What is Ontology Mapping?

Updated October 2026

Ontology mapping is the process of stating how concepts in one ontology or data model relate to concepts in another. Each mapping says that two terms are equivalent, that one is broader than the other, or that they are otherwise related. The result lets software query and reason across systems that were built with different vocabularies, such as a CRM and a billing database.

What is ontology mapping?

Ontology mapping connects meaning across models, not just field names. Two teams can model the same business in different ways. One ontology may call a paying account a Client, while another splits it into Customer and Subscriber. A mapping records those relationships explicitly so tools stop guessing.

The research literature separates two ideas. According to the ontology alignment overview on Wikipedia, matching predicts how similar two terms are and returns a score. Mapping produces logical statements, such as equivalence or inclusion, that a system can use. In practice the words overlap, and most teams say mapping for the whole job.

The same overview lists the main relationship types. These are equivalence, subsumption, disjointness, and part of. Mappings can also be simple, linking one term to one term, or complex, linking one term to a combination of several.

How is ontology mapping different from schema mapping?

Schema mapping moves data between structures, while ontology mapping aligns what the data means. A schema mapping might say that the column cust_id in one table feeds customer_id in another. It does not say whether a trial user counts as a customer.

An ontology mapping works one level up. It links concepts that carry hierarchy, properties, and rules. That is why it is the step that settles questions like whether a reseller is a kind of customer. Once concepts are aligned, you can bind them to real tables with SQL to RDF mapping tools such as R2RML engines.

Most enterprise projects need both layers. The ontology mapping decides that two models agree on Customer. The data mapping then decides which rows in which systems become instances of Customer.

What do ontology mappings look like?

A mapping is a small statement with a subject, a relationship, an object, and ideally a reason. The relationship comes from a shared vocabulary so any tool can read it. Three formats cover most real work.

SKOS mapping properties

The W3C SKOS Reference defines mapping properties for linking concepts across schemes. They include exactMatch, closeMatch, broadMatch, narrowMatch, and relatedMatch. The spec makes exactMatch transitive and leaves closeMatch non transitive, because chained approximate matches lose precision. That distinction matters when mappings pass through a hub vocabulary.

OWL axioms

OWL 2 is the W3C ontology language with formally defined meaning. Mappings written as OWL equivalence or subclass axioms can be used by a reasoner. That lets software infer new facts and detect contradictions, but it also means one bad mapping can make a whole merged model inconsistent.

SSSOM and the Alignment format

SSSOM, the Simple Standard for Sharing Ontological Mappings, stores each mapping with a predicate, a justification, a confidence score, and the author. It suits teams that review mappings in spreadsheets and version them in git. The Alignment API from INRIA calls itself "an API and implementation for expressing and sharing ontology alignments." It defines an RDF alignment format plus EDOAL, a language for complex correspondences.

Which techniques are used for ontology matching?

Modern matchers combine lexical, structural, logical, and learned signals, then filter the results. No single technique works across domains. The systems that did best in the 2025 OAEI campaign each used several.

Lexical matching

Lexical methods compare labels, synonyms, and definitions. They normalize text, compare strings, and use synonym lists, so Client can match Customer. They are fast and precise when labels are rich. They fail on homonyms and on terse business names like acct_typ.

Structural matching

Structural methods look at where a concept sits. If two classes have similar parents, children, and properties, they are more likely to match even when labels differ. Matcha, for example, adds instance based and cross reference matching on top of lexical methods.

Logic based repair

Logic based methods use a reasoner to check whether a set of mappings creates contradictions. LogMap is the best known example. It combines matching with repair to minimize logical inconsistencies, which keeps a merged ontology usable.

Embedding based matching

Embedding methods turn each concept into a vector and compare vectors. Matcha uses a pretrained sentence BERT model for this, according to its 2025 OAEI paper. Oxford's DeepOnto packages BERTMap, a BERT based matcher, plus the Bio-ML datasets used at OAEI.

LLM based matching

LLMs are now used to propose and validate mappings. LLMs4OM tested a retrieval plus zero shot prompting design on 20 datasets and found LLMs can match or beat traditional systems, especially on complex cases. Agent-OM uses two LLM agents, one to retrieve candidates and one to match them.

