Is semantic translation the same as converting XML to JSON?

No. That changes representation; semantic translation addresses the meaning expressed by the data, though both operations can occur together.

Must it use an ontology?

No. Ontologies are one approach; sufficiently precise dictionaries and mappings can support some tasks.

Can it always be reversed?

No. If distinctions are merged or information is lost, the reverse translation may be ambiguous or impossible without retained context.

AI Glossary / Data engineering and architecture / Intermediate

Semantic translation

Definition

Semantic translation transforms information from one conceptual representation into another while aiming to preserve its intended meaning. In data engineering, this can involve mappings between vocabularies, schemas, or ontologies and rules for converting values. It is more than changing a file's syntax or translating field labels. This entry uses the data-interoperability meaning, rather than the separate use of the phrase in human-language translation studies.

What it means

A source system and a target system may describe the same situation differently. One may store a temperature with an explicit unit, while another expects a specific unit and property name. Semantic translation must identify what is being measured and how the target expresses that meaning. If the target cannot represent an important distinction, the process should expose that limitation instead of producing a superficially valid but misleading record.

How it works

Establish source and target definitions, then specify correspondences and transformation rules. These may map concepts directly or pass through an agreed intermediate model. Apply the rules to records, validate the target representation, and check whether the intended interpretation survives. Retain provenance and information about unresolved cases. A technical translator executes the mappings it receives; it cannot assume that two concepts are equivalent just because their names look alike. Changes to either model require review of the translation.

Three practical examples

Sensor integration: a service converts measurements into an agreed observation model while preserving the measured property, unit, time, and device identity.
Catalog exchange: a supplier's product classifications map into a buyer's categories, with explicit handling when one category is broader than another.
Organizational reporting: a team translates local status codes into shared concepts, retaining an “unmapped” outcome when the source status has no valid counterpart.

Think of it this way

Think of explaining the same game rules using two different rulebooks. A successful translation preserves what actions are allowed and what outcomes mean, not just the words on the page. The analogy is limited because data translation often uses executable mappings and formal constraints. It must also handle incomplete records and representational limits that a human translator might resolve through conversation.

Worked example

Source A contains a delivery record with status code 2, defined as “package handed to carrier.” Target B has statuses Prepared, InTransit, and Delivered. The approved mapping translates code 2 to InTransit, not Delivered. A record with source code 9, defined as “awaiting investigation,” has no exact target status. Mapping it to Delivered merely because the target requires a value would change its meaning. The integration instead sends it to an explicit exception path or uses a supported qualified representation. If the target later adds UnderReview, the mapping can be revised after confirming that the definitions match. This example illustrates conceptual translation rather than a cosmetic code replacement.

Illustrative case study

Fictional example: Poplar Sensors connects two building-management platforms. Both record humidity, but one value describes relative humidity and another represents an absolute moisture measure. Engineers reject a direct property rename and document the different quantities. They only convert records when the necessary inputs and a justified transformation are available. The remaining records retain their original meaning and a clear limitation. This invented case illustrates cautious mapping, not a claim about a named platform or a universal conversion formula.

Common misconceptions and mistakes

Do not assume syntax conversion preserves semantics. Do not treat broader and narrower categories as identical. Avoid replacing unknown values with plausible defaults unless the policy explicitly permits and labels them. A translation can also be directional: the target may omit distinctions needed to reconstruct the source. Finally, a formal mapping can preserve the wrong interpretation if the source definition was misunderstood. Domain evidence and representative examples remain essential.

Business use

Use semantic translation when systems must act on each other's data despite different conceptual models. Begin with a narrow exchange and clearly defined outcomes. Review mappings with owners of both models, test exceptional records, and measure unmapped or lossy cases. For AI tools, retain the source meaning behind translated fields so the assistant can explain qualifications. Do not let a natural-language model invent missing correspondences merely to produce a complete-looking response.

Check your understanding

A source distinguishes “requested cancellation” from “completed cancellation,” but the target has only “cancelled.” Is mapping both source states to that target always acceptable?

Answer: No. It may erase a consequential distinction. The receiving task must explicitly accept that loss or the integration must preserve the difference through another supported mechanism.

Related terms

Learn next

choose two small code lists with overlapping meanings. Write a mapping table containing exact matches, broader matches, and unsupported cases. Test one record from each group and explain what the receiver can legitimately conclude after translation.

Sources

Primary sources consulted: aerOS Semantic Translator documentation, https://docs.aeros-project.eu/en/latest/data/autonomy_semantic/translator.html — excerpts describing direct and central-ontology mappings consulted. Dou, McDermott, and Qi, Ontology Translation on the Semantic Web, https://ix.cs.uoregon.edu/~dou/research/papers/JDS04.pdf — research excerpt consulted. ITU-T Y.4563 publication excerpts also informed context. Examples are original; full editorial review remains pending.

Previous: Semantic technology Next: Semantic triple