No. A CAT workflow can include conventional retrieval and editing features as well as AI-generated suggestions.

Is post-editing the whole process?

No. It is one activity that may occur within a wider assisted-translation workflow.

Does a glossary replace translation memory?

No. Preferred terms and previously translated segments provide different kinds of assistance.

AI Glossary / Natural language and speech / Advanced

CAT

Definition

CAT stands here for computer-aided translation, also called computer-assisted translation: a workflow in which software supports a human translator's work. Assistance can include translation memory, terminology resources, editing tools, and machine-translation suggestions. CAT is broader than automatically generating a translation and is not the name of one particular AI model.

What it means

A translator often encounters recurring phrases, product terminology, and documents with similar structures. CAT tools help retrieve earlier work and keep these resources available while the translator decides how to express the current meaning for the intended audience.
The human role is substantive. An earlier translation may be unsuitable in a new context, and a fluent machine suggestion may misrepresent a requirement. Computer assistance can reduce repetitive work without eliminating the need to understand the source, the target language, and the purpose of the document. This entry uses the translation meaning of CAT, rather than unrelated uses of the acronym.

How it works

A project imports source files, presents manageable text segments, and makes relevant resources available beside an editing area. A translation memory stores paired source and target segments for retrieval. A terminology resource helps identify preferred expressions, while a machine-translation service may generate a proposed target segment.
The translator can accept, adapt, reject, or independently replace suggestions. The workflow may include checks for missing content, inconsistent terminology, numbers, or formatting tags, followed by review of the assembled document. The exact features depend on the tool and project configuration.
Translation-memory matching and machine translation perform different jobs: one retrieves stored examples, while the other generates output. They can work together, but neither a close match nor a generated sentence is automatically approved for publication. The project needs a clear rule for when an edited segment becomes reusable, approved material.

Three practical examples

A product team updates a user guide and retrieves translations of unchanged instructions. Reviewers still check that the surrounding workflow and product version support reusing them.
A website team uses an approved terminology list to keep navigation labels consistent across pages. Translators review each label in the interface because available space and intended action can affect wording.
A service organization supplies machine-generated drafts of routine notices to translators for post-editing. The translators check meaning, tone, names, and numbers before the notices enter the normal approval process.

Think of it this way

Think of CAT as a translator's organized workbench. Earlier examples, terminology notes, and draft suggestions are within reach, but the person using the workbench decides which materials fit the current job. A well-stocked workbench helps; it cannot decide by itself whether a borrowed component is appropriate.

Worked example

Consider a fictional 100-segment update. Assume all segments are equally weighted for this simplified calculation. Thirty have reusable translation-memory suggestions requiring one minute of human checking each. Forty have suggestions requiring three minutes of editing and checking each. The remaining thirty require six minutes each to translate and check.
Segment-level work totals 30×1 + 40×3 + 30×6 = 330 minutes. Add 40 minutes for project setup and final document review, giving 370 minutes overall.
Suppose translating and checking every segment without those suggestions takes six minutes per segment, with the same 40 minutes of overhead. That total is 100×6 + 40 = 640 minutes. The difference is 270 minutes, or 270/640 = 42.1875%, approximately 42.2%.
These are stipulated planning times, not measured productivity findings. They do not show that all CAT projects achieve this saving, and the comparison is meaningful only if both workflows meet the same quality requirements. Segments also vary greatly in length and difficulty in real projects.
Now suppose twenty of the apparently reusable segments actually need six minutes each because their context changed. Each takes five minutes more than planned, adding 100 minutes. The revised total is 470 minutes, and the saving against 640 becomes 170/640, approximately 26.6%.
The exercise shows why reuse categories should guide estimates rather than substitute for review. It also keeps setup and final review visible: counting only editing time would give a misleading account of the complete workflow.

Illustrative case study

Fictional example: LocalePath localizes a booking website. A stored translation for a button labeled “Return” came from a page about sending an item back. A new page uses the same English label to mean going back to the previous screen. Matching the source text exactly does not make the old target wording suitable.
The translator checks the screen context, supplies a context-appropriate translation, and records where that wording should be used. The team reviews the finished interface rather than relying only on isolated segment approval. This invented case illustrates the difference between matching text and matching meaning; it makes no claim about a particular tool's behavior or measured error rate.

Common misconceptions and mistakes

CAT and machine translation are not interchangeable terms. CAT can support translation through memories and terminology even without a machine-generated draft. Conversely, automatic translation without a translator-centered workflow does not illustrate all that CAT involves.
A translation-memory similarity score is not a probability that its target text is correct. Even an exact source match can need changes because of context, audience, or terminology updates. Another mistake is treating segment completion as proof that the exported document's layout, links, and reading flow have been reviewed.

Business use

CAT can support recurring multilingual publishing when organizations maintain useful translation memories and terminology resources. Define who approves reusable translations, how outdated entries are revised, and what final review covers. Evaluate productivity alongside meaning preservation and consistency. For any connected translation service, establish whether the project's content may be sent to it before making that service part of the workflow.

Check your understanding

Why is the first planning total 370 minutes?

Answer: The 330 minutes of segment work are combined with 40 minutes of setup and final review.

Why can an exact text match be unsuitable?

Answer: The same wording can have a different meaning in its new context.

What changes the revised saving?

Answer: Twenty segments require five extra minutes each, adding 100 minutes.

Related terms

Machine translation
Translation memory
Localization
Post-editing
Human-in-the-loop
terminology management
fuzzy matching
translation quality

Learn next

Compare a retrieved translation-memory suggestion with a generated machine translation and identify what each contributes. Then design a small localization workflow that records context, approval, terminology decisions, and review of the finished document or interface.

Sources

Jurafsky and Martin, Speech and Language Processing, chapter 13 introduction, CAT and post-editing: https://web.stanford.edu/~jurafsky/slp3/13.pdf OmegaT project, translation-memory, glossary, matching, and machine-translation interface features: https://omegat.org/

Previous: Case-control study Next: Catalog server

CAT