Google Translate and other machine translators have made a lot of progress since the Cold War. An internet browser can do in seconds what it once took a super computer to do in hours. This inevitable progress leaves many of my fellow translators looking like Chicken Little, always one news article from the apocalypse. The newest scare comes in the form of Google’s Neural Machine Translation and its use of deep neural nets. First introduced in 2016, this update to Google Translate uses probability to increase the accuracy of its translations. Journalists in various publications, including The New York Times and The Economist, have suggested that this newly updated version of Google Translate signals the end of human translation.

I had to know if this “deep learning” version of Google Translate was as revolutionary as those journalists claimed. Would it topple the translation kings like Deep Blue toppled the chess masters? Could it spell the end of human translation altogether? With these questions in mind, I decided to go ahead and test it out. And, I don’t think the sky is falling just yet.

WHAT DEEP REALLY MEANS

Before jumping into the test, I should clarify something here. Deep is an ambiguous word with multiple definitions. These run the gambit from “very intense or extreme” to “profound or penetrating in awareness or understanding.” When people hear that Google bought a company called DeepMind that specializes in “deep neural networks” many people assume that the word “deep” means wise. But, in fact, the word “deep” is being used in a more literal sense. These networks are deeper than their predecessors because they have 12 layers instead of two or three. This does not mean, however, that deep neural networks are any more profound than their shallower counter parts.

With that cleared up, I’d like to share a few of my personal beliefs about Google Translate.

THE SKINNY ON ME AND GOOGLE TRANSLATE

If you’ve been to my blog before, you’ve probably seen a post or two lamenting the quality of Google Translate. Every translator knows that Google Translate has little use outside of making funny internet videos. But, I quickly found that belief didn’t extend to the general public. Many otherwise intelligent people are enamored with machine translation, finding it above reproach. To some degree, I understand. It’s a free tool that speaks more than 96 languages. It’s also the closest many people will ever come to being bilingual.

As a linguist and lifelong translator, however, I do not believe that machine translators are above critique. Simply put, Google Translate is a piece of software which focuses on decoding, instead of understanding, the text put into it. Therefore, the result is typically a garbled, inelegant mess. As you’ll soon see, this still holds true for its newest incarnation.

THE TEST

To test the new and “improved” Google Translate, I translated three English phrases into Swedish and then back into English. What follows is a table of my results:

Original Phrase	Swedish Result	English Conversion	Critiques
Let Me Quarterback Your Transaction	Låt mig kvartalsbaksida din transaktion	Let me quarter behind your transaction	Google Translate does not pick up on cultural references. It also fails to realize that “quarterback” is a term from American Football that means “to lead.” As a result, the Swedish translation is gibberish.
One swallow does not make a summer	En svälj gör inte en sommar	A swallow (the verb to swallow food or drink) makes not a summer	Google mistakenly used the verb form of the word swallow. The resulting sentence structure is strange and literal
Well-written, well-optimized content is a wonderful way to build a reputation	Välskriven, brunn-optimerade innehållet är ett underbart sätt att bygga upp ett rykte	Well-written, optimized well (to get water from) content is a wonderful way to build a reputation	Google misinterpreted the adverb well for the noun well. As a result, the sentence becomes useless.

While Google Translate can bulldoze its way through simple, literal sentences, it’s practically useless for more creative, culturally-oriented texts. It not only tries to tell readers that swallowing is not key to summer, but that soggy pieces of paper drudged up from your well are a fantastic way to boost your website’s reputation. Though the addition of deep-learning will likely improve the software for the general user, it’s still lightyears away from replacing human translation.

CONCLUSION

A lot of translators are terrified of deep neural networks and what they mean for their job security. After testing them out, I quickly realized that there’s nothing to be afraid of. Much like the Google Translate that came before, this “improved” version is still horrible at picking up on subtlety and nuance. It will be years, perhaps decades, before a practical alternative to human translation becomes available.

If you don’t want your Swedish content coming out like the phrases in my test, you’ll need to hire a professional translator. You can contact me directly at tess@swedishtranslationservices.com or visit swedishtranslationservices.com to learn more.