The most practical pattern is narrower. The LogMap team's 2025 paper describes Gemini 2.5 as an "Oracle for the cases where LogMap is uncertain." That keeps LLM cost and variability limited to the hard cases.

Which ontology matching tools are current in 2026?

LogMap, Matcha, DeepOnto, OntoAligner, and Agent-OM are the actively developed open-source options. The Ontology Alignment Evaluation Initiative has benchmarked matchers every year since 2004, and its 2025 campaign ran more than 20 tracks. The table summarizes where each tool fits.

LogMap and LogMapLLM

LogMap can match ontologies with tens or hundreds of thousands of classes and repairs mappings that cause contradictions, per its GitHub repository. It scored the top F1 of 0.64 in the OAEI 2025 conference track. The limit is that LogMapLLM runtimes are only approximate because LLM API call times fluctuate, as the anatomy track results note.

Matcha

Matcha comes from the LASIGE team in Lisbon and carries over the key algorithms of AgreementMakerLight on a broader core. Its 2025 paper reports the highest F-measure in 10 of 32 OAEI tasks. It also led the anatomy track with 0.941. Its limit shows on small business style models, where its conference track F1 was 0.63.

AgreementMakerLight

AgreementMakerLight is an Apache 2.0 matcher built around efficient element level techniques. It still works for existing pipelines. Its repository copyright runs from 2013 to 2020, and active development has moved to Matcha.

DeepOnto and BERTMap

DeepOnto from Oxford's KRR group is an Apache 2.0 Python package with BERTMap and evaluation tooling. It is a strong base for research on learned matching. Its limit is accuracy against hybrids. LogMapBio and LogMapLLM outperformed BERT based systems on the 2025 Bio-ML track, per the LogMap paper.

OntoAligner

OntoAligner is an Apache 2.0 Python library that wraps lexical, retrieval, LLM, retrieval augmented, and embedding based aligners behind one pipeline. It won the Best Resource Paper Award at ESWC 2025. It is a toolkit rather than a tuned matcher, so quality depends on the pipeline you assemble and evaluate.

Agent-OM

Agent-OM applies LLM agents to matching and was published at VLDB 2025. It placed in the top three on the 2025 anatomy track in its first year. On the conference track it scored an F1 of 0.61, behind classic LogMap.

How does an ontology mapping project work step by step?

Scope the concepts, generate candidates, filter with logic, review with experts, then store and maintain the result. Tools speed up the middle steps. The first and last steps decide whether the mappings stay correct.

Step 1. Define the target model and scope

Decide which ontology is the hub and which concepts matter first. Start with the entities behind your key questions, such as Customer, Account, and Contract. Mapping everything at once produces thousands of candidates nobody reviews.

Step 2. Enrich labels and synonyms

Matchers depend on text. Add human readable labels, definitions, and known synonyms to both sides before running anything. This one step often does more for quality than switching tools.

Step 3. Generate candidates with several matchers

Run a lexical matcher and an embedding or LLM matcher, then combine their outputs. Keep the similarity score and the method for each candidate. Those fields become the justification and confidence in your mapping file.

Step 4. Repair with a reasoner

Load the candidate mappings with both ontologies and check for contradictions. LogMap does this automatically. Removing mappings that break consistency prevents one bad equivalence from spreading through the merged model.

Step 5. Review with domain experts

Route low confidence and high impact mappings to the people who own the meaning. A finance lead should approve what counts as revenue. LLMs can draft a rationale for each candidate, but a person should accept it.

Step 6. Store, test, and version the mappings

Save accepted mappings in an open format such as SSSOM or OWL, under version control. Add tests that check known correct pairs and flag new contradictions. Rerun matching when either ontology changes.

Where does ontology mapping matter in enterprise data?

It matters wherever several systems describe the same business things differently. Common cases are mergers, multi system customer views, and AI agents that query across departments.

Vendors now ship their own semantic models, which adds another mapping target. Microsoft's Fabric IQ ontology, still in preview, can generate ontologies from existing Power BI semantic models. Each vendor model covers its own platform well, so mapping between them and your own ontology becomes ongoing work.

