Deverbalisation: what AI is not able to understand

THE IDEA AND THE WORD

The translation process necessarily goes through deverbalisation, which is probably the most important mechanism for human translation. Deverbalisation means taking apart the source text and working with the ideas in it, for which we must search for the right words that convey its meaning. When we work with similar languages, like romance languages or even English, this process tends to go unnoticed, as the linguistic structures are often similar. The bigger the difference between the source language and the target language is, the more clearly the weight of deverbalisation is noticed. When two languages aren’t in contact, they develop different ways of “saying things”, which means that limiting ourselves to substituting words will produce a message that’s completely different to what we aimed for. On this subject, the problem with IT tools doesn’t lie as much in their inability to carry out these tasks as it does in their inability to detect that something’s slipping through their fingers.

Cultural references

Cultural references go beyond mentioning a character, a book or a historical event that’s unknown to the reader of the target language. Culture is rooted in the way we say things and, as time goes by, it has increasingly less to do with words. A Japanese novel that talks about “how beautiful the moon looks tonight” says quite a bit less about the moon than the emotions of the characters. For machine translation, it only talks about an element (the moon) with a characteristic (beautiful). For the human translator, however, there is a traditional resource to make a hidden declaration of love. While the machine translation is correct because moon = moon and beautiful = beautiful, at the moment humans are the only ones able to detect: a) the cultural load of this message, b) the weight of the context to know if the sentence is literal or metaphoric, and c) which words need to be added in English so that the English reader receives the same information as the Japanese reader.

Text types

If we were talking about cultural references at first, it’s now worth mentioning text typology. Instruction manuals, academic articles, mystery novels, children’s stories, comedy monologues, advertising texts, campaign slogans… The features of each of these texts differ between one language and another. A board game manual isn’t written in the same way in English and in Japanese, if we’re following the same example of distant languages. The humour of a comedy dialogue isn’t found in the same places. The source language structures the information in a certain way, specifies some elements, repeats others and removes those beyond as unnecessary, while the target language can have acquired, through tradition, a different set of rules to make the same type of text take shape. The human translator knows this and adapts it so that the reader feels like “they’re reading a novel” or “they’re reading a set of rules”, while for the moment, computerised tools lack the ability to understand the text as a whole and apply these changes. As a result, the reader can get lost in a text rhythm that seems strange to them, with repetitions that seem to contradict themselves, or they might even be unable to understand the game instructions because things simply aren’t explained in that way in their language.

Contextual elements

Under this improvised umbrella we collect all those elements that, strictly speaking, are outside the text and, therefore, a translation programme, focused on the analysis of words, can’t detect. It could be distorted faces in a comic to express sudden humour, motivational music in a film, or any sentence that cannot be read in just one way because it’s been written to create ambiguity. In the first case, maybe it would be necessary to forget about the original words to translate jokes into other languages, like we’ve seen recently in The Last of Us, which has been translated into many languages. In the second, the effort can be focused on creating a discourse that adapts its energy to that of the cinematographic context (shots, movement, music, sound effects, and so on) through different mechanisms that the original discourse uses. And when it comes to ambiguity, we’ve come up against a brick wall. We tend to think that each unit of meaning conveys a unique message: there’s only one possible way of reading a combination of words. However, riddles, mysteries, and double meanings are also a part of language. Deverbalisation is more important than ever in this case, as it allows us to analyse the different options of meaning after that phrase and find the way of covering not just one, but all the possibilities. Yes, we could spend hours discussing Gandalf’s last words before falling to Moria at the hands of Balrog.

In short, understanding a text doesn’t equate to the addition of understanding each and every word of it. On the other hand, the versatility and richness of language lies in all that we don’t see but we still perceive. Until machines find a way to analyse these invisible cracks in language – where, poetically, we could say that the human soul sleeps through – what is transmitted through machine translation will only cover a tiny percentage of what the author tried to convey.

Writer: Maite Madinabeitia
Translator: Katie L. Wright
Proofreader: Elisabet Pina