Knowledge Graph Mediated Translation (KGMT)

For most language pairs, fluency is no longer the hard part. Control is: keeping domain terms, regulated language, and brand voice right across every language, at volume. Knowledge Graph Mediated Translation (KGMT) gives a translation model that control. It grounds the model in a knowledge graph of your domain, so the output stays tied to verified knowledge rather than to whatever sounds plausible.

Where translation still breaks

In regulated and specialist content, the risk isn't clumsy phrasing. It's a model that swaps in the wrong term, drops a qualifier that carries legal weight, or adds a detail that was never in the source. The longer and more domain-specific the text, the more often we've seen it happen. In a blog post that's an annoyance. In a label or a contract it can be a liability.

Peer-reviewed: what the knowledge graph does for compliance

In 2026, researchers at Ionian University published the first peer-reviewed evaluation of KGMT in a regulated domain. On a controlled corpus — a 500-word retinol product description seeded with 30 controlled compliance errors, translated from English into Greek and French — they compared KGMT against two GPT-5 baselines: a plain prompt, and a prompt with the full glossary and compliance tables pasted in.

• Plain GPT-5 prompt. Detected 0% of the source-compliance violations — and gave different outputs on different runs.
• GPT-5 with the glossary and compliance tables pasted into the prompt. Still 0%. The “just give the model the documents” shortcut did not deliver compliance.
• KGMT. Detected 100% of the source-compliance violations, with zero penalized translation errors — and a stable, auditable output.

The evaluation also separates errors a translation introduces from non-compliance already sitting in the source. KGMT flagged source violations instead of silently carrying them into every target language — in one case, a retinol concentration above the EU limit was translated faithfully and flagged as a source defect. That distinction names the real risk in many regulated translation workflows: not a bad translation, but a non-compliant source propagating unchecked into every market.

Under the study's stricter compliance-based scoring, KGMT scored 100.0 on translation quality in both target languages, against roughly 72–74 for the baselines. It is a controlled, single-domain result — one source text, two language pairs — not a claim about every workload. But it demonstrates the pattern KGMT was designed around: the model knows language; the knowledge graph is what lets it hold the law.

Gene, V. & Sosoni, V. (2026). Dual-Metric Compliance and Quality Evaluation of Knowledge Graph Mediated Translation in Regulated Domains: An Enhanced Architectural Framework. Proc. NeTTIT 2026, pp. 109–118. DOI: 10.26615/issn.2815-4711.2026_015. The knowledge graph behind the study was engineered by LangOptima and Lead Semantics, credited in the paper's acknowledgements.

How KGMT works

• Ground the model in your knowledge. A knowledge graph holds your domain's entities, terminology and relationships as verified facts. As it translates, KGMT works from that known knowledge instead of guessing. That constrains the fabrication ungrounded models tend to produce.
• Control the context, deliberately. Rather than hoping the model picks up the right context, KGMT delivers it on purpose. That helps most where plain models struggle: long-form and narrative text, specialist domains, and lower-resource languages.
• Keep provenance and version. KGMT separates your organization's own position from text it merely quotes, versions content by date, and carries brand characteristics through every language. What ships is traceable, not only translated.

Optional add-ons

Automatic Post-Editing (APE) and COMET quality-estimation scores are available where you want an extra automated editing pass, or a measurable quality signal on every segment rather than a spot check.

Where it fits

KGMT runs standalone, or it connects to your existing stack: a translation management system, or another retrieval-augmented generation (RAG) workflow. It's additive. It makes the stack you already have more controllable, without replacing any of it.

Where it comes from

KGMT is a formal specification, developed by Edwin Trebels, Founder of LangOptima, with Lead Semantics, and implemented in their knowledge-transformation tool, TextDistil. The knowledge-graph grounding is the same approach LangOptima brings to enterprise data, applied here to translation.

See what it does to your own content.

One question for a first call: where does translation quality actually cost you today, and what would "controlled" look like across your languages?

Book a call →

Working in regulated translation? The free GDPR Compliance Calculator includes a regulated-translation exposure module and shows a result in about two minutes, with no signup.