Unravelling Babel’s tongue: AI within transcreation

To say the pace of development within AI is blistering is a little like saying the surface of the sun is ‘a bit warm’.

Within a matter of what seems like minutes, AI has silently overtaken a whole host of global issues, pirouetted neatly and taken centre stage, becoming a major talking point that suddenly, imperceptibly, has begun to warrant questions, concerns, excitement and an appetite for deeper understanding.

The amount of airtime it has enjoyed is second only to the number of questions posed in its general direction, which have ranged from ‘Will AI take all of our jobs?’ to ‘Can AI learn to respond like a human?’ and ‘Should I be worried about anyone naming a company Skynet, even as a joke?’.

We’ve seen advocates and adversaries, naysayers, game-changers, first-adopters and stalwart rejectors. Speculation, in every avenue, is rife. But the devil has always made a habit of residing within the detail; it’s all too easy to respond to the inevitable question of ‘What can AI do?’ with a beaming ‘What can’t AI do?’. So the lens we’d like to use to focus today is that of language.

Machine translation (MT) engines like Google Translate have been around for some time now, as have more advanced platforms utilising ‘neural MT’, where we’ve seen sequentially better translations generated for a variety of applications sitting firmly within the space of general translation and expanding to reach the fringes of marketing and advertising.

From the perspective of computational logic, it makes a lot of sense: dictionaries plugged into heuristic language engines able to find linguistic equivalents, applying grammatical rules to verbs, declining and augmenting nouns and cases as necessary. The linguistic jigsaw kit has a lot of pieces in the box to complete the puzzle, which it often does with aplomb.

Putting AI to the test

But what about when we get a bit clever? When we feint and riposte, vie and obfuscate, conjure castles out of clouds with words for stone and syntax for mortar? What happens when we couple AI with linguistic engines, point it toward creative and idiom and let it loose with some ‘transcreation’? (Transcreation being the adaptation of concepts, not just words, into another language and culture.)

So, where to start? It’d be all too easy to throw idiom at AI and poke fun at it when it fails to equate it or merge it into a structure that works fluidly. Instead, we wanted to give it a fair chance, to give it the tools available to a copywriter when we provide a brief; structured, specific and indicative of the desired tone.

When there’s a direct equivalent, with the same meaning and temperature in both source and target language, it’s a simple equation.

The setup was straightforward enough – plug in a set of instructions, call out any campaign specifics, particularities of language/culture etc, and wait to see what AI would generate in terms of ingest and output.

In a scenario whereby we have tessellating pieces of language, i.e. equivalence of expression across locales, the AI’s job becomes that of matching source to target. This is something the machine can achieve with proficiency. Providing, of course, that there are no larger cultural nor linguistic obstacles at play.

Our first example comes from copy to accompany an elite snowboarding event, namely the headline ‘Hitting new heights’. The AI output in Spanish (‘Alcanzando nuevas cumbres’) preserved both the literal and figurative interpretations of the source, playing on co-existing concepts of altitude and aspiration tidily. Job done, it would seem…

But what happens when it gets a little trickier?

Linguistic nuance

Take, for example, the verb ‘sizzle’, excerpted from a headline displaying a new cooking oil for a prolific dairy brand. Noteworthy, here, is the function of the voiced fricative (‘z’), intended to typify the sound most commonly attributable to pan-frying. However, when moving into Simplified Chinese, we are presented with AI’s current limitations. Namely, an inability to identify minutiae of nuance and apply it correctly.

The AI’s first attempt rendered an onomatopoeic effect incorrectly reflecting the given context, resulting in a snake’s hiss as opposed to a sizzling sensation. However, in subsequent attempts (with minor change in brief composition and guidance from human expertise), the AI was able to conjure the correct effect.

In the same campaign the metaphor ‘golden touch’ featured prominently, for which the AI raced to find an equivalent expression. But what it found didn’t cover the duality in meaning present in the English, namely the golden, buttery image of the food itself juxtaposed against something wondrous or superb. In fact, the Chinese expression it landed upon, translating as ‘the golden stone’, instead appeared oddly misplaced when used in the context of cooking.

The human copywriter, conversely, actively chose to avoid the ‘golden touch’ metaphor from the get-go, knowing it just wouldn’t do to simply lift and shift it, and instead went for a more straightforward approach, focusing on the fact that Chinese copy in advertising favours a more direct, easily accessible tone, especially for product benefits and key features.

Interestingly, the copywriter accounted for the omittance of wordplay by introducing the concept of intricacy (akin to the ‘golden touch’ metaphor) earlier in the piece. By doing this, they were able to keep the sentiment and feel by flavouring the rest of the copy appropriately, something the instruction-led AI with no strategic human guidance would find impossible to do.

Metaphor is hard enough on its own, but what about when it’s culturally inferred? In a script for an international personal care brand, the line ‘Use our four-step guide to help detox your child’s feed’ seems simple enough surely? Aside from the notion of numerical steps being quite a flexible construct in English, and the AI picking out a well-meaning equivalent in the Czech language that sounded quite strange, there was a bigger problem.

