On Google Translate’s sixth birthday this month, its developers had plenty of reasons to celebrate. Two hundred million in fact: that’s the number of people who use its free international language services each month.
“In a given day we translate roughly as much text as you’d find in one million books,” Google Translate engineer Franz Och wrote in a blog post. It’s roughly the same amount all human translators produce in a year.
There’s no doubt that’s an impressive feat. The web service now covers 64 different languages, including many smaller ones such as Basque, Yiddish and even Esperanto.
But of course, numbers aren’t everything. Google doesn’t make the same claims for quality as quantity. The recent furore when English court staff had to rely on the online service, due to a lack of human interpreters, suggests machines aren’t yet up to the same standards.
And it’s not set to replace other international language services any time soon. The market for human translators is also growing, with more British companies looking overseas for new business opportunities.
Unlike earlier translation programs, Google Translate has little knowledge of grammar or sentence structure. Instead it works on the assumption that the sentence you’re trying to translate (from Arabic to French, say) has already been translated. It sifts through huge volumes of translated material, from novels to legal documents, to find a similar sentence.
This approach is pretty successful for common combinations. But it does run into problems for unusual combinations, such as Finnish to Swahili.
A study in Translation Journal found that its accuracy varied widely between languages. Researchers concluded it was fairly accurate between European languages, but “relatively poor” for Asian languages. This could be due to lack of material, or differences in the structure of the languages.
They found that for most language combinations, it produced a result that was “acceptable for reading comprehension”. But they didn’t take into account the fluency or style of the results. With no knowledge of context or tone of voice, machines often produce results that sound stilted or awkward to a human reader (if not garbled).
As Google itself says: “Google Translate can make intelligent guesses as to what an appropriate translation should be.” But it admits: “Not all translations will be perfect”.
Researcher Andreas Zollmann, who works for the company, has expressed doubts that more and more data will make it more accurate. He told the Guardian: “If you ever solve machine translation perfectly, then you have something that is properly artificially intelligent. Language is not separate from who we are.”
Of course there are plenty of innovative ways machines are assisting humans with international language services. Our post-edited machine translation service uses human editors to provide acceptable (and cost-effective) translations of large volumes of materials. And smart software can identify repetitions in documents, saving time for translators.
So what do we recommend? If you’re sending a quick email to your Italian friend, or want to figure out what that Swedish article says, then online translation tools can be a huge help! But for any client-facing material, then there’s still no substitute for professional (human) translators.