Does this require moving every source into a graph database?

No. Virtual approaches can query underlying sources through mappings, while other approaches materialize integrated data.

Is an ontology the same as a database schema?

Both describe structure, but an ontology focuses on domain concepts and their formal relationships; mappings connect those concepts to particular storage structures.

Can it work with imperfect data?

Yes, but imperfections must be represented or handled explicitly. Integration does not make missing or contradictory observations disappear.

AI Glossary / Data engineering and architecture / Intermediate

Ontology-based data integration

Definition

Ontology-based data integration uses an ontology as an explicit model of meaning to connect information from different sources. Mappings relate source structures to shared concepts and relationships, allowing queries or combined data to be interpreted through that conceptual model.

What it means

Different systems often describe related things with different fields, codes and assumptions. One may store a course description, another a scheduled delivery and another a learner's registration. Integration needs to establish how those records relate, not merely copy them into the same database.

An ontology provides a vocabulary and formal relationships for the intended domain. Mappings explain how source records express that vocabulary. The resulting access can be materialized, with integrated data stored, or virtual, with queries translated to source operations when needed.

The approach is broader than matching ontology labels. It connects actual source data to the model and must address identity, units, missing information and source-specific meaning. The ontology does not resolve these issues automatically.

How it works

Designers establish the questions the integrated system must answer and define the relevant concepts. They inspect source schemas and create mappings from rows, fields or other structures to those concepts and relationships.

For relational sources, a mapping can associate a query result with RDF statements. R2RML is one standard language for expressing relational-to-RDF mappings. In virtual knowledge graph systems such as Ontop, queries expressed in SPARQL can be translated into SQL over the underlying relational data.

Reasoning can use supported ontology relationships to broaden or interpret a query. For example, a request for equipment may include records mapped to a declared subclass. Which inferences are supported depends on the system and ontology language profile.

Validation checks whether mappings express the intended facts and whether identifiers connect the right entities. Performance and availability still depend on the sources and execution approach. A virtual view does not guarantee a common snapshot across independently changing databases, while a materialized view needs a refresh policy.

Three practical examples

A training organization connects catalog, scheduling and enrollment data through a shared model of courses, deliveries and registrations.
A maintenance team maps different asset-system field names to common equipment concepts while preserving source identifiers.
A research project exposes relational observations through a knowledge graph vocabulary so users can query domain concepts.

Think of it this way

Think of an interpreter's guide that explains how several departments' forms express the same or related business ideas. The guide connects meanings; a separate process retrieves records.

The analogy is limited because formal mappings can generate machine-readable statements and support inference. An incorrect mapping can systematically produce wrong answers even when every source record is accurate.

Worked example

Preserving meaning while converting units

An invented integration combines two temperature sources. Source A records a reading of twenty with a Celsius unit. Source B records sixty-eight with a Fahrenheit unit. Both describe temperature observations, but their raw numbers are not directly comparable.

A mapping into a shared Celsius representation converts B using (68 − 32) × 5 ÷ 9 = 20. The integrated representation can now express both observations with the same unit, while retaining provenance and the original values where required.

Simply renaming both source columns temperature would not perform that conversion. Nor would assigning both records to a TemperatureObservation class. The mapping or transformation must explicitly implement the agreed representation.

If source B omits its unit, the system should not infer Fahrenheit solely because the number looks plausible. The example illustrates a semantic integration requirement rather than a claim that every ontology engine performs unit conversion automatically.

Illustrative case study

A fictional learning network integrates course information from regional providers. One source uses course_id for a reusable curriculum, while another uses it for a particular scheduled class. Early combined reports treat them as interchangeable and miscount available offerings.

The team defines separate CourseDescription and CourseDelivery concepts and maps each source accordingly. A delivery is related to its description, and registrations refer to deliveries. It tests examples in which the same description is delivered several times.

The network then offers queries through the shared model, retaining source links so disagreements can be investigated. An AI assistant may help users formulate questions, but the mappings establish what the answers mean. Conceptual distinctions matter before adding a natural-language interface.

Common misconceptions and mistakes

One vocabulary makes all source data equivalent. Different scope, units and identifiers still need explicit handling.
Virtual integration means no data is ever copied. Implementations can cache or materialize results; the actual architecture determines movement.
Reasoning repairs incorrect mappings. It can instead propagate their consequences.
A common query interface guarantees one consistent moment in time. Cross-source consistency requires separate consideration.

Business use

Ontology-based integration can be useful when several applications need a stable conceptual view despite heterogeneous source structures. It can reduce the need for every consumer to rediscover the same semantic relationships.

Plan for mapping ownership and change management. A renamed column may be easy to detect, while a changed business definition can silently invalidate a mapping that still executes successfully. Validate both technical execution and domain meaning.

For AI applications, the ontology can supply structured context and relationships for retrieval or analysis. It does not guarantee truthful generated answers or complete source coverage. Preserve provenance and explain the scope of the integrated view so users can assess what a result includes.

Check your understanding

Two source columns are both mapped to temperature, but one contains Celsius and the other Fahrenheit. Has naming them alike made their numbers directly comparable?

Answer: No. The units and any required conversion must be handled explicitly, with the meaning and provenance retained.

Related terms

Ontology
Semantic integration
mapping
ontology-based data access
virtual knowledge graph
R2RML
SPARQL
provenance

Learn next

Read semantic interoperability and data mapping. Choose two small sources and document how their identifiers, units and entity meanings relate before writing a shared query. Test one case where similar column names conceal different concepts.

Semantic interoperability

Sources

W3C, R2RML: RDB to RDF Mapping Language: https://www.w3.org/TR/r2rml/. Primary specification allowing mappings to be used for materialised or virtual access.

Ontop, Key concepts: https://ontop-vkg.org/guide/concepts.html. Primary implementation explanation of ontologies, mappings and virtual knowledge graphs.

Mapping Patterns for Virtual Knowledge Graphs: https://arxiv.org/abs/2012.01917. Primary research on defining and maintaining mappings. The temperature calculation and learning-network case are original illustrations.

Previous: Ontology alignment Next: Ontology engineering