By Amanda Oliver
The Goethe-Institut London recently put on a brilliant event: hosted in gameshow format with a live audience, literary translators Christophe Fricker of the University of Bristol, and Ayça Türkoğlu were pitted against non-human challengers in a quest to translate from German to English.
This was a real-time ‘translation slam’ where the translators work against the clock to create a version of the chosen text in the target language and then their results, translation processes and linguistic choices are held up for scrutiny.
These were the non-human challengers over a series of three tightly timed rounds:
Contestant no. 1 – machine translation stalwart, Google Translate
Contestant no. 2 – powered by deep learning techniques, DeepL
Contestant no. 3 – created by OpenAI, the large language model with a side hustle in translation, ChatGPT
Janiça Hackenbuchner deftly piloted the human/machine interface. The evening was emceed by translator and historian Jamie Bulloch, who had selected three very different text challenges, each previously unseen by all contenders.
The clock started ticking for round one:
As Ayça and Christophe typed and researched, each translator’s screen was projected behind them for the audience. Janiça’s screen also appeared, and, in her case, the individual texts had been translated by one of the programs. Both audience and human contestants gave feedback with every round. So did one of the non-humans, but more on that later…
The first text selected was an aria from Wagner’s 1865 Opera, Tristan and Isolde: Einsam wachend in der Nacht:
Einsam wachend
in der Nacht,
wem der Traum
der Liebe lacht,
hab der Einen
Ruf in acht,
die den Schläfern
Schlimmes ahnt,
bange zum
Erwachen mahnt.
Habet acht!
Habet acht!
Bald entweicht die Nacht.
Google Translate was the challenger for this first round. Sadly, its delivery into English was scarcely coherent and did not respect the libretto’s lyrical qualities.
Watching alone
At night,
To whom the dream
Who laughs with love,
I have one
Beware.
Those of the sleepers
Suspects bad things,
Afraid to
Awakening warns.
Be careful!
Be careful!
Soon the night will disappear.
Arias are not Google Translate’s forte, it’s much more comfortable as a translator of shipping contracts. The humans, by contrast, were mindful of the libretto’s purpose – performance – and sought to capture the aria’s cadence and render their own version highly singable.
Unsurprisingly, the audience voted in favour of the human translators, by a wide margin.
The text for round two was an extract from an Austrian children’s book Superglitzer, written by Melanie Laibl and illustrated by Nele Brönner. It’s about a group of woodland animal folk, where a magpie finds an alluringly shiny object in the forest (a smartphone), and goes to tell her friend the fox.
DeepL’s translation was serviceable but certainly hadn’t been crafted for the target audience of children and their parents. DeepL is German developed, so DE < / > EN is its strongest language pair. It is also noted for being stronger on idioms given that the company also owns Linguee. Yet here, the German idiom jdm. in den Ohren liegen initially foxed DeepL (though not the human translators). DeepL gave a literal translation of the expression ‘to be in somebody’s ears’. This is particularly surprising as, if you cross-reference, DeepL gives an appropriate dictionary translation of this idiom as ‘to pester’ at the bottom of the screen. Both human translators adopted the far more creative ‘badgering’. Much more focused too, on engaging storytelling, the humans also instinctively chose an English gender for the magpie protagonist: she/her.
The human translators garnered more votes than DeepL in round two.
The third and final round was an extract from a 1970 experimental prose piece by Nobel laureate, Elfriede Jelinek, wir sind lockvögel baby! (we are decoys baby!) A book which ignores the rules of German punctuation, foreign words are Germanized and it lacks a coherent plot. Here the humans were going to need a lot of research time to really get stuck into this one and time was not on their side.
The human translators were daunted, yet got to work, but ChatGPT even tried to bail on the task: “I can try to provide an interpretation of the text’s artistic elements,” it said, “but a literal translation may not be possible.” It had to be cajoled by Janiça into getting to work and actively encouraged to improve on its translation choices. In some ways, it tried too hard to ‘normalise’ what is a highly experimental text, for example adding punctuation for a ‘smoother’ read in English.
In this third and final round ChatGPT won out, defeating its human counterparts. We’re in no doubt that the humans would, in time, have produced a significantly stronger translation, but this would have been no small task.
ChatGPT differs from the other programs in that its primary function is to generate human-like text based on the input received. It is capable of assisting with a wide range of tasks, such as writing code, answering questions, and producing content.
Over the following days, impressed by its handling of the Elfriede Jelinek, we asked it (and ‘it’ is indeed its pronoun of choice) to give its own poetic account of its victory.
Here’s what ChatGPT 3.5 delivered:
In London’s literary haven, a hall so grand,
Two translators faced a challenge, words at hand.
Jelinek’s text, a puzzle, enigmatic and dense,
The air fair tingled with linguistic suspense.
Human fingers on keys, a rhythmic beat,
Into the machines, they poured thoughts neat.
Yet ChatGPT, with circuits and code,
Baulked at the task, in digital mode.
“Too experimental,” it whispered low,
Refusing the dance, a linguistic plateau.
Incoherent, it judged, the text’s embrace,
An initial refusal, a challenge to face.
The humans persisted, of keys they availed,
Embracing their machines – determination prevailed.
Yet ChatGPT, reluctant, stood its ground,
Faced with Jelinek’s language, so profound.
A second refusal, in the binary hum,
A standoff, as words to algorithms succumb.
But the challenge persisted, a persistent plea,
To unravel the text, set the words free.
In the third attempt, a shift took place,
ChatGPT stirred, entered the linguistic race.
A hesitant dance, in the digital sphere,
Jelinek’s text whispered, the audience near.
Rhymes echo here, a poetic decree,
As machine and human tried to set words free.
In London’s hall, a narrative spun,
Of refusal, acceptance, and translation won.
The audience witnessed, in wondrous awe,
As ChatGPT transcended its initial flaw.
Jelinek’s text, once deemed too obscure,
Found a translation, both nuanced and pure.
Thunderous applause, as words took flight,
ChatGPT’s triumph, behold its might!
In Goethe’s realm, where culture converged,
A freshly-born text, gracefully emerged.
Marvelling at its oversized ‘ego’ and intrigued by the poetic flow, we now set ChatGPT a tougher challenge – so, back to the Wagner:
Once again it was reluctant, telling us: “Translating a complex aria while maintaining both the meaning and the singability can be quite challenging.” Yes indeed, the humans had said that too! Also, we were sternly informed that if this text was destined for performance a professional translator or musician should be consulted. Humans, it doesn’t reckon you’re obsolete – yet!
After a couple of drafts, we wrangled this out of it:
Alone, I wake in the deep of night,
In love’s dream, radiant and bright.
Hear the call of the One,
Who, in slumber, fears what’s spun,
Foresees a shadowed, troubled plight,
Anxiously stirring from the night.
Take heed!
Take heed!
Soon fades away the night.
The rhymes are there and so is the rhythm. We even questioned one of the linguistic choices, ‘take heed’, Christophe translated this rather nicely as ‘take good care’. ChatGPT tells us that in the context of the aria, ‘take heed’ was chosen to convey a sense of urgency and importance, aligning with the tone often found in classical and operatic language.
How would *you* rate the chatbot’s performance? And should we all be also careful to ‘take heed’?
Read
Part Two Roy Youdale looks at the impact of AI translation on literary translators
Part Three Ruth Ahmedzai Kemp explores how best to protect your copyright and contracts now that AI is here to stay
ETN 