On this page
The components at a glanceThe TMS is the hubMemory and terminology are the assetsMachine translation: configure, do not just switch onQuality estimation and QATesting in contextThe order most teams add thingsA growing company buys a translation platform because a competitor uses it. A year later it has three glossaries in three tools, a machine translation engine nobody configured, and Korean reviewers working in a spreadsheet because the platform's review screen has no screenshots. The tool was fine. Nobody had decided which job each part of the stack was supposed to do.
"Translation stack" sounds like a single product. In practice it is a set of components, some bundled together and some separate, and growing teams add them in roughly the same order. This post lays out the components, what each one does, when you need it, and what to check before choosing. Product names appear only as examples; the components are what matter.
The components at a glance
| Component | What it does | You need it when | Check before choosing |
|---|---|---|---|
| Translation management system (TMS) | Holds source and target text, tracks versions, assigns work, runs the workflow | More than one language or more than one release a month | Repository and CMS connectors, review in context, export of only approved text |
| Translation memory (TM) | Stores every approved segment pair and offers matches for new text | From day one; it is usually built into the TMS | Can you export it in a standard format if you leave? |
| Termbase / glossary | Approved terms per language, with notes and forbidden variants | From day one | Shows terms to translators while they work; feeds the MT engine |
| Machine translation engine | Produces drafts, generic or customised | When volume exceeds what reviewers can translate from scratch | Glossary support, style adaptation, data terms |
| Quality estimation and QA checks | Scores drafts, flags placeholders, numbers, terms, length | When you need to decide what gets human review | What it checks for Korean specifically; thresholds you can set |
| Connectors and automation | Moves text between repository, CMS, design tool and TMS | When copying files by hand starts delaying releases | Pull requests rather than direct commits; sync schedule |
| In-context review and testing | Screenshots, live previews, pseudo-localization, device testing | As soon as you ship a UI | Can reviewers see where a string appears? |
The TMS is the hub
Everything else connects to the TMS, so choose it first and choose it for workflow, not for its built-in machine translation. The deciding features for most growing teams are connectors and review. A repository connector should detect changed source strings and send translations back as a reviewable change. Crowdin's GitHub integration, as one documented example, syncs source changes continuously and delivers translations on a schedule, hourly by default, through a pull request from a separate service branch. Its GitHub Action covers the same tasks inside your own workflow, with options to upload sources, download translations and create the pull request.
Then check review. Korean reviewers need to see a string in its screen, with its character limit, to choose between words that a list view cannot distinguish. A TMS that cannot show screenshots or a live preview pushes reviewers back to spreadsheets, and the stack quietly falls apart.
Memory and terminology are the assets
Tools change; your translation memory and termbase should not be lost when they do. Before signing, confirm you can export both in standard exchange formats and that the export includes metadata such as approval status and notes. A Korean termbase built over two years, with decisions about loanwords versus native terms, speech level and product names, is worth more than any single tool. See building a termbase step by step.
Machine translation: configure, do not just switch on
Engines differ in how much they let you steer them. Google Cloud Translation, for example, offers glossaries and an adaptive translation mode that uses your example translations to steer output toward your style. LLM-based tools accept instructions and examples in the prompt. Whatever the engine, connect it to the termbase, test it on your own Korean content rather than a benchmark, and read the data terms. Google's documentation for its Cloud Translation API states that content sent to it is not used to train its translation features; consumer tools may work differently.
Decide what human step follows. ISO 18587, the standard for post-editing machine output, distinguishes light from full post-editing; write down which level each content type gets. Our post on light versus full post-editing covers the guidelines.
Quality estimation and QA
Quality estimation scores drafts so you can route them. Phrase's Quality Performance Score, for example, can be used with thresholds to decide which segments skip or receive human editing. Rule-based QA checks catch mechanical problems. For Korean, add checks that generic rules miss:
- Variables placed directly before a Korean particle, which will be grammatical for some values and not others.
- Numbers converted into Korean large-number units (ten thousand, one hundred million), checked digit by digit against the source.
- Mixed speech levels within one screen or document.
- Glossary terms present in the source but missing in the target.
Testing in context
The last component is the one most often skipped. Pseudo-localization catches hard-coded text and layouts that break when text grows, before any translation exists; Android, for instance, includes built-in pseudolocales for debug builds. Device testing with real Korean text then catches line breaks, truncation and font issues. We show how these fit into a release pipeline in an illustrative SaaS localization pipeline.
The order most teams add things
- TMS with built-in memory, plus a glossary.
- Repository or CMS connector, so nobody copies files by hand.
- Machine translation connected to the glossary, with a written post-editing level per content type.
- In-context review with screenshots.
- Quality estimation and extra QA checks once volume justifies routing.
The people who run this stack matter as much as the tools; see translation ops roles for small teams and the costs in a cost model of a hybrid AI pipeline.
Choosing or fixing a translation stack?
We work inside your TMS as your Korean team, and we can tell you which parts of your current setup are slowing Korean down. See our services →