Must source ontologies be discarded after a merge?

No. They may remain authoritative for their original uses, with the combined model maintained separately.

Can merging be automated?

Tools can suggest operations and detect some conflicts, but semantic decisions often require domain judgment.

Is importing an ontology always a merge?

Not necessarily. An import can reuse another ontology's statements without performing the reconciliation usually meant by merging.

AI Glossary / Data engineering and architecture / Intermediate

Ontology merging

Definition

Ontology merging is the construction of a combined ontology from two or more source ontologies. It reconciles overlapping concepts, relationships and logical statements so that the resulting model can serve an agreed purpose without silently changing their intended meaning.

What it means

An ontology is more than a list of terms. It can describe classes, properties, individuals and rules about their relationships. Combining two ontologies therefore involves more than copying their labels into one file or deleting similarly named entries.

Merging differs from alignment. Alignment records correspondences while the source models can remain separate; merging produces a combined model. An alignment can inform a merge, but accepting a correspondence is not the same as deciding how every definition and restriction should appear in the result.

There may be no single correct merge independent of purpose. A shared reporting ontology might need a different scope from a detailed engineering model. The team must decide what to preserve, what to reconcile and which differences remain meaningful.

How it works

The process begins by examining source scope, identifiers, definitions and modeling assumptions. Candidate overlaps are identified through labels, structure and other evidence. Domain experts determine whether similarly described entities are equivalent, related or genuinely different.

A merge plan specifies how entities and statements will be retained or transformed. Identifiers may be preserved, replaced or connected through documented mappings. Provenance records which source supplied a statement and why a conflict was resolved in a particular way.

Logical checks then examine the combined consequences. Two source ontologies can each be consistent yet conflict when their statements and mappings are combined. A class may become impossible to instantiate, or an asserted individual may create an outright inconsistency.

Validation also uses practical questions and sample data. A model can pass a consistency check while answering the intended business questions incorrectly because a definition was misunderstood. Versioning and regression checks help maintain the merged model as its sources evolve.

Three practical examples

Two research projects combine overlapping domain ontologies into a shared model while retaining the origin of important definitions.
An organization reconciles product concepts from acquired businesses before establishing a common knowledge graph vocabulary.
A learning platform constructs a combined course ontology from separate content and scheduling models, preserving the distinction between a course description and a delivery.

Think of it this way

Think of combining two technical dictionaries that also contain rules about how their terms relate. Identical words may need different entries, while different words may describe the same idea.

The analogy is limited because logical statements can generate consequences that are not obvious from reading definitions separately. A successful merge requires checking those consequences, not merely editing the prose into a consistent style.

Worked example

Finding a conflict introduced by combination

Consider two invented ontologies. The first says that PhysicalItem and DigitalItem are disjoint: nothing can belong to both classes. It also states that product P is a PhysicalItem.

The second states that its product record Q is a DigitalItem. A proposed merge declares P and Q to be the same individual because their labels match. The combined model now says that one individual belongs to both disjoint classes.

Under these stated OWL-style assumptions, the combination is inconsistent. Each source could have been consistent alone; the identity mapping created the conflict. The team must investigate whether the records are actually the same item, whether a label refers to a package containing different things, or whether a source definition is wrong.

Simply deleting the disjointness statement would remove this particular logical contradiction but could erase an important domain rule. Repair should follow the intended meaning, with the decision documented.

Illustrative case study

A fictional educational publisher merges a glossary ontology with a catalog ontology. Both contain a model class. In the glossary it describes a computational model; in the catalog it describes a physical demonstration object used in lessons.

A preliminary matching step suggests combining them. Editors reject that equivalence after reviewing definitions and examples, retaining distinct identifiers and adding more specific labels. They do combine another pair of genuinely equivalent topic concepts, preserving the source mappings.

The merged ontology is tested against questions about digital tools and physical teaching materials. The publisher also records which source versions were used. The case illustrates why fewer terms is not necessarily a better merge: preserving a real distinction can be essential to correct retrieval and reasoning.

Common misconceptions and mistakes

Merging is concatenating files. Conflicting identifiers, definitions and logical statements require reconciliation.
The smallest resulting ontology is the best. Removing legitimate distinctions can damage meaning.
Consistent sources guarantee a consistent combination. New mappings can make their statements interact incompatibly.
Passing a reasoner proves the business model is correct. Logical consistency does not establish that the chosen meaning matches the domain.

Business use

Ontology merging can reduce fragmentation when teams need a shared conceptual model. It also creates governance work: someone must own the combined definitions and decide how future changes are incorporated.

Agree on purpose and scope before resolving conflicts. Record mappings to legacy identifiers so that existing data can be interpreted or migrated deliberately. Review whether source restrictions and usage conditions permit the intended reuse.

For AI-assisted merging, treat generated correspondences and rewritten definitions as proposals. An assistant may make the text look coherent while weakening an important restriction. Test the logical consequences and representative business questions, especially where the model supports automated decisions.

Check your understanding

Two consistent ontologies become inconsistent after a merge equates two individuals. Does that prove one source was already inconsistent?

Answer: No. The added identity claim can create a contradiction between previously separate statements. The mapping and intended meanings must be examined.

Related terms

Ontology alignment
ontology integration
logical consistency
disjointness
identity
provenance
ontology versioning
knowledge graph

Learn next

Read ontology alignment and logical consistency. Combine two small concept models and list every identity or equivalence decision. Test a case in which similar names have different meanings before accepting the merge.

Ontology alignment

Sources

Noy and Musen, PROMPT: Algorithm and Tool for Automated Ontology Merging and Alignment: https://www.cs.ubc.ca/~rap/teaching/534a/readings/Noy-00.pdf. Original research on interactive assistance and conflict handling.

Stanford Protégé, PROMPT: https://protegewiki.stanford.edu/wiki/PROMPT. Primary project documentation; cited as historical method context, not a recommendation to install the legacy tool.

W3C, OWL 2 Primer: https://www.w3.org/TR/owl2-primer/. Primary explanation of identity, equivalence and disjointness. The conflict example and publisher case are original illustrations.

Previous: Ontology learning Next: Opacity