A terminology alert should help a linguist find a real error, not force an approved word into every sentence.
Large multilingual projects can contain thousands of approved terms, yet a simple exact-match check cannot tell whether a term is correct in a particular sentence. A source word may be a noun in one context and a verb in another. An approved Greek term may also appear in different inflected forms.
This methodology article adapts the problem-and-solution structure of an RWS terminology-QA case study to OMADA’s English-to-Greek practice. It does not claim that OMADA deployed RWS’s product or achieved its reported pilot results.
Why exact matching is not enough
A glossary can contain several kinds of weak entries:
- duplicates
- broad, general words
- outdated entries
- terms that fit only a specific domain
Rule-based alerts can produce false positives, while a fluent AI suggestion may quietly alter meaning. In legal, medical and institutional content, both missed terminology errors and unnecessary corrections can be costly.
English words such as “default”, “charge” or “claim” can shift function and meaning by domain, and Greek inflection adds further variation. A system that matches strings without considering part of speech, subject field and context can generate excessive alerts that distract from genuine problems.
A human-led QA workflow
First, a terminologist checks the glossary. Each entry should record:
- one concept per entry
- source and target forms
- subject field
- a definition or context
- approval status
- an owner for updates
Next, a controlled QA pass compares source and target segments and flags candidate issues. A language model may help assess context or suggest a correction, but it should provide the evidence and uncertainty behind the flag.
A Greek linguist then accepts, edits or rejects each flag against the source, the client-approved term and the grammatical setting. The decision is recorded so repeated alerts and glossary ambiguities can be resolved. The same principle underpins MTPE and AI-output review: the model assists, the reviewer decides.
- 01
Prepare the glossary
Deduplicate concepts, distinguish senses and capture subject field, grammatical information and approved variants.
- 02
Detect candidates
Compare source and target segments against the controlled terms, recording the context and the reason for each alert.
- 03
Assess and propose
Use a constrained model only as an aid to evaluate whether the term applies and to suggest a possible correction with evidence.
- 04
Review and learn
A Greek linguist validates or rejects the finding; the team categorizes false positives, updates the glossary and retests on a held-out sample.
The reviewer keeps authority
The reviewer must retain authority over meaning, grammar and the final text. A model can surface patterns at scale.
A terminologist resolves ambiguous entries, and a project lead ensures that findings become durable process improvements. See how we support AI teams with linguistic review.
How a pilot should be measured
This is a proposed OMADA methodology, not a report of a deployed AI product. Before any claim of reduced errors or saved hours, a controlled pilot should sample correct and incorrect segments and define a baseline, a test set and human reference decisions.
The pilot should then report:
- precision and recall
- false-positive categories and their causes
- correction quality
- reviewer time and effort against the baseline
OMADA should publish performance figures only after such a pilot, with permission to disclose them. RWS’s published figures belong to its own system and clients.
The intended benefit is a more focused human review, clearer terminology records and an auditable decision trail—not automatic approval of AI output. For a way to classify findings, see a useful taxonomy for linguistic QA.
The strongest part of an AI-assisted terminology check is often the quality of the glossary and the feedback loop around it. Automation is useful when it directs human attention to well-evidenced issues.
Sources and scope: RWS, AI + Human Expertise: Smarter Terminology QA at Scale ↗. A methodology study, not a claim of OMADA deployment or measured OMADA results.