Gains:
- Understanding the concepts of translation memory (TM), fuzzy matches and segments, and being able to use the differences in fuzzy matches by adapting them rather than blindly.
- Ability to apply the discipline of writing only approved translations to TM, keeping customer TMs separate, and performing TM maintenance
- Ability to protect privacy by controlling where data goes in MT/LLM integration within CAT
Professional translation is most often done in a CAT tool, not in a plain text editor. In this unit, you will learn CAT tools, the translation memory (TM) at the heart of translation, how they work together with machine translation and LLM, and how to use this ecosystem efficiently and safely. The goal is to become a translation technology user who manages an entire project, not individual sentences, in a consistent, fast and traceable manner.
Basic concepts
CAT tool (Computer-Assisted Translation) is professional software (such as Trados, memoQ, Phrase, MateCat) that organizes the translation sentence by sentence (segment) and connects the translation memory and termbase. Shows source and target side by side, catches repeats, preserves formatting.
TM (Translation Memory) is a database that stores the source-target sentence pairs you have previously translated. When a new sentence arrives, if there is a similar sentence in TM, the agent suggests it to you. The key concept here is fuzzy match: the similarity of the new sentence to a sentence in the TM (100% = exactly the same, 85% = largely similar). High matches speed up the work because you reuse your previous translation.
A segment is the basic unit of translation—usually a sentence. The CAT tool divides the text into segments and edits each separately.
Importance of TM: MT produces raw drafts, but TM returns your approved, human-quality past translations. The two are different: MT is a prediction, TM is a memory.
Tip: Do not confuse TM and MT. A 100% TM match (your previously approved translation) is generally reliable; MT recommendation should always be verified. CAT tools show the two with different colors/labels; always see this difference.
Integration of MT and LLM into CAT
Modern CAT tools call MT engines and increasingly LLMs internally: if there is no match in the TM, the agent automatically returns an MT recommendation for the segment; you post-edit it (i.e. MTPE flows in CAT). Latest generation tools also integrate LLM: you can do operations like "rewrite this segment with this tone", "correct this term to termbase" etc. in the tool.
Advantage of this integration: TM, termbase, MT and LLM are combined in one screen; For each sentence, the flow "look at TM first, otherwise suggest MT, term QA automatically" is established. Downside and caution: where does the data go? Cloud-based MT/LLM integration can send text to external servers; In confidential work, this may be a breach of contract. Be sure to know which engine the vehicle is connected to and with which data policy.
Feeding and clearing translation memory
The value of TM is as much as the quality of the translations in it. Garbage in, garbage out. Two disciplines:
- Just write the certified translation to TM. If you save the raw MT output to TM without validating it, it will return to your future jobs as bad "memory".
- Perform TM maintenance. Over time, terms change and errors enter. Periodically clean TM, correct worn/incorrect pairs. In this work I'm using the LLM to ask "are there any inconsistencies/term bias in these TM pairs?" You can use it as an auditor.
Caution: Using one customer's TM in another customer's business is both a breach of confidentiality and a risk of term confusion. TMs are kept separate on a client/project basis. A mixed "everything TM" destroys both confidentiality and quality.
three mini cases
Case 1 — TM flew the replays. A manufacturer was having a product family translated where 70% of its manual remained the same year after year. Thanks to TM, 100% and high fuzzy matches came automatically in every new version; Only ~9,000 words of the 30,000 word work were actual new translation. Time and cost were reduced by one third.
Case 2 — Dirty TM propagated the error. One team had saved the raw MT output into TM without verification. A year later the same mistranslation came back again and again, appearing “reliable” as a 100% match; The error was spread across 14 different documents. The team instituted a "certified translation to TM only" rule and a clean-up tour.
Case 3 — Confidentiality leaked in integration. A translator did confidential work in a CAT project that was automatically connected to cloud MT; The text went to an external engine whose data policy is unclear. Once noticed, MT integration for secret projects was turned off and only enterprise engines that "do not store/train data" were used.
