What is an Ontology?

The standard definition in information science comes from Tom Gruber, who in 1993 described an ontology as "an explicit specification of a conceptualization." In 1998, Rudi Studer and colleagues extended this to "a formal, explicit specification of a shared conceptualization." The latter is now the most widely cited, and each of its four terms states a requirement.

• Formal: processable by machines without relying on human interpretation
• Explicit: everything is written down, nothing is left to tacit understanding
• Specification: a definition that leaves no room for interpretation
• Shared: based on agreement among multiple parties, not the convenience of one

The fourth is the heaviest in practice. A definition that only works inside one company is not an ontology.

The word comes from philosophy. It is an early-17th-century Latin coinage (ontologia); the work of Jacob Lorhard in 1606 is regarded as one of its earliest appearances. The question itself is older: since Aristotle, this branch of metaphysics has asked what it means to exist and what kinds of things the world is made of.

Artificial-intelligence researchers borrowed the word in the 1980s, and the definitions above took hold in the 1990s. To have a computer deal with a world, one has to write down in advance what kinds of things exist in that world and how they relate. What philosophy asked as inquiry, information science implemented as design.

Ontologies are often confused with glossaries and taxonomies. In fact there is a ladder of expressive power, built up from the bottom.

1. 1Glossary: a list of terms and their explanations. No relationships
2. 2Taxonomy: a hierarchy of broader and narrower terms, such as product categories or biological classification
3. 3Thesaurus: a hierarchy plus relationships such as synonymy and association
4. 4Ontology: classes, attributes, relationships and constraints, defined formally. Because rules such as "every A has exactly one B" or "A and C cannot both hold" can be written, machines can reason over it
5. 5Knowledge graph: actual data (individual instances) loaded on top of an ontology used as its schema

A taxonomy shows what belongs where. An ontology also shows what does what to what, and what is not allowed. That difference decides whether a judgment can be handed to a machine.

One more distinction matters in practice: the difference from a data model. A data model is the blueprint for containers of data. It decides which tables and fields exist and connects directly to actual data. An ontology is a dictionary and relationship map for meaning. It decides what a field name refers to and how it relates to other fields. The former is strong on storage and processing, the latter on consistency of meaning. In practice, the realistic approach is to use the data model as the skeleton and deepen definitions with an ontology where differences in meaning cause problems. Six basic elements are common to both: what the data is about, what it relates to, what identifies it, at what granularity, as of when, and what it means.

The standards for description are set by the W3C (World Wide Web Consortium). RDF is the data representation format, RDFS and OWL are the ontology languages, SKOS is a lightweight vocabulary for taxonomies and thesauri, and SPARQL is the query language.

Representative ontologies in operation include the Gene Ontology in biology (established in 1998 to describe the functions of gene products in an internationally unified vocabulary), FIBO (Financial Industry Business Ontology) in finance, and schema.org for the web at large (launched jointly by Google, Microsoft and Yahoo! in 2011 to convey the meaning of pages to search engines).

In industrial data sharing, the information model of the International Data Spaces Association (IDSA) and the semantic models of Catena-X, built by the European automotive industry, serve as a common language among participating companies.

There are two reasons.

First, data sharing across companies and industries has become a practical issue. In arrangements where many companies bring data to a shared platform, such as Japan's Ouranos Ecosystem or Europe's Catena-X, there must be a guarantee that the same word is used with the same meaning. Even with the lines connected, data cannot be matched if the meanings of terms differ. An ontology provides that guarantee. Alongside technical interoperability, this is called semantic interoperability.

Second, the spread of generative AI. Large language models are strong at statistical association but do not strictly guarantee that "A" in one document and "A" in another refer to the same thing. In systems that search internal documents to answer questions (RAG) and in agents that carry out work automatically, that ambiguity leads directly to errors. Placing an explicitly defined system of concepts underneath statistical inference is becoming a common way to give it firm footing.

When a company tries to provide the data it holds to an outside party, the first obstacle is not technical. It is definitional.

Data that has long been used internally often has field names whose meanings differ by department, inconsistent granularity, and home-grown code systems. Tacit understanding keeps it working inside the company, but that understanding is lost the moment the data is handed to an outside party. This is why data that can be handed over as it is hardly exists. Even when an industry standard or shared platform is provided from above, a company cannot connect to it unless its own data definitions are in order to receive it.

Whether data can be monetized, that is, whether it can be priced and a transaction completed, depends on whether this work has been done. What is recorded, at what granularity, under which definitions, and how will the other party interpret it? The first thing we do in our Data Asset Advisory service is to write down this tacit understanding.

Whether or not the word ontology is used, every data transaction begins with the work of aligning meaning.