The Magic of Machine Translation and the
of  Translation
Is machine translation magic?

By Christopher Reid

While attending a localization conference in Warsaw a few years back, I heard the chief technology officer of a well-known machine translation (MT) company refer to the results of his engine as just that — “magical.”

I found this troubling because, for a chief technology officer, MT should not be magic — in fact, it should probably be the exact opposite of magic. Still, the question of whether MT is magic cannot be easily dismissed. Speculations about magical MT or AI run rampant throughout the internet, and the suggestion even appeared recently in the New York Times.

As a former translator and MT entrepreneur, I have witnessed a form of MT’s “magic:” Time and again, I have seen how it is able to convince unsuspecting translators that its results are correct, even when they’re not. This trickery means that MT is not just magical, but potentially diabolical — what medieval peasants might call “black magic,” perhaps. We therefore need to reflect on the implications of MT being so good that it is even able to fool native speakers. It’s a question of existential relevance not just for the translation industry, but for translators themselves.

The magical nature of MT recalls the Disney animated film “The Magician’s Apprentice.” The novice played by Mickey Mouse shirks his water-carrying duties by surreptitiously using his master’s magic cap to bring his broomstick to life. Not only does this result in the broom doing Mickey’s work for him, but the broom begins to multiply of its own accord and breaks free of Mickey’s control. Ultimately, the duplicating brooms cause a flood of water, bringing to mind the prospect that Mickey could drown.

Taking this metaphor further: MT is already carrying our water. And if we cede the most human aspect of our existence to MT, we may one day find ourselves similarly submerged by a torrent of ones and zeroes. Recently, I translated a text from German to English with the phrase “Erlebnis-Individualisierung,” which MT translated as “experience individualization.” This translation sounds about halfway correct, but it’s also clunky and, in this case, didn’t fit the context. Instead, I consciously uncoupled the German compound and translated its second half as “personalization.” But, what if I had been lazy like Mickey? What if untold numbers of linguists simply deferred to the machine over and over again, over an extended period of time? As in the popular film, we could wind up in the matrix.

Germans have a phrase for English terms that circulate in the German language: They say that the word was Germanized (eingedeutscht). There has been no little amount of hand wringing over the decades about how this is ruining German. As near as I can tell, however, there is nothing systematic about this process. Although there’s some speculation, few would be able to pinpoint the unhappy moment when the English word “handy” became cellphone in German. Considering the steady churn of textual production, but also that the written text is still the medium by which language is learned and propagated, the assault on language is potentially even greater.

But we don’t have to resign ourselves just yet to an MT-induced demise of language. That’s because not all MT is the same. In fact, the two different types of MT at the cutting edge promise two different futures for translation: custom MT and interactive MT. The rise of the latter should be more beneficial for the profession and language in general.

Custom MT uses a company’s language data accumulated over the years to train an engine so that the work of translation is predictable and accomplished in the very first processing step. It makes it possible for companies to present uniform corporate branding across multiple countries, to ensure that terms in technical or other standardized documentation is consistent and that there is little room in the translation stage for incorrect translations. Businesses in all industries are rejoicing, because custom MT gives them the standardization that they need while reducing the translator’s role to being a veritable checker.

I spoke with a young woman recently who completed a PhD on MT and post-editing, and she envisions a future in which translators essentially audit the machine. If the text passes the quality standard, it gets a stamp of approval. This calls to mind Charlie Chaplin in the film Modern Times as he frantically tries to tighten bolts on a sped up assembly line. The Little Tramp is then sucked into the machinery to become part of the automation. In this future, the machine has fully usurped human production. Interactive MT, by contrast, promises a productive merger of man and machine. It entails the possibility of preserving language’s integrity rather than letting MT run roughshod over it with unknown long-term effects.

With interactive MT, when the linguist clicks on a word they are also presented a list of other possible translations. This feature shifts the responsibility to the translator to reflect on the suitability of a translation and to make a decision. The translator needs to think about which term is the correct one in a particular circumstance in light of the context and her wealth of experience. Here, the machine prompts the linguist to engage in the most human of acts — the act of reflection. This is the first and most constitutive moment of the translation process.

This drop-down list is an instance where the machine unabashedly admits its imperfection and presses the translator for help. It is up to the linguist to choose wisely. It remains to be seen whether translators will always have the linguistic fortitude to correct unruly MT like a schoolmarm of old — as in my example above with Individualisierung. If we reach a point when translators cannot decide between an array of linguistic options, then we are likely already speaking a language that is more machine-made than man-made.

It’s been observed that the great books of world literature should be translated each generation anew. This is because every era is informed by new experiences, sensibilities, and vocabulary. As the German writer Goethe said, to translate is to venture as far as the untranslatable. It is and will always be the task of the translator to interpret our humanity for the page. The translator realizes their most essential function in the process of producing language, in the interpretive act of understanding the original author’s intent.

Some may speculate whether MT might someday translate Dostoevsky. But an algorithm will never be able to understand the peculiar fates of mankind. MT does not even know what it appears to know. Its predictions entail neither inspiration nor certainty. There is no cognitive sub-layer beneath the programming or beyond the neural network. MT lacks the subconscious reservoir that writers and translators alike draw from in their creative endeavor.

In this context, it is important to think about who the guardians of language are. Who are the ones charged with preserving its integrity so that the good terms (in a virtuous process) are maintained and the bad terms are rejected? Today’s linguists must therefore be called on to keep language alive — to keep it from becoming a “dead letter” that conveys basic information rather than a medium where human creativity can be fully explored and put on display.

If we are honest, it is no longer the case that translators simply translate. In a sense, we are all “post-editors” now. I want to argue to all those conflicted souls in the profession that it’s okay. This merger between man and machine, linguist and MT, does not portend the end of their essential function as language gatekeepers. My appeal to translators would therefore be to protect their respective languages against the blitzkrieg of high-quality, high-powered machine translation. Working linguists are defenders of the realm, which means that they need to have more in-depth and intimate knowledge of their languages than ever before.

It seems certain that MT’s flaws in the translation process are going to become even harder to detect. It is important that these quasi-translations, which represent subtle but unmistakable shifts in meaning, do not become ingrained in our speech. It’s not difficult to imagine a world where, if this happens enough, humans will no longer be communicating like humans. The magic of MT will be a recursive loop in which the machine speaks us and we speak the machine. May linguists everywhere work together to stave off this diabolical future.

Assuming that linguists can uphold their end of the bargain with MT, then the future of language may (ironically) be more at risk because of a translation problem. Humans, of course, also come up with bad translations, and the term “post-editor” is no exception. The “post” suggests that the linguists are not essential to translation but an afterthought. The phrase “post-editor” is not just off-putting — as I’ve tried to show, it’s inaccurate. What in fact may end up keeping people out of the profession is this ignominious moniker. Understanding the linguist’s urgent role, they certainly deserve a befitting title. Let us ponder this, then, before we fall too deeply under MT’s spell.

Christopher Reid, PhD, is an MT entrepreneur and academic translator based in Frankfurt, Germany.



Subscribe to stay updated between magazine issues.

MultiLingual Media LLC