I’ve been a translator for just shy of 15 years. Yet I’ll be the first to state it’s not the most glamorous of careers. We toil behind the scenes, giving of ourselves so our clients’ style, tone, and meaning shine—just in a different language.
But heaven forbid robots take credit for our work. That’s when we get noisy.
In late 2018, that’s just what happened. Murmurs grew into roars when a Shanghai interpreter took to social media to protest the misleading marketing of “AI-powered translation” at an international conference. In actuality, the “translation” was a voice-to-text transcription of a human interpreter’s work. The post went viral on Chinese social media, creating a scandal for iFlyTek, the promoter of the mislabeled technology.
The public response was revealing. It showed the extent to which people believe machines have already replaced humans. Many were shocked to learn that today, interpreting still requires human professional to perform specialized work.
Turns out, they didn’t.
It’s true that AI consistently outperforms humans at chess, poker, Jeopardy, and driving cars (not to mention diagnosing cancer, shooting free-throws, and playing Go). But when it comes to language, the most sophisticated technology on earth is still very much the human brain.
How come? Here are 3 reasons why machines won’t replace human translators any time soon.
Big data just isn’t big on humor
Any translator will tell you that puns, sly innuendo, idioms, and jokes (as well as nuanced cultural references) are among the toughest hurdles to jump to cross the language barrier. And without them, our quality of expression becomes much poorer. From an interpreter’s standpoint, body language and tone of voice directly inform a speaker’s intent and must be accurately analyzed and conveyed in the target language.
While this is challenging for humans, it’s currently impossible for machines.
The move from phrase-based machine translation to neural networks has yielded a significant improvement in overall quality. Yet, neural machine translation is even more dependent on vast sets of training data than its predecessors. And since the biggest bilingual datasets available are from official translations of religious texts and government documents, these algorithms have a pitifully low exposure to non-verbal expression, wordplay, and humor.
Perhaps more disturbingly, neural machine translation doesn’t confess its mistakes. Instead—like an ill-prepared student—it often tries to fudge through them. When Google Translate started offering biblical prophesies in exchange for junk input, experts attributed the errors to neural networks’ preference for fluency over accuracy.
These “false positives” are extremely clumsy and far more insidious than more obvious mistakes, as audiences in the target language may never realize a glitch occurred and might attribute the outlandishness of the rogue translation to the original text itself.
Language is subjective
Artificial intelligence typically excels at tasks imbued with objective reality. Whether this means navigating road conditions or identifying patterns, machines function best when given clear rules.
By contrast, natural languages are subjective constructs invented by humans to communicate with each other. Machines do exhibit rule-like behavior (like grammar and conjugation). But those rules are grounded only in convention and not objective reality. What’s more, they’re constantly evolving.
We may have forfeited our lead in assessing credit risk or recognizing tumors, but we still have—and may always have—the final say over what is or isn’t “natural” in a natural language. This authority is reflected in the metric of choice for evaluating machine translation algorithms, the BLEU (Bilingual Evaluation Understudy), which scores candidate translations based on their similarities to a human professional’s work. The framework’s inventors concede that the closer a machine translation is to a professional human translation, the better.
Human translation doesn’t just set the standard. It is the standard.
Listen very closely, bots
The two challenges above make it difficult enough to perform machine translation on a static piece of text. But asking a computer to interpret live speech simultaneously makes for a Herculean task. As such, Automatic Speech Recognition (ASR) is still fallible in many ways.
Alexa and Siri may seem like skilled conversationalists, but their witty banter has limits. Conversations with robots tends to fit a narrow set of conditions. Robots fare best with short, command-based interactions involving a controlled environment and a finite vocabulary.
On the other hand, most live conferences and business discussions feature spontaneous, highly context-dependent, and continuous speech—traits that make the error rate of most ASR programs skyrocket.
There are many hilarious and humiliating examples of ASR falling short. Earlier in 2018, hedge fund guru Ray Dalio reflected on his mis-forecasts as a young trader and machine translation failed spectacularly. “How arrogant!” he bellowed to the crowd. “How could I be so arrogant?”
The real-time subtitling program struggled in vain to render his rhetorical device.
“How?” the subtitles asked. “Aragon, I looked at myself and i.”
While recent advances in the field are promising (and many experts predict that the word error rate of ASR software will reach parity with human transcribers in the near future), not all word errors are the same. Fudging “alright” into “all right” may be an inconsequential mistake, but confusing “today” with “Tuesday” could cause a serious mix-up. Even with fewer errors, machines remain far more likely than humans to commit semantic errors that misrepresent the intended meaning of speech. That’s a huge part of why machines won’t replace human translators any time soon.
Humans and robots can work together
Mankind has long made a pastime of reflecting on our perceived superiority—over each other, over animals, and more recently over machines. It’s a dark pastime and an inevitably foolish one.
The day very well may come when computers develop human-like command over natural language. One day interpreters and translators, along with radio hosts, editors, copywriters, and others in the language economy, may find their jobs on the robot’s chopping block.
But that day is further away than most people think. Part art, part science, language work is surprisingly defensible against these early iterations of artificial intelligence.
As with so many other industries, we language professionals should focus on harnessing artificial intelligence and neural language processing to increase our labor’s efficiency, quality, and cost-competitiveness. CAT (Computer Assisted Translation) tools are already widely used among translators, and while many bristle at the notion, no doubt simultaneous interpreters would benefit from some form of speech recognition combined with translation memory technology. However, for the foreseeable future, these tools would serve to complement, not overtake, human output.
And as long as there are humans interpreting in booths or toiling over a desk, let’s be decent and give them the credit they deserve.