The term ‘detox’ in English holds both positive and negative meanings, but in Czech it has a strong, immediate link to the literal context of food and diet, so the result is rather clinical and lacks the warmth the English tries so hard to capture.

Similarly, when faced with adapting the expression ‘Make waves’ (ie the idea of making a splash as well as the notion of disruption and fuelling positive change) into Russian, the AI struggled again. There is no inherent expression for this in the Russian language, so it simply indexed the closest construct it could, which lost the imagery as well as the metaphor and sounded very stilted indeed.

Unfulfilled promise

So, what determines the machine’s word choice? Why can it hit the nail on the head sometimes but miss entirely on others?

When there’s a direct equivalent, with the same meaning and temperature in both source and target language, it’s a simple equation. But when there is no match, or the master copy is imbued with more nuance or layers that sit outside the basic definition of the word/phrase, it cannot grasp it, nor map its way to a successful adaptation. But the story it tells you when it sells the line back is where we see a significant trick-of-the-light.

In the transcreation world, back-translations and rationales are two core mechanics to illustrate what has taken place during the alchemy of linguistic exchange. The first showcases the mechanical truth and the second demonstrates the truth of feeling – what has shifted creatively to allow the line to come to life – both of which are assembled by expert copywriters to illustrate all the sentiment needed to make sense of (and, critically, to approve) the line.

We found that the rationale provided by the AI to substantiate its language output was overwhelmingly a reformulation of the initial brief. To the layman, it sounds very much like a tick in the box; it sells us on the promise of linguistic credibility, which is all too easy to believe when taken at face value.

Sadly, and far too often, the AI claims to have achieved something that simply isn’t true, and that, without being blunt, is the fact of the matter. It’s this blindfolded oracle – albeit unconscious – that exemplifies the most imminent danger for the world of marketing. What happens when large brand campaigns place blind faith in a promise that doesn’t ring true?

For AI alone to yield anything akin to what a human copywriter would have come up with, you currently have to feed it all the linguistic jigsaw pieces to spell out the answer, warn it away from false idiom and steer it through onomatopoeia to get something close to correct, and that’s with the foresight of a trained, native-speaking copywriter holding the wheel. Many global brand managers will not have the luxury of speaking all the languages within their regional remit, nor the time to feed the line to the machine.

Sadly, and far too often, the AI claims to have achieved something that simply isn’t true.

The pools of data that shore up the internet of things come from a decidedly Anglocentric core, the patterns of which we have seen with word placement and phrase formulation as the system tries to bend foreign idiom around them. As we move into more esoteric language pairs, the corpus of data shrinks and copy options fall into common phrasal patterns.

We saw minor variations in output when switching between platforms (Google, DeepL, Microsoft) with our test material, but all showed the hallmarks of the mechanic we have witnessed in motion. AI builds syntax layers upon referenced constructions, mapping phraseology based on rules, but never upon feeling, and therein lies the rub. Language is equal measures science and fiction combined; it is not a binary equation.

Connection to culture

To learn a language is to pay respect to the culture that birthed it. A language’s past and present form a blanket of reference from which to spin yarn: popular films, quotes, songs of yesteryear, current affairs, historical events; they all feature heavily in the way a language functions today, just as much as yesterday.

For example, many British English expressions have their roots in naval warfare, such as ‘to show one’s true colours’ or ‘the cut of one’s jib’, whereas in Japan the roots are intertwined with swordplay. When you reach reconciliation, you ‘return the blade to the scabbard’ (moto no saya ni osamaru), and when you just don’t get along with someone, you ‘don’t match the blade’s curves’ (sori ga awanai).

As Isaac Asimov said: “In life, unlike chess, the game continues after checkmate.” The extension of the cultural impact on language cannot be downplayed if fluidity is what you are aiming for.

Will systems mature to the point they can discern difference, impact and taste? It’s a deep question with no immediate answer today.

There is undoubtedly a necessary place in the world of today for AI within language. When coupled with the right tiered approach in terms of localisation, AI advancements have already begun to provide time and cost efficiencies for brands – namely within the realms of long copy and general localisation – but when we are looking at layered, nuanced copy which needs to evoke feel and provoke a reaction, that’s a different angle entirely. We must remind ourselves that this technology is very much in its infancy, and we must allow it to develop and mature, testing and learning as we go.

So back to our original question. Will AI take all of our jobs? No, we don’t think it will. Whilst it’s an understandable reaction, pushing AI aside out of fear is not the answer. Instead, the origami trick will be to turn AI into a tool that works for us, not instead of us. Something that augments us, but not replaces us.

As Kyle Reese once said: “The future’s not set, there’s no fate but what we make for ourselves.”

Rik Grant is transcreation partner and Beth Holding is transcreation manager at Tag.

Learn more about at Tag and our global production services here