Four copyable templates
1) Fuzzy match evaluation (to LLM):
Below is the new source sentence, the similar sentence in TM and its approved translation. Tell me exactly what I need to change to adapt the TM translation to the new sentence; Do not suggest unnecessary changes. Show the difference in meaning clearly. New source: [...] | TM source: [...] | TM translation: [...]
2) TM consistency check:
Below are the source-target pairs from TM. Mark inconsistencies and term deviations where the same or very similar source sentences are translated DIFFERENTLY. Just list the problems.Couples: [...]
3) Segment tone rewriting (LLM in CAT):
Make the following target segment [desired tone] WITHOUT CHANGEING the meaning. Comply with the list of terms: [...]. Keep numbers and names. Export only the rewritten segment. Segment: [...]
4) Privacy/integration pre-check:
I will use MT/LLM integration in a CAT project. I will describe the type of text in this project; Tell me whether it is appropriate to send this text to a cloud-based external engine, what questions (data retention, training, location) I should ask the provider.Text type/privacy status: [...]
Weak prompt / Strong prompt
Weak: "Use the translation in TM." (It is unclear what match rate and what to adapt; blind copying introduces errors.)
Strong: "This new sentence matches the sentence in TM 90% of the time. Build on the TM translation but reflect this difference: the source has 'annual' instead of 'monthly', so it should be 'annual' instead of 'monthly'. Don't change anything else."
Difference: strong prompt pinpoints the match difference; It prevents "leaving the old translation as is", which is the most common mistake of fuzzy matching.
CAT ecosystem components table
component
What does
relationship with AI
TM (translation memory)
Returns approved past translations
Reliable; precedes MT
Termbase
Ensures term consistency
It is given as a prompt to LLM
MT recommendation
Raw sketch into empty segment
Must be post-edited
LLM integration
Tone/term/rewriting
Controlled, attention to privacy
QA module
Scans number/term/tag error
automatic control
Common mistakes
- Blindly accepting the fuzzy match. An 85% match can hide a 15% difference in meaning.
- Writing raw MT output to TM. Dirty TM carries the error into the future and proliferates it.
- Mixing customer/project TMs. Confidentiality and risk of term confusion.
- Not knowing where the integration data goes. Hidden text may leak out without you being aware of it.
- Forgetting the TM/MT difference. MT is an estimate; TM is a memory. The two are not equally safe.
Pre-translation and alignment
Two advanced CAT functions further accelerate workflow in the AI era. The first is pre-translation: at the beginning of the project, the tool automatically fills all segments with TM and MT; Instead of an empty file, you start with a file where 100% matches are approved, fuzzy matches are waiting to be adapted, and empty ones are MT drafts. This changes the job from “writing from scratch” to “reviewing and editing” — but you need to evaluate each segment rather than blindly approving the pre-translation.
The second is alignment: if you have a source-target document pair that has been translated before but has not entered the TM, the alignment tool matches them segment by segment and converts them into TM. Thus, your past translations are added to the "memory". Caution in alignment: automatic matching is not always correct; one segment slippage disrupts the entire alignment. Reviewing aligned pairs before TMing prevents the risk of dirty TM in the first place. These two functions turn your current assets (past translations) into investments in future businesses.
In summary
CAT tools; It combines translation memory, termbase, machine translation and LLM into a single production line. TM is a memory that retrieves your approved past translations and provides great speed in repetitive tasks; MT is a prediction that always needs to be verified. The savvy user carefully adapts the difference in fuzzy matches, writes only approved translations to the TM, keeps customer TMs separate, and protects privacy by knowing where the data goes in the integration.
Application task
Set up a small project in a CAT tool (a free online CAT also works): add a termbase and an empty TM. Translate a 10-sentence text, save the approved translations to TM. Then translate a second text very similar to the first text and adapt the fuzzy matches returned by TM with the "fuzzy match evaluation" template; Note how many sentences you would make a mistake if you blindly accepted TM.
checklist
- [ ] I connected TM and termbase to the project.
- [ ] I used the difference in fuzzy matches adaptively rather than blindly.
- [ ] I only wrote to TM the certified translation that I have verified.
- [ ] I kept customer/project TMs separate.
- [ ] I checked where the data goes in the MT/LLM integration.