top of page

THE ROLE OF TECHNOLOGY IN TRANSLATION

BY MICHELLE BULDOC

 PROFESSOR AND DIRECTOR OF TRANSLATION STUDIES

As a translator and as director of the MA in Translation Studies, one of the most frequent questions I hear is about Artificial Intelligence and neural machine translation. “Aren’t you afraid that machines will replace translators?” someone, usually a prospective student, will ask. Or, from the person sitting next to me on a flight: “Isn’t machine translation a more efficient and cheaper way to get translations done?” And with Google’s new Translatotron unveiled just this month, I anticipate the questions will become more frequent.

 

Besides sounding like something out of the Matrix series or some other sci-fi flick, Google’s Translatotron is heralded as a direct speech-to-speech sequencing system. Rather than using the three-step process of 1) converting speech to text (using automatic speech recognition), and then 2) using machine translation (such as Google Translate or DeepL) to translate the transcribed speech, and then 3) converting the text back into speech (via text-to-speech synthesis), Translatotron skips the middle step entirely. Text-to-speech translation apps (like TripLingo) already exist, but the reason techies are so excited about Translatotron is because of the quality of the speaking voice, which although still synthesized, includes the hesitations and repetitions of natural speech as well as the speaker’s tone.

 

In short, with Translatotron, you can translate what sounds like your voice into another language. And this has techies a-shiver. Writers from The Silicon Review (17/05/2019) declare that “Google has undertaken a new one-of-a-kind project that can help lower the language barrier.” Stephen Johnson on the BigThink.com (24/05/2019) asserts that “Translatotron could soon make foreign-language interactions run more smoothly.” And AJ Dellinger, from engadget.com, begins his article on Translatotron (15/05/2019) with the heady proclamation that “Speaking another language may be getting easier.”

 

But are the translations good? Will neural machine translation and Translatotron put translators and linguists out of a job?

 

I am not so much a Luddite (although I still prefer my reel lawn mower to an electric version) that I do not recognize the utility of neural machine translation and even of Translatotron. Not only can neural machine translation translate large quantities of material and quickly to boot, but because it uses AI, it can also create its very own algorithms based on the response it gets from us on its prior results. In a sense, it automatically improves its translations. The European Commission, the largest employer of translators in the world, uses a neural machine translation system known as the eTranslation service for its policy documents. Even the French “Association pour la promotion de la traduction littéraire” (Atlas), whose focus is on literary translation, will be putting into place an “Observatory of automatic translation” at its upcoming meeting in November.

 

And Translatotron? Translatotron can even mimic the speaker’s tone, which means that it can point out the speaker’s emotional state. And Translatotron’s translations will undoubtedly improve because, like any neural machine translation programme, it depends on algorithms that become more sophisticated as the quantity of data increases.

 

Translatotron may be, at least initially, convenient for some telephone conversations (upon which Google engineers based their research) and for tourism; neural machine translation is especially useful when translating in a context in which phrasing is formalized and meaning is unambiguous.

 

So what do we human linguists and translators have over neural machine translation and Translatotron? I think it can be boiled down to three concepts: quality, understanding, and context.

 

Neural machine translation programmes cannot interpret or even engage with ideas, particularly when they are ambiguous or complex, nor can they produce translations that are attuned to cultural context. Even in disciplines in which expression may be relatively straightforward, humans are still needed to ensure that the translations produced by neural machine translation make sense. In fact, one of the developing career paths for translation graduates these days is in machine translation post-editing.

 

The same is true of Translatotron. As a recent editorial (17/05/2019) in The Times has pointed out, Translatotron “can’t strictly do translation at all. Instead, it works by taking gambles on statistical probabilities.” Translators, especially those with formal, Masters-level training, do not gamble on how to express meaning in a different language. They have developed the skills and understanding necessary to make sense of ambiguous, complicated ideas, particularly in texts from the humanities and the social sciences. Translators are trained to recognise how these ideas are bound to a culture, time and place; what’s more, both translators and interpreters know just how in tune they must be with their readers and their listeners to convey the nuances of the original.

 

As sophisticated as they are, neural machine translation and Translatotron are in the end only tools. They may be able to generate a selection of possible translations, and will (again, based on statistics) select one that may be probable, but it may not be correct or even accurate. A qualified interpreter will be your best ally in any kind of important negotiation, political or business; a translator trained in the particular linguistic and cultural domain in which you would like to disseminate your ideas or your research will ensure that they are conveyed precisely and accurately.

bottom of page