Does semantic integration require a central database?

No. It can be implemented through virtual access, shared mappings, or other architectures as well as copied data.

Must every source adopt the same schema?

No. Explicit correspondences can connect different schemas, provided the relationships are sufficiently precise for the task.

Is it the same as semantic interoperability?

Integration is work that connects information; interoperability is the ability of systems to exchange and use it with the intended meaning. The concepts support each other.

AI Glossary / Data engineering and architecture / Intermediate

Semantic integration

Definition

Semantic integration connects information from different sources while explicitly addressing how their concepts and values relate in meaning. It goes beyond placing records together or translating their syntax. The work may align vocabularies, reconcile definitions, transform values, or preserve distinctions that should not be merged. Its success depends on the intended question and the evidence supporting the proposed relationships.

What it means

Consider combining maintenance records with equipment sensor readings. One source may identify a physical machine, another a replaceable component, and a third a production line. Semantic integration establishes the relationships needed to use these records together without pretending that all three identifiers name the same thing. A shared storage platform can make access easier, but it cannot by itself establish these domain relationships or determine which measurements are comparable.

How it works

Identify the task and relevant source concepts. Define a shared model or explicit mappings among the existing models. Check identities, units, classifications, time meaning, and relationships. Transform values only under justified rules, preserve provenance, and flag unresolved correspondences. Implement the integration through a materialized dataset, a virtual query layer, a graph, or another suitable architecture. Validate outputs using examples where the correct interpretation is known. Revisit mappings when source definitions or the shared model change.

Three practical examples

Product discovery: a catalog links different supplier terms for the same product category while retaining important distinctions in material and intended use.
Research discovery: a repository connects datasets and publications through documented project and subject identifiers, helping users find related evidence across collections.
Operations analysis: engineers relate faults, inspections, and sensor observations to the correct equipment component and time interval before comparing them.

Think of it this way

Think of assembling a team of interpreters who agree on what each organization means by its key terms. Their task is not simply to translate words, but to preserve distinctions that affect decisions. The analogy stops at automation: a semantic integration system needs explicit representations and rules that can be executed or checked. Human intuition about a term is not automatically available to the software.

Worked example

Source A records a temperature of 20 degrees Celsius for sensor S1 at 10:00 UTC. Source B records 68 degrees Fahrenheit for device D7 at the same instant. A documented asset mapping says S1 and D7 refer to the same sensor, and both values describe air temperature at the same location. Converting B gives (68 − 32) × 5/9 = 20 degrees Celsius, so the observations agree numerically. If D7 instead names the surrounding room's thermostat, treating the records as duplicate observations would be unjustified even though the numbers match. Semantic integration needs identity, measured property, location, time, and units together. Equal converted values alone do not establish that two records describe the same observation.

Illustrative case study

Fictional example: Juniper Manufacturing combines inspection and maintenance data. Its engineers discover that “unit” means a whole machine in one source and a replaceable module in another. They model the part-of relationship and link events to the appropriate level. Reports can then separate machine-level downtime from module-level repair activity. The team retains source identifiers for traceability. This invented case illustrates a modeling choice and does not claim that the integration proves a causal relationship between repairs and later performance.

Common misconceptions and mistakes

Do not assume that sharing JSON, XML, or RDF automatically establishes shared meaning. Do not collapse concepts merely because their labels are similar. A broad category and a narrow category may overlap without being equivalent. Avoid silently discarding conflicts to produce a clean-looking combined result. Also distinguish integration from truth verification: consistent mappings can connect sources that contain incorrect observations. Provenance and quality checks remain necessary.

Business use

Use semantic integration when important questions cross system boundaries and ordinary field matching leaves ambiguity. Start with a limited set of valuable questions and validate the mappings needed to answer them. Include domain specialists and maintain ownership of the shared definitions. For AI assistants, use the integrated representation to support evidence retrieval and explain relationships, while retaining source citations and uncertainty. A connected graph of claims is useful only when its relationships are justified and its limitations remain visible.

Check your understanding

Two sources use “vehicle” and “car.” Can a mapping simply declare the categories identical?

Answer: Not without checking their definitions. If vehicle includes trucks and motorcycles, car is a narrower concept. Treating them as equivalent can cause incorrect inclusion or exclusion in queries.

Related terms

Semantic heterogeneity
Ontology alignment
Semantic interoperability
data integration
knowledge graph
provenance
schema mapping

Learn next

integrate two tiny datasets about the same domain. Write the identity, unit, category, and time assumptions before joining them. Include one deliberately incompatible record and confirm that the integration flags or separates it instead of forcing a match.

Sources

Primary sources consulted: NIST, Semantic Integration through Invariants, https://www.nist.gov/publications/semantic-integration-through-invariants; Foundations of Semantic Interoperability and Integration, https://www.nist.gov/itl/ai/foundations-semantic-interoperability-and-integration — abstract and project excerpts consulted. W3C, Best Practices for Publishing Linked Data, https://www.w3.org/TR/ld-bp/ — vocabulary excerpt consulted. Examples and implementation workflow are original; full editorial review remains pending.

Previous: Semantic heterogeneity Next: Semantic interoperability