From TMS To AI – The Next Evolution In Translation Workflows

Is AI a terrifying dragon or a tool to be mastered in translation?

Have you ever felt like you’ve been fighting a dragon in the last few years? I know I certainly have. Whether it’s AI, automation, or shifting client expectations, it’s easy to feel like the industry is under siege by a dragon or 10. But what if instead of an uphill battle, we could ride that dragon?

We’re probably no strangers to doomsday predictions about our industry - we seem to get a fresh crisis every decade or so. TMS-es, machine translation, and now, LLMs and generative AI. They were all heralded as the thing that will destroy our industry, and yet, we are still here, aren’t we?

An article by Zalán Meggyesi, Chief Solutions Engineer of easyling.com

Fortress to Firestorm

Since the early ’90s, when translation memories (TM) were invented and incorporated into mainstream CAT tools, TMs were considered to be our rock, our fortress that would secure our position in the world of translation.

Fortress of the old world

• Meticulously-crafted TMs
• Hand-curated glossaries
• Solid, boring first-pass MT, post-edited by human linguists

Translators relied on meticulously-crafted translation memories, drawn from years of cooperation with their clients; on hand-crafted glossaries that were curated painstakingly by production teams, and on revisions of initial translations to ensure alignment – all the things that made the output great in quality and expensive in time and labor. Essentially, for the last thirty or so years, translation has been a tranquil city. But now, there’s a dragon outside the walls.

Firestorm of the new era

• “LLMs are coming for us!” – not wrong, but an oversimplification
• Fears of losing the human touch are justified

For the last 2-3 years, LLMs have been getting really good, good enough to actually start being useful in many professional contexts. Translation is one of these contexts, since LLMs can interpret meaning and recreate the message by passing it through their intermediary layer. Of course, while this is very powerful, LLMs still fall short of actual human translators where nuance like cultural highlights or wordplay is considered.

What can we do about the dragon?
Run / Fight / Ignore?

So what do we do? Do we run and abandon localization to the machines? Probably not. Do we just ignore the dragon and hope it goes away? Maybe. This sounds like a great way to get burned sooner or later, though… Do we fight it and put shackles and restrictions on AI and its use? Perhaps, but in the end, I suspect this only ends up hamstringing us rather than securing our position.

Taming the Dragon - harness the power of AI

What if we were to put a yoke on the dragon and have it work for us rather than against us? Now that sounds like a plan! Instead of fighting it or running from it, we can harness its strength to make our workflows faster, smarter, and more scalable. So, how do we go from theory to practice? How do we actually train the dragon to work for us? Let’s break down a proposed workflow that has the potential for savings of 90%.

Smarter workflow setup

• Streamlines setup for pre-translation
• Reduces later workload
• Improves quality

Traditional pre-translation workflows rely on manual glossary creation, supplemented by statistical document analysis, which is nevertheless still a time-consuming task. However, it is vital for maintaining brand voice and consistency. Automating this to be part of the intake process allows the workflow to take a running start and speeds up all subsequent steps. Providing domain-aware pre-translation via specialized LLMs accelerates the final curation and creates a high-quality translated approved glossary faster than traditional human methods.

AI-powered translation efficiency

• Very fast first-pass translation
• AI-powered quality evaluation of first-pass content
• Post-editing suggestions
• Review based on LLM scoring
• Allows linguists to contribute where impact is the greatest

Machine translation (MT) has a huge advantage: it’s fast, blazingly fast. This means we can create a first-pass translation quickly, but that still needs to be second-guessed. This should even be done in a language by language basis. We recently found that for some language pairs, especially low-resource ones, LLMs can produce translations that outrank traditional MT engines.

This pre-translation is followed by an LLM-powered quality evaluation engine (similar to xCOMET) that ranks all translated content and provides feedback and notes for linguists to consider during review. This ranking allows the production team to focus on the part of the content where a human touch can have the greatest impact on the quality – after all, improving the bottom 10% is much more impactful than tweaking the top 10%.

Human expertise, AI assistance

• Augment, don’t replace
• Reduced load on linguistic team leads to higher quality
• Focus on the greatest impact

Human linguists are still very much necessary: while AI can understand context, evaluate quality, and suggest alternatives, it can’t understand true meaning the way a human can. However, it can handle the simple edits, where a human touch doesn’t add much value, freeing up people for the more meaningful work, like high-impact content, tone refinement, and final validation.

So far, we’ve looked at how AI can be harnessed and how to put a yoke on a metaphorical dragon instead of letting it burn the city to ashes. But here’s the real question. Does this actually work? Is this just theory or are we seeing some real tangible benefits in real world projects?

The AI-powered website translation workflow

Challenges

• 1,000,000+ words of highly technical content
• 32 languages (including low-resource ones)
• frequent updates and dynamic content made manual translation expensive
• tradition MT engines lacked the nuances needed for accurate tone and style

Results

• 11% of the total content had to be touched by a human
• 90% cost savings

The AI-powered translation workflow can be the deciding factor between a project reaching orbit or failing to launch. We recently published a real Brobdingnagian project, featuring low-resource languages like Croatian, Thai, Bulgarian, or Hungarian; where traditional workflows would have been prohibitively expensive due to the amount of content involved. Furthermore, the requirements included specific style guidelines that traditional MT would not have been able to match.

With Easyling’s AI-powered workflow, however, we could

• prepare the initial glossary candidates automatically in less than an hour,
• handle the translation via LLMs utilizing the client-approved glossary and the style guidance developed in conjunction with our client, and
• only sent content for human review that the evaluation model flagged as “needs post-editing” - as a result, barely 11% of the total content had to be touched by a human.

That’s the difference between clutching a contract and losing a client.

The way forward

Ride the dragon, don’t fight it

• Yes, AI is here to stay
• No, translators are not being replaced
• The future is brighter with AI
• We do more meaningful work, not less

There’s a saying that’s been going around for a while, “This is the dumbest AI will ever be”. I’m inclined to believe that – AI and LLMs are here to stay. But this doesn’t mean that human translators will be replaced. In fact, if anything, we’ll need more translators to keep up with the expanding market. What we do believe is that AI will free our time up to do more deep work, more of what’s impactful, to soar above the land rather than get stuck at ground level.

With most customers demanding native language experiences, the volume of content requiring translation is increasing globally. This rise in content naturally leads to a heavier workload for our linguists. To manage this ever-growing demand, we must accelerate our processes while simultaneously maintaining, or ideally enhancing, linguistic quality and cultural insight.

Ethical Translations [closing remarks]

Of course, our job is not finished when it comes to spot-checking and proofreading AI translations. There’s a lot more to be done for the development of autonomous systems, from collating and labelling quality training data, often specialized for a single client, through ensuring the fine-tuned systems are free from systemic biases, all the way to making sure our business leaders communicate honestly and transparently about the use and capability of AI systems to foster trust in the output and contribute to the most important goal of localization: building human connections.

This article is an extract of a GALA webinar, the recording is available here.