“The more things change, the more they stay the same.”
—Jean-Baptiste Alphonse Karr, French novelist
Some years ago, I wrote a post for the Our Languages blog about how translation technology has changed the translation profession (opens in new tab). With the advent of even more effective software based on advanced machine learning techniques called “deep learning” and especially the recent arrival of large language models powering popular chatbot and virtual assistant tools, it’s past time for an update.
Optimistic versus realistic views
While public attention has been drawn only recently to the natural language processing capabilities of these programs, which have made headlines for their ability to generate text, images, audio and video on command, translators have been impressed by their capabilities since the first machine translation tools to use deep learning came on the market in the late 2010s.
No doubt, the new machine translation tools produce significantly better translations than their predecessors (statistical-based programs). They create a more idiomatic target text and more accurately reflect the source text. To the untrained eye, the translations these tools produce may be entirely satisfactory, and they can even deliver better results than some untrained or beginner translators. In many more circumstances, then, these new tools do an acceptable job.
Given such results, many are calling these new tools “artificial intelligence” (AI) and even claiming that they show signs of intelligence or consciousness.Note 1 Others argue that what is now widely known as AI is no such thing.Note 2
Translators have realized from their first uses of this new generation of machine translation tools that they aren’t intelligent. While they can seemingly miraculously overcome typos in the original language to translate the intended word, they often use different terms from one sentence to the next. They also still fall into homonym traps. And they continue to produce mistranslations that can invert the author’s line of argument. The AI children clearly still need human adult supervision.
As for large language models, translators use them not only for machine translation, but also to conduct research, find parallel texts and brainstorm potential solutions. For instance, these tools can be used much like a thesaurus, reverse dictionary or idioms dictionary. They can even propose edits to translators’ initial drafts. Indeed, for better or worse, they’ll do more or less whatever they’re asked.
Necessary human expertise
One advantage of machine translation has always been that it saves translators from having to type out a document. Unfortunately, translators often had to retype the translation anyway because the machine output was so unidiomatic or inaccurate. Today, that problem is much less of a worry: it’s often more efficient to let a machine produce a draft and then to revise it.
As always, the more technical the translation, the less likely a machine will provide usable output. So subject-matter knowledge and experience are still required to craft a viable translation.
Correctness and clarity remain essential in most contexts. For example, governments have a duty to provide clear and accurate information. Legal contracts must be written carefully, for disputes can turn on the placement of a comma.Note 3 Media outlets generally want to publish articles that reflect reality. And companies might think twice before letting their chatbots run wild. Just the public relations damage these tools can do is potentially immense, as AI has become a buzzword and stories about its foibles often go viral.
Translators have to watch out for similar problems in all their work. For example, inclusive writing is an important standard that’s being increasingly applied.Note 4 Machine translation tools, because they make pronoun errors or because they regurgitate word patterns from other times or places, sometimes fall short of that standard. Misgendering someone has always been a fear for translators, but now such errors are more insidious—sitting disguised in nice-sounding, pre-written, computer-generated text. Being ready and able to overrule the computer’s decision is essential.
Unresolved issues and prospects
In short, the overall translation landscape has shifted dramatically in some ways but little in others. People and organizations continue to employ the new software in use cases that vary from ideal to inadvisable. An age-old problem remains: the person seeking a translation may fail to recognize whether the product they paid for (or obtained for free on the Internet) is any good, because they don’t speak the other language. As a result, an inaccurate or incomprehensible translation can slip through.
The bottom line is always whether a text is fit for purpose—for the author and for the reader. But because it’s usually the former who commissions the translation, it’s the latter who sometimes suffers the consequences.
With that in mind, do you use deep learning software to translate text or large language models to generate content? If so, how do you deal with the issues they raise? And what do you think the future holds?
