LearnLanguageswith Us!

Start Learning!

Language News

How Google Translate Works, and Why It Doesn’t Measure Up Posted by on Sep 2, 2015 in Language Learning

With over 200 million daily translations, there’s no denying that Google Translate is a wildly popular translation service. Indeed, machine translation has come very far since its infancy in the early 2000s. Instead of translating words at face value, machine translators have developed complex algorithms to deliver more accurate translations, and some even take into account colloquial language and idioms. Still, the very nature of machine translators prevents them from ever doing a human’s job. Let’s take a look at how machine translators (such as Google Translate) work, what their limitations are, and why they can’t replace the quintessential human touch.

There are a lot of things that this computer can do, though producing an accurate translation is not one of them. Image via Pexels

There are a lot of things that this computer can do, though producing an accurate translation is not one of them. Image via Pexels

How machine translation works

Google Translate, as well as other machine translators, operate on statistics rather than rules. That is, they look for patterns in hundreds of millions of documents that have already been translated by human translators. Google Translate makes special use of UN documents, which are translated in all six official UN languages, and thus provide ample linguistic data. This way, they can weigh a plethora of options for phrases presented by various different (human) translations, and select an educated guess based on the one that occurs most frequently. For example, they detect that, in Spanish, the phrase “darse cuenta” is usually translated as “realize” in English. Therefore, based on statistics, Google Translate will correctly translate the phrase as “realize”, rather than a word-for-word translation, which would appear more like “give account”.

Finding linguistic data large enough to create legitimate statistical analyses is no easy feat. Given that more documents are available in English than in any other language, the data almost always uses English as an intermediary step when translating between two languages that aren’t English. For example, when translating from Russian to Spanish, Google Translate will first translate the text from Russian to English, and then from English into Spanish. As a result, when translating in languages other than English, machine translations actually involve two iterations.

In fact, some language pairs involve even more iterations. If you want to translate some text from, say, Catalan to Japanese, Google will translate it first into Spanish, as most existing Catalan translations are in Spanish. Then, this translated Spanish-language version of the original Catalan text will be translated into English. And finally, the English version of the Spanish version of the Catalan text will finally make it to Japanese — and if you’re lucky, it will still bear some resemblance to the original meaning.

Why it doesn’t make the cut

Google Translate does a good job with very basic translations — especially those whose target language is English — and now even offers alternative interpretations for certain words and phrases. However, the very methodology upon which Google Translate is based prevents it from ever competing with human translators. Here’s why:

Statistics don’t have feelings. Google Translate is based on statistics — it chooses the “best” translation based on how certain words and phrases have been translated in other documents. As a result, machine translators choose the most probable translation, but not the most interesting or poetic one. As a result, even if translations are accurate (which they often aren’t), they adopt a robotic, lifeless tone. It takes a human translator, with feelings and creativity, to reproduce the tone, color, and vibrancy of the original text.

Machine translations struggle with complex grammar. Language is based on rules, and as a result, a statistics-based translator like Google will struggle with complex grammatical concepts, such as the difference between the imperfect and preterite past tenses in Romance languages. This is especially true given that Google almost always uses English — a language that does not grammatically distinguish between preterite and imperfect tenses — as an intermediary step when translating into Romance languages. Therefore, Google Translate often incorrectly translates the imperfect past as the preterite past (and vice versa), making ongoing or habitual acts seem like one-time, completed events.

Google can’t write for an audience. Every translator knows that you need to tailor your work to whom you’re writing for. For example, if this article were written for a casual blog, my use of the word “whom” in the previous sentence may come off as overly formal. However, given that this article appears on a language interest blog, grammarians and language experts may applaud my correct distinction between “who” and “whom” (though they may scoff at my decision to end a sentence in a preposition). Machines cannot make such judgment calls — Google cannot take into account who the intended audience is for the article it translates. Only a human translator can make that kind of decision.

Google Translate vs. a human being

To illustrate the difference between Google Translate and a living, breathing human translator, I will employ both to translate the following text in English, which appears on a website selling Argentine wine. Try to guess which one was written by a human, and which was produced by a machine (spoiler alert: it won’t be hard).