Mappings also improve AI answers. In the data.world benchmark, GPT-4 answered 16 percent of enterprise questions correctly over raw SQL and 54 percent over a knowledge graph of the same data. The full evidence is covered in how ontology powers AI analytics.

Ontology mapping aligns concepts, while entity resolution decides which records are the same real customer. Most projects need both. For the graph platforms that store the result, see the knowledge graph platforms guide.

Who should own ontology mappings?

Your company should own its mappings, in open formats, on infrastructure it controls. Mappings record what your data means, which is institutional knowledge. SKOS, OWL, SSSOM, and R2RML are all open, so nothing forces that knowledge into one vendor's format.

Some teams build this with a partner on their existing stack. Galaxy is one such partner, using internal tooling that speeds up ontology, mapping, and entity resolution work. It builds in the open, starting with Filament by Galaxy, and plans to open source more of that tooling.

Frequently asked questions

What is ontology mapping in simple terms?

Ontology mapping is the work of stating which concepts in one model mean the same thing, or something related, as concepts in another. If one system says Client and another says Customer, a mapping records that they are equivalent. Software can then query, merge, and reason over both systems as if they shared one vocabulary.

What is the difference between ontology mapping and ontology matching?

Matching produces candidate pairs with a similarity score. Mapping is the reviewed result, a set of logical statements such as equivalence or subsumption that systems can act on. In practice a matcher proposes correspondences, people and reasoners filter them, and the accepted set becomes the alignment that downstream tools consume.

How is ontology mapping different from schema mapping?

Schema mapping connects table and column structures so data can be moved or transformed. Ontology mapping connects concepts and their meaning, including hierarchy and logical rules. A schema mapping might copy cust_id into customer_id. An ontology mapping states that a Client in one model is the same kind of thing as a Customer in another.

What techniques are used for ontology matching?

Most matchers combine several techniques. Lexical methods compare labels and synonyms. Structural methods compare parents, children, and properties. Logic based methods use reasoning to remove mappings that cause contradictions. Embedding and LLM methods compare meaning rather than spelling. Strong systems in the 2025 OAEI campaign mixed at least three of these.

Can large language models do ontology mapping?

They help, mostly as validators and for hard cases. LogMapLLM asks Gemini 2.5 only when classic LogMap is unsure, and Agent-OM uses LLM agents for retrieval and matching. In the 2025 OAEI conference track, classic LogMap still scored the top F1 of 0.64, so LLMs work best inside a hybrid pipeline.

Which ontology matching tools are still maintained in 2026?

LogMap and its LLM variant, Matcha from the AgreementMakerLight team, DeepOnto from Oxford with BERTMap, and the OntoAligner Python library are all active and open source. LogMap, Matcha, and Agent-OM all competed in the 2025 OAEI campaign. The Alignment API remains useful for storing and evaluating alignments, but it is not a matcher.

What is the OAEI?

The Ontology Alignment Evaluation Initiative runs a yearly benchmark campaign for ontology matching systems and has done so since 2004. The 2025 campaign covered more than 20 tracks, from anatomy and conference ontologies to biomedical, multilingual, and instance matching tasks. Results were presented at the Ontology Matching workshop at ISWC 2025.

What format should ontology mappings be stored in?

Use an open format that records who made each mapping and why. SSSOM is a simple tabular standard with fields for predicate, justification, and confidence. OWL axioms suit mappings you want reasoners to use. The Alignment format and EDOAL from INRIA handle richer and complex correspondences. Avoid storing mappings only inside one vendor tool.

How accurate is automated ontology matching?

It depends heavily on the domain. On the OAEI 2025 anatomy track the best system reached an F-measure of 0.941. On the conference track the best was 0.64. Ontologies with rich labels and synonyms match well, while small business models with terse names need far more human review.

Who should own ontology mappings in a company?

The company should own them, in open formats under version control, because mappings encode what its data means. A data or knowledge engineering team usually maintains them, while domain experts approve the meaning. Tools and partners can generate candidates, but the accepted mappings should live on infrastructure the company controls.