24 April 2025
Neglected translation memories undermine AI
The translation industry is at a pivotal juncture. We are all grappling with how to integrate artificial intelligence (AI) into our workflows. But few translators – and even fewer clients – want to start from scratch. After all, years of hard work have gone into creating the bedrock of translation: our beloved translation memories (TMs).
For years, TMs have been the comforting digital quilt we wrap ourselves in – a patchwork of past projects, client-approved phrasing, and various reviewers’ takes on things. There’s a romantic notion that TMs are golden repositories of linguistic excellence, carved in stone. We’ve hoarded them, treasured them, like a collection of fine wines in a vault.
But here’s the thing: TMs are not fine wines. They don’t improve with age.
Since TMs are repositories of language, they must be dynamic and go with the times. Terms fall out of use, expressions become outdated, in many fields English terms take over. For example, who in German-speaking countries still refers to a button on a user interface as a “Schaltfläche”? Or in France, who says “Zone de défilement” when “scroll box” rolls off the tongue so much easier?
But it’s not just the IT industry where globalization has brought about the anglicization of terminology. In the automotive industry, “dashboard” has largely replaced “Armaturenbrett” in German. “Tempomat” and “Régulateur de vitesse” are giving way to “cruise control”. Even “Luftsack” and “coussin gonflable” are disappearing in favour of “airbag”.
The printing industry has seen similar shifts. “Offset printing” used to be “Flachdruckverfahren” in German and “impression offset” in French, but the English term has now wiggled its way into both languages. Slogans are often left in English *Levelling up your Labelling”, “Inkjet-Drucker”.
And it’s not just terminology. Entire phrases and concepts have evolved, especially in sales and marketing. “Sales funnel” used to be “Verkaufstrichter” in German and “entonnoir de vente” in French, but these terms are falling out of favor. A “lead” was once an “Interessent,” but now it’s just a “lead.” A “pitch” is no longer a “Verkaufspräsentation” or “présentation de vente” – it’s simply a “pitch”. And a newsletter is a “Newsletter”, “Product news” is “Produktnews”.
In the world of banking and finance too, English terms have become prevalent. Think of Anti/Monez Laundering (AML) as opposed to “Geldwäschebekämpfung”, or Compliance, Cashflow, Return on Investment (ROI), Due Diligence, Initial Public Offering (IPO), Fintech, to name but a few.
We take TMs for granted, recycling their contents endlessly without thinking twice. If we’re brutally honest, we haven’t really curated them. Certainly not in the way we originally intended (“Let’s only clean files into the Master TM that have been carefully reviewed and approved by the client”). There just hasn’t been the time. No one wanted to pay for proper care. Sure, we’ve all made the odd valiant attempt to update a segment we know is wrong, only to be met with resistance, sometimes among our own ranks, sometimes from the client side: “But this was approved by the then marketing expert in-country 12 years ago – we can’t suddenly change it!” And that of course is often true. Because what we are working on are just updates of the manual or the brochure. And these have to blend in with what’s already there!
The result? TMs have become a liability rather than an asset. If we don’t address this, our machine translations will be stuck in a time warp. AI will churn out words by the million, but they won’t reflect how people use language today. Instead of being a tool for efficiency, it will become a source of frustration.
It’s a common misconception to blame poor translation quality squarely on AI. While AI tools are not infallible, the root of many translation errors lies in the human-generated data they are trained on. If the foundational data is flawed, AI systems will mirror those imperfections. In many cases, the blame lies squarely with the lack of rigorous TM and TB maintenance.
It’s important not to fall into the trap of believing that “more data” is necessarily “better AI”. Quality beats quantity here, too. If you want your translations to be up-to-date and relevant to your corporate branding, they must reflect modern usage, current terminology, and your current tone of voice.
Postponing the curating and improving of legacy material until after machine training is a recipe for frustration and extra (human) effort. It’s far better to work with a smaller, high-quality dataset than to rely on massive volumes of outdated segments.
Before launching into glorious AI-driven translation workflows, it’s absolutely essential to do a full spring clean of your TMs. This means:
For example, some companies are shifting from a formal to an informal tone. This doesn’t just mean changing pronouns like “Sie” to “du” in German and adjusting verb forms – it also requires adopting a more casual tone of voice and simplifying sentence structures.
Improving translation quality is a collective responsibility. Language service providers, translators, and client organizations must work together to prioritize the spring-cleaning of TMs. Give them a good dusting, take them to the dry cleaners, and be prepared to discard “historic” entries.
Investing time and resources into this process not only benefits AI training but also supports human (or machine) post-editing efforts. Needless to say, terminology databases (TBs) must undergo the same rigorous overhaul.
It seems to me that one of the most important activities right now is cleaning, curating, and sanity-checking the data used for training AI.
The challenges faced by the translation industry today are not just the result of AI’s breathtaking advances – they’re also the consequence of years of neglect, or perhaps we should call it naivety.
The future of translation depends on whether we can get our act together and treat training data like the precious resource it is, not like some bottomless junk drawer. Feeding machine translation workflows with outdated, recycled segments will only lead to irrelevant output and wasted effort.
Translators shouldn’t have to spend their time fixing machine output based on bad decisions from 20 years ago. If we are to embrace automation and produce relevant, high-quality translations, we must finally buckle down and tackle that long-overdue spring clean.
Click button bellow to apply to work as a freelancer.
Apply as a freelancer