If you’re selling wine, you’d better hire a professional translator. Image via Davide Restivo / Wikipedia

If you’re selling wine, you’d better hire a professional translator. Image via Davide Restivo / Wikipedia

Original text:

Después de una excelente cosecha como la que le precedió, la cosecha 2009 muestra sus virtudes en este vino base Cabernet Sauvignon, mas el ensamble de tres variedades de gran personalidad que encontraron en San Rafael el terruño ideal para la expresión de sus mejores cualidades. Vino aun de color rojo violáceo intenso a pesar de los años en botella, ya en la copa se nos muestra intenso y seductor con aromas especiados que se entremezclan con nítidos y frescos aromas a frutas de ciruelas, cerezas negras y moras, mientras que se van desprendiendo lentamente los aromas tostados que recuerdan a granos de café molidos.

Translation 1:

After a bumper harvest as that which preceded the 2009 vintage shows its virtues in this cuvée Cabernet Sauvignon, but the assembly of three varieties of great personality that found in San Rafael ideal for the expression of his best qualities terroir. Wine intense purplish red even though the years in the bottle color, and in the cup shows intense and seductive with spicy aromas mingle with crisp, fresh fruit aromas of plums, black cherries and blackberries, while van slowly peeling roasted aromas reminiscent of ground coffee beans.

Translation 2:

After a great harvest like the one that preceded it, the 2009 harvest shows its virtues in this cuvée Cabernet Sauvignon. It’s an expressive mixture, articulated by three varieties of great personality that are found in San Rafael, the perfect region to bring out its best qualities. The wine still preserves a strong purplish-red color, in spite of the amount of years gone by since it was bottled. Once poured into the glass, it remains intense and seductive. Its spiced scents mix together with clear and fresh fruit aromas of plum, black cherry and blackberry, while its toasted scents release slowly, reminiscent of the delicious smell of ground coffee beans.

You probably guessed right: the first translation was done by Google; the second, by a professionally trained bilingual translator. As you can see, the machine translation is barely comprehensible: it contains grammatical gaffes such as run-on sentences, wrong adjective order (e.g., “wine intense purplish red”), and sometimes just failed to translate words altogether (e.g., “van”). On the other hand, the human translation flows smoothly, is organized coherently, and matches the elegant tone of the original article.

Machine translation isn’t without its perks. It can be a life-saver in a pinch, when you need to get a rough idea of what a certain phrase means, or when you need to decipher a street sign while traveling to a foreign country. However, when it comes to translating important documents, the limitations of machine translation prevent it from being a viable option. As the above examples demonstrate, a translation that is based on statistical patterns will never match the quality of one created by a professional, who understands the rules and nuances of language. Indeed, machine translation has come a long way, but it’s still far from replacing the human touch.

The following post is from Paupaul_thumbnaill, an English teacher who lives in Argentina. Paul writes on behalf of Language Trainers, a language teaching service which offers foreign language movie reviews as well as other free language-learning resources on their website. Check out their Facebook page or send an email to paul(at)languagetrainers.com for more information.

By the way… want more free language learning resources, advice, and news from Transparent Language? Sign up for our newsletter!

Share this:
Pin it

About the Author:Transparent Language

Transparent Language is a leading provider of best-practice language learning software for consumers, government agencies, educational institutions, and businesses. We want everyone to love learning language as much as we do, so we provide a large offering of free resources and social media communities to help you do just that!


  1. Matthew Rothenberg:

    “Never”? That’s a bold assertion based on only a few decades of machine translation. I’ve certainly noticed in recent years that Google Translate has become far more comprehensible and handles many more idioms in the languages I can read.

    I do think there are choices a human translator makes that a machine would be hard-pressed to match. And certainly the current state of the science is limited in the ways you describe. However, machine translation is helping people get the gist of writing far more complex than street signs. (I often use it to get a general understanding of what’s being discussed in Arabic newspapers, for example. It’s not pretty, but it’s much better than nothing.)

    So … All good points, but IMO short-sighted about what “never” really means.

