AI translation: how to train ‘the horses of enlightenment’
‘Translators are stage horses of enlightenment,” the poet Alexander Pushkin wrote in the margin of one of his manuscripts. Two centuries later, the political scientist Steven Weber similarly compared translation to transportation: not of people and goods but of ideas and knowledge. Just as the world swapped horses for mechanical means of transport, multilingual communication has accelerated too – and now, with the use of AI tools, translation can happen faster than ever.
But faster doesn’t always mean better – the use of AI comes with various risks. This week the European parliament adopted the Artificial Intelligence Act, the world’s first comprehensive piece of AI legislation. It requires developers to be transparent about the data used to train their models, and to comply with EU copyright law.
Meanwhile in the UK, the Artificial Intelligence (Regulation) Bill is scheduled for a second reading in parliament next week. Literary translators, like other artists, continue to campaign for proper recognition. It is vital that we are heard.
Since the adoption of neural networks in 2015, translation algorithms have greatly improved. Academic publishers have been using AI since 2018, when Massot éditions released a French version of Deep Learning by Ian Goodfellow, Yoshua Bengio and Aaron Courville, translated from English by DeepL. Trade publishing soon followed suit. The next frontier for the technology to cross is literary translation.
A unique creative practice, literary translation has always involved myriad choices. This word or that? Fidelity or freedom? Paper or screen? Dictionary or database? The advent of AI introduced another dilemma: should a text be translated from scratch, or could it be run through AI translation software first?
Unreadable translations of pirated books flooding the internet show that the process cannot currently be fully outsourced to computers. Instead, publishers have begun to use AI-assisted translation. Some European presses work with Nuanxed, which employs humans to edit machine-translated books. The company aims “to maintain the quality of traditional translations”, offering savings to publishers and “market rates” to linguists.
The threat of being undercut by machines is an ongoing concern for translators. In its recent survey, the European Council of Literary Translators’ Associations recommends that professionals avoid editing AI-generated texts or charge translation rates for such work. In our age of jet-propelled information, translating by hand could become the new green travel; the appeal of sustainably translated books would then grow, hopefully benefiting their creators.
When asked about quality, even some of the proponents of AI believe it may be good enough for a potboiler but not for a poetry collection. Kristoffer Lind, the CEO of Swedish publisher Lind & Co, told me that his company only uses machine translation for genres such as crime and romance. While the suitability of AI is often debated in economic terms, for Eva Ferri, the publisher of Europa Editions UK and the Italian press Edizioni E/O, the choice is ethical. “Hiring a human being is the right thing to do,” she says, “even when the alternative is much more cost-efficient.”
Like many of my fellow translators, I have no interest in farming out my literary projects to AI. Mark Polizzotti, who has translated more than 50 books from French, including works by Gustave Flaubert, André Breton and Raymond Roussel, spoke for most of us when he told me: “I do the work not so much to express as to discover what I think, where it will take me. By doing this work for me, AI would remove the aspects I most value, and I’d rather poke along at my analogue pace and come out the other end with an increased knowledge of both the work and myself.”
Related: ‘Translation is an art’: why translators are battling for recognition
Meanwhile Nichola Smalley, who translates Swedish and Norwegian fiction, had three deadlines on the afternoon we spoke. Still, she says, “There’s no point in going faster and faster – that’s not what literature is about.”
Sometimes, though, it is hard to reach places poorly served by technology without speeding things up. “I can imagine scenarios where AI helps translate from languages for which there aren’t a lot of translators available,” Ferri says. However, most AI tools “will not have enough data on hand” to translate from or into such languages, Anton Hur, a translator of Korean literature, points out. Attempts to remedy that situation include a model being developed by researchers at the University of Massachusetts. Designed for anyone to contribute to its training dataset, it is expected to “aid literary translators in sharing more minority voices”.
How do we see our relationship with technology evolving? Tim Gutteridge, who translates literary and non-literary texts from Spanish, believes that in both cases using computer-assisted translation can be a “reasonable decision”, and if tools include AI features that give us more choice and control over the results, so much the better. Edwin Frank, the editorial director of New York Review Books Classics series, calls the work of human translators “crucial”, given that the advance of AI is “absolutely inevitable”.
“I have no principal opposition to the use of [such] tools,” he says, “any more than I do to the use of screwdrivers.”
When translating, we need all the tools we can get. Referencing texts created by humans and processed by machines might be a new way of consulting existing translations. AI could be regarded as another vehicle allowing us to navigate the multilingual expanses of knowledge. But having this option shouldn’t prevent us from travelling at our own chosen speed.
• Anna Aslanyan is a journalist and translator, and the author of Dancing on Ropes: Translators and the Balance of History
