Evening Star as two different stars, although both names refer to a single planet, Venus (Jarrar 2022).

Ontology learning from text, pioneered by Maedche and Staab (2001) and building on lexico-syntactic patterns from Hearst (1992), aimed to reduce the cost of manual modelling. Typical pipelines extract candidate terms statistically and then hand them to experts for validation. Al-Khalil, an OWL ontology of Arabic grammar attached to the GOLD linguistic ontology, used a repeated-segments extractor whose proposals an expert approved (Aliane et al. 2010). An Arabic infectious-disease ontology compared TF-IDF, C-value and YAKE term extraction on a corpus of 115 documents covering about 25 diseases. It also mapped lay terms gathered from roughly 100 people onto clinical concepts (Alsudias & Rayson 2020).

How it is used today

Linguistics and lexicography are a major field of application. Dictionaries and wordnets are published as RDF linked data.

• Dbnary turns Wiktionary editions into RDF lexical networks. Its 2012 version used an LMF-inspired model, and its English network then had about 2.1 million nodes (Sérasset 2012).
• A later version of Dbnary moved to the Lemon model (Sérasset 2015).
• The OntoLex-Lemon core is being redesigned to connect ontologies and dictionaries more closely (McCrae et al. 2026).
• Representing morphology in ontology lexicons remains challenging (Klimek et al. 2019).
• Treebanks and annotation formats are also being expressed in RDF (Mambrini et al. 2022; Chiarcos & Glaser 2020).
• The Research Constructicon models Construction Grammar with a modular RDF/OWL ontology that reuses OLiA and FOAF (Winckel et al. 2026).
• Azhary, a WordNet-style Arabic lexical ontology, holds 26,195 words in 13,328 synsets (Ishkewy et al. 2014).
• Cross-lingual mental-health ontologies for Indian languages aim to connect how patients express themselves with clinical understanding (Kandala et al. 2025).

Event ontologies support information extraction. FrameNet's frame relations helped build a partial upper-level Event Ontology (Palmer et al. 2014). Other work models implicit causal relations between events (Segers et al. 2018). When the SynSemClass event ontology was extended, machine suggestions from a fine-tuned classifier gave statistically significant gains. These held both for inter-annotator agreement and for agreement with an expert's gold sample (Straková et al. 2026).

Large, often community-built knowledge graphs act as broad ontologies. Their data quality has been compared across DBpedia, Freebase, OpenCyc, Wikidata and YAGO (Färber et al. 2018). Ways of reifying RDF statements have been tested on Wikidata (Hernández et al. 2015). Wikidata has also been proposed as an ontology for NLP (Spaulding et al. 2023).

Ontologies also serve as inputs and anchors for language technology:

• The WebNLG benchmark asks systems to verbalise sets of DBpedia RDF triples. Its data contains 25,298 data–text pairs from 15 DBpedia categories (Gardent et al. 2017).
• Semantic parsers can match their outputs to a target ontology on the fly (Kwiatkowski et al. 2013).
• OG-RAG grounds retrieval in a domain ontology. Across four LLMs it raised fact recall by about 55% and answer correctness by about 40% compared with RAG, RAPTOR and GraphRAG baselines (Sharma et al. 2025).
• Ontology-guided reverse reasoning over knowledge-graph labels reached a Hit@1 of 89.43 on WebQSP, meaning the top-ranked answer was correct for 89.43% of questions. This was state of the art among the compared systems, and Hit@1 was at least 25.43% higher than when the LLM answered alone (Liu et al. 2025).
• In the Brazilian health graph, about 700,000 of its more than 1.4 million triples were produced purely by OWL 2 RL inference (Pires et al. 2026).

Language models are now entering the engineering workflow itself. GPT-3.5 has been used to grow is-a hierarchies from a seed concept, with verification queries to curb hallucination (Funk et al. 2023). In ontology alignment, an LLM can serve as an oracle for the uncertain mappings flagged by LogMap. That combination ranked second overall in the OAEI 2025 bio-ml track (Lushnei et al. 2026).

Open questions

Several problems remain unresolved:

• Evaluating automatically built ontologies is hard when no gold standard exists. LLM-built hierarchies, for example, could be judged only by manual inspection. They followed everyday rather than scientific taxonomies and showed cultural bias (Funk et al. 2023).
• Automatic formalization is fragile. An LLM-to-OWL pipeline for judging syllogism validity reached only 48.69% accuracy. That is below the 50% expected from random guessing on this two-way task, and well behind a simple logistic-regression classifier at 72.25%. Translation and serialization errors held it back (Lopez-Ponce et al. 2026).
• Alignment with LLM oracles still falls clearly short of a perfect oracle. It also raises concerns about training-data leakage (Lushnei et al. 2026).
• Many under-resourced languages lack domain ontologies that capture informal vocabulary (Alsudias & Rayson 2020).
• Ontology-grounded retrieval cannot be applied in domains that have no well-defined ontology yet, so the cost of building one remains. Once an ontology exists, however, OG-RAG stayed stable even with about half of its attributes randomly removed (Sharma et al. 2025).
• How far upper-ontology rigour can scale to large, crowd-built resources remains an open design question. Even the carefully aligned Arabic Ontology investigated only about 1,300 concepts thoroughly, with about 11,000 more only partially validated (Jarrar 2022).

Key papers

The works this page cites, the most cited first; citations as counted by OpenAlex plus the papers of this library that cite the work, read 2026-10-23.

Related concepts

• Knowledge representation and reasoning — Knowledge representation and reasoning is the area of artificial intelligence that studies how facts, concepts and rules about a domain can…
• Lexicography — Lexicography is the discipline of compiling, describing and designing dictionaries and related lexical resources. It covers both the…
• Lexical semantics — Lexical semantics is the study of word meaning. It asks what words and multiword expressions denote, how a word's senses are organized, how…
• Named entity recognition — Named entity recognition locates stretches of text that refer to specific real-world things, such as people, organizations, places…

More papers

Topics: 🕸️ Knowledge graphs & ontologies

Shelves: 🤖 AI & machine learning

This page in Markdown → https://md.research.elamur.ai/free/en/concepts/ontology-engineering/