  2. Chris:

    Not exactly on-topic, but “Every translator knows that you need to tailor your work to whom you’re writing for”
    even as a native English speaker, I had to read that twice!
    –> “…tailor your work to the person for whom you are writing” maybe
    “”tailor your work to your audience” would be even clearer, but misses out your deliberate use of “whom”.

    Just playing devil’s avocado,

  3. paul avermaete:

    “…tailor your work to the person for whom you are writing”,,,

    “c’est le Ton qui fait la Chançon”,,,(French proverb)

    Very interesting explanation, I think the best we-users-can do, is ALWAYS first translate into English, and then in the choosen language; is this an agreement ?

    thanks, paul Antwerp Belgium

    • Jon:

      @paul avermaete Le Ton doesn’t like that either, lol.
      .tailor your work for the person to whom you are writing..
      ..tailor your work to whomever you are writing for..
      just having fun

  4. chris:

    ” c’est le Ton qui fait la Chançon ”
    does not work well in Google Translate.

  5. Terry E Ferrell:

    Google Translate doesn’t work that well all of the stuff on my phone still comes out in Korean I can’t even get into the music programs that I want from google translate all for nothing to Korean and kpop stuff I cannot read what is the problem

  6. Lily W:

    Google Translate is good for single words. Not full sentences or paragraphs. The translations are completely wrong and Im not going to pretend I know why. Dont use it for important conversations. Youll just look like a fool. Shell out some money and purchase Rosetta Stone or something.

  7. Lily W:

    P.S. This was a great article!!

  8. Elena:

    It was interesting to read about 2 intermediaries in the translation from Catalan to Japanese. When I explored this Google’s feature in 2011, I found that Ukrainian and Russian were “on equal terms”, Russian was not an intermediary language to translate to and from Ukrainian. It seemed a bit political))
    Translations from Russian to Ukrainian and vice versa were 95% correct (which gives hope that translation between very close grammatically languages can be already automated for the most part, e.g. Turkic, Eastern/Southern/Western Slavonic languages).


  9. Janmejai:

    nice article but I needed a technical article in detail.

  10. Stephsnie:

    I think google translate is better than not trying at all. Have you ever spoken to someone who spoke very little English but you could understand what they needed even if you wanted to giggle st how they asked? Americans usually don’t even try to express themselves in another language, if they are encouraged to try through the tool of Google-it is priceless!

  11. Toni Molik:

    I tried to translate one word from Malayalam (Thirovosthi) to English. I don’t understand the procedure. How does it work? Toni

  12. P De Mario:

    I think that google give us an idea what we wanto to translate, but the final conclusions it depends from us.

  13. Sam:

    What is missing here is the fact that not all human translators are excellent linguists translating into their native languages. Human translations can be awkward and inaccurate as well. Google Translate makes it possible to read news or other websites in foreign languages. Sure, the output is not perfect, but it’s usually much better than what you could get by trying to look up each word one by one in a dictionary. Much faster, too.

  14. fari:

    google can not translate at all ….It does not make any sense ..very disappointed ,,,,

  15. Swift:

    Example: https://translate.google.com/translate?hl=en&sl=auto&tl=en&u=http%3A%2F%2Fotorten.ru%2Fdocheri-otortena.html entire text translated word by word, often wrong words that are assumed typos or omonyms, no syntax at all.And it doesn’t accept any corrections, from AI’s point of view this is 100% sure translation

  16. Michael:

    This thing about Google Translate using intermediary languages… how do you know that? What’s your source?

  17. Adrian Wallwork:

    The example given makes no sense. No human would translate a document on the fly and then not spend a few minutes revising it. Google Translate simply provides you with a first draft which you can (must) then work on.

  18. Patrick:

    Good info…some thoughts…
    As a theologian we have had to deal with biblical translations and versions since forever…for example the popular Jerusalem Bible in English is considered a version since it is a translation from the original French translation which was translated from the original languages. In other words the English JB is a translation of a translation…and in these cases serious biblical students know something is usually lost when that happens. However key meanings usually are not.
    AI gets better the more it is used. I use Google Translate when I get stuck writing to my French family and friends in French. I will try a phrase or a sentence in various ways before hitting the send key. I use the one that seems to be the clearest. But the author is right…the human touch along with kindly corrections is the best teacher.

Leave a comment: