Translations + AI: The Italian innovation that’s changing the game

Artificial intelligence is revolutionising the world of translation, yet Machine Translation is not always reliable. Language Check, an Italian innovation, aims to spot AI translation errors, changing the rules of the game.

AI can be seen as a shining sun of opportunity.

There are huge expectations surrounding artificial intelligence. Undeniably, AI will soon transform every profession, particularly having a profound effect in the translation sector.

Agencies and freelance translators look to artificial intelligence as a threatening cloud looming on the horizon. Tools such as DeepL or Google Translate can produce significantly qualitative translations in a short amount of time. Though human translators are increasingly impacted by falling per-word rates, AI can be seen as a shining sun of opportunity as it can be used to improve their work, rather than hinder it.

If the text is less complex, translators can rely on AI to produce an initial draft and subsequently focus on editing. The work isn’t disappearing so much as changing shape, as, in theory, post-editing requires less time than a translation from scratch.

A natural question therefore arises: Why are translators still needed if AI can produce more efficient results? When translating brief documents or business correspondence, this question may ring true. However, professional translators continue to remain essential in contexts where one cannot afford to make mistakes. For legal contexts, technical manuals, and pharmaceutical instructions, a mistranslated text is simply unacceptable.

If we turn to the BLEU score, which measures the quality of machine translation, very clear data emerges:

The BLEU score shows that Machine Translation (MT) is generally reliable, with scores above 60% (varying widely by language pair and domain). However, MT performance drops drastically when it comes to translating technical documents. This perfectly illustrates why human translators remain fundamental in certain fields. Technical, legal, pharmaceutical, automotive and mechanical sectors cannot rely on the 45% accuracy rate offered by MT.

As aforementioned, artificial intelligence has further contributed to a decline in per-word translation rates, trapping professional translators in a dilemma: how can they maintain quality at ever-lower prices?

There is another often overlooked issue tied to the use of AI in translation. Even if MT systems reach 80% accuracy, that leaves 20% of errors to be spotted. It is challenging for translators to spot these errors as machine translated texts tend to flow naturally, making errors harder to detect as the fluency maintains linguistic coherence.

In such conditions, revision becomes difficult and time-consuming, as the translator must tediously go over the entire document. This is not, however, a dead end.

Enter Language Check, a unique Italian innovation, developed by Aqrate.

“Having worked in the translation world for many years, and considering how AI is changing the sector,” says Marco Baglioni, CEO of Aqrate, “We created a solution to facilitate the work of agencies and translators.”

The idea took shape after CTO Michele Fariselli encountered a research paper by Christian Federemen, a researcher who worked in the Microsoft Translator team, and Tom Kocmi, a mathematics and physics professor at the University of Prague.

“I was struck by the data showing that, while AI can make mistakes or hallucinate when translating, it is much more effective at checking its own output,” recalls Fariselli.

The paper showed that ChatGPT-4, when used to review machine or human translations, reached 90% accuracy. “At that point,” continues Baglioni, “we had the idea to create a tool for translators, a system where you could upload a translated document, whether by MT or a professional, and receive a detailed report flagging mistranslations, terminology drift and suggestions for improvement.”

It may seem counterintuitive, as the same machine that commits the errors is also able to detect and correct them, but it’s one of the many mysteries surrounding AI.

As AI pioneer Geoffrey Hinton has put it, we often observe what these models can do before we fully understand why they do it:

“We know what happens, but we don’t know why AI makes a specific decision. For example, if we ask a physicist what happens when a leaf falls from a tree, he can explain it depends on wind, angle, and other factors. However, even with this knowledge, he still cannot predict the exact spot where the leaf will land. That’s where we are with AI.”

At present, there are no other tools like Language Check on the market, capable of providing such a large amount of useful information to translators and concretely facilitating their work.

Translators can increase productivity by combining MT to create a first draft, and Language Check to identify errors and make corrections (using the suggestions), without having to go over the whole document.

Aqrate plans to showcase the much-anticipated Language Check in the United States at two major industry gatherings: LocWorld in Monterey and ATA 2025 in Boston.

__

Sign up for our newsletter to get the latest innovation news and more sent to your inbox.