COVID-19 has changed the world, and for many, the biggest change came in how we use technology. The pandemic has convinced more businesses to allow remote work than any cogent argument. So, one must ask: can technology apply to translation? Could a computer possibly create translation as good as a person’s?
It’s not an unreasonable argument. Reports have been coming out for a while now that AI translation’s been coming closer to human levels; The Verge suggested that Google’s AI translation system was nearing “human-level accuracy” in 2016, for example, and Microsoft has made huge strides forward. But how good is it, really?
Can a Computer Translate Exactly as a Person can?
The short answer is no. The long answer: A lot of the hype is based on the advent of Neural Machine Translation, which tugs on neural networks—meaning a system that can be coded to recognize data patterns—for machine translation; computers can also process enormous data sizes.
From the computer’s end of things, it converts symbols from one language into another, using neural network models to learn what the symbols originally were and what they became. More importantly, it uses previous data—often created by human translators—to do so; things like European Parliament minutes, for example, get translated by hand into 21 languages, which is a gold mine. In throwing-out a rules-based approach, Neural Machine Translation has made waves.
A Smart Computer Program is Still a Program
But the problem is that it’ll still make errors, as a computer program is made-up of a series of codes. Each algorithm has a specific function. And that’s just not how language works.
Humans are not logical and cannot be reduced to a series of mathematical symbols, and language, perhaps humanity’s greatest achievement and a key factor that distinguishes it from other species’, is constantly evolving.
Slang is one thing, but consider how fast the world is evolving. There are new buzzwords, keywords, and technologies being introduced every day in engineering, medicine, business, politics, etc. In 2020 we started to interface via ZOOM as a result of COVID-19—a sentence that wouldn’t have made sense a year ago.
Languages borrow phrasing from each other. Jordan’s most popular show is Spanish, affecting Arabic slang. Neologisms are introduced. In a way, there is an irony to translation technology: it can’t quite keep up because of
how fast the technology itself is moving.
What Human Translation is Distinguished from Computers’?
speed aside, what can people bring to translation that computers can’t?
1- Reliability
Computer programs don’t actually know what a word means, just that the odds suggest where they should go, and the results are still often surreal, sometimes offensive, and largely unintelligible. A favorite: A sign at a famous chain supermarket in Amman advertised its mixed fruit as “fruit problems”—the word مشكلة was the guilty homograph.
2- Localization
A phrase like “You’re having me on” would make perfect sense to someone from Britain, but would confuse Americans; a phrase like “on the ball” would confuse someone from Britain (or New Zealand or Australia), but make perfect sense to someone in the United States. We’ve discussed the importance of good localization before, and a good translator will tailor a text to suit its audience. A computer cannot do this—not yet, anyway.
3- Consistency
Neural Machine Translation isn’t good at context, even though it’s designed for it. Astonishingly, computers do learn, but we can’t always guarantee they learn the right things. And that’s also because computers don’t understand.
4- Subtext and Tone
Deciphering an idiom is one thing—that’s something Neural Machine Translation largely took care of—but computers simply cannot puzzle out things like subtext. A formal letter can be misinterpreted as a threat; a sincere marketing campaign will suddenly seem sarcastic. These are components that it takes people to work out because they’re people; they’re entirely based on being human. A computer has no use for sarcasm; programming common sense is unthinkable.
What is The Core Problem with AI Translation?
Rather more insidious is when translation mistakes have real consequences. We’ve talked about translation mistakes from people before, but each situation was resolved because the people there had enough common sense to communicate (with one notable exception during World War II).
This isn’t possible when it’s machines doing the translation.
1- Computers work with data, and data creates patterns
A famous example of this is attempting translation between gendered and ungendered languages; if a data set suggests that most of the engineers are males, the translation will treat any female engineer as a male. Worse, because Neural Machine Translation lacks consistency, one page of the text may describe someone named (for example) Sara as a woman, but the program won’t take that into account apart from that one sentence.
2- Computers don’t understand meaning
Consider the following headline from Corriere della sera, an Italian newspaper, as done by Google Translate (which uses Neural Machine Translation): “Covid, doctors and nurses: trauma like war veterans. Psychic distress for 4 out of 10.” It’s nonsensical. And that’s just one paragraph: a correct translation of the headline () would read: “Medical Staff and Nurses on COVID: ‘War-Like’ Trauma; 4 in 10 Staff Experiencing Mental Distress.” In an age where news is shared on Facebook and fake news is prevalent, a mistranslated headline can have serious consequences.
3- Computers have no common sense
A person announces, “I’m a PC.” A program would translate this as a person stating they were a computer. A person says, “I’m a fan of metal,” and the computer might think they were a cooling device wrought out of, well, metal. A human translator would correctly conclude that the first person was stating they were a Police Constable, and the second a listener of, well, metal music.
Will AI ever fully replace translators? Maybe. Only time will tell. But right now the technology just isn’t there, and programs have simply become another tool in a translator’s arsenal. Luckily, we know a thing or two about good translation. If you want your work properly translated, localized, or edited, reach out to Gibran today.
Leave a Reply