Artificial Intelligence May Only Take Us So Far: The Abject Failure of Google Translator

I like to experiment with language services (Hindi versions) of Google Products. Granted that I don’t need to do this, since everyone I communicate with over email is perhaps as much or more fluent in English than in Hindi. So call it inverse snobbery, too much free time or whatever you please.  The tale that follows reveals that my endeavors are not that fruitless after all.

A friend, to whom I often write emails in Hindi, decided to go one up on me. He replied to me in Bengali.  I decided out of curiosity to enter the text in Google translator and asked for a Hindi translation. The fun began.

‘হর্স ‘( ‘Harsh’), my name written in Bengali script was translated by Google as ‘घोड़ा’ the Hindi word for ‘Horse’.  By no means, the letters making up my name in Bengali combine to mean “Horse.”  Puzzled, I decided to ask Google for a Bengali – English translation and it had indeed translated হর্স  as Horse. Then I realized that Bengali doesn’t have a hard “a” sound and instead uses “au” and instead of a hard “sh” often pronounces it as “s”. So “Harsh” can sound like Horse. And then Horse in Hindi is घोड़ा (the word for the animal) and hence the output by Google. This reveals two major flaws in Google Translator.

First, that Google is really ‘fooling’ users when it offers translation from any source language to many other target languages. For instance in this case the translation was really being made from Bengali to English and then to Hindi. Similarly I checked Bengali to Spanish, the same word was translated as Caballo – the Spanish word for the animal horse. Perhaps Bengali to Spanish being mediated via English is still understandable but Bengali to Hindi via English is a very inefficient way of translating. It is almost like translating between Arabic and Urdu via English. More importantly the service conveys the impression that it directly translates from the source language to the target language.

The second flaw suggested by this incident is even more grave. That is, if Google does not have the meaning of the word in the input language in its database (for instance my name here in Bengali) , it translates the ‘sound’ into English. Now if that sound happens to be spelled as a legitimate English Word, as was “Horse’  in this case, it assigns the ‘meaning’ of the word in English to all subsequent translations.  This completely distorts the original meaning ,of course.

In this case I was reasonably close to the three languages to ascertain what was going on. It may not always be the case. I would perhaps go to a real person to put me wise than rely on artificial intelligence. Big Brother may desire to simplify our lives, but he is not so wise yet, after all.

Addendum: A conversation with someone who read this one. And an update to respond to all previous reactions

Rohan MurarkaHow can you gauge it based on translation of proper nouns?
HarshT :Rohan, good question, but one I had anticipated all along. Of the two arguments I made – the first one ( about translation being mediated through English stands irrespective. The second one is perhaps a problem because of the word being a proper noun ( hence not in the extant database of the source language). But understanding it as a common noun (in the mediating language) for further translation is what changes the meaning completely. It is an error that can be easily fixed – they just need to flag it as a word they cannot translate and retain it. Assume I knew no Bengali or English here -then I had no way to decipher that why did someone call me a ‘ghoda’. Instead the translation could have been ‘haurs’ with a quote or something around it to signifiy that the word was not ‘comprehended’ by the machine
About these ads

About harshT

Assistant Professor at the Missouri School of Journalism
This entry was posted in Language Internet, Media Technology and Society, Personal Accounts, Random Ruminations, Social Media. Bookmark the permalink.

10 Responses to Artificial Intelligence May Only Take Us So Far: The Abject Failure of Google Translator

  1. bp says:

    ঵৆৉ what is their functions

  2. Pingback: The Economist concurs with some of my thoughts on Machine Translation | Harsht's Journal of Random Ruminations

  3. Pingback: Responding to responses on my earlier posts on Google Translate | Harsht's Journal of Random Ruminations

  4. Akshay Madan says:

    Very interesting observation! It seems their program is a usual data based program rather than an artificial-intelligence one.

  5. Pingback: Algorithmic Translation and Human Intelligence « A Spoonful of Win

  6. Pingback: Words of the mouse can mislead: Google Translator exposed through more fundamental evidence | Harsht's Journal of Random Ruminations

  7. sprabs says:

    Google translator uses probabilistic models to figure out best translations, I am not surprised therefore that an unknown word is picked up to be horse (more likely) than your name (less likely). That said, it is clearly a bug :-)

    To type in a non-English language there is a thing called transliteration (Google also has it, so do other sites) where you phonetically type stuff in English and it gets converted to native font for ex ghoda would apepar in Hindi font and so on. While it does not do any translation, it works in an awesome fashion.You would see that would more faithfully reproduce your name.

    • harsht says:

      Sprabs,
      I know transliteration – There is no intelligence in that – it is fairly straightforward.
      What about my first point that came across – that translation is being mediated through English – that is never made obvious by them – but seems to be the case!

  8. harsht says:

    I am no computer scientist to be mindful of how that community uses the term reflexively. In the social world, the term has a connotation which makes Google Translate what ordinary beings perceive as ‘Artificial Intelligence’ ( as task which involved the human brain which a machine tries to do now). In the title, I am voicing an opinion – a computer scientist can differ as can any other reader. Perhaps this blog may prompt Google to correct some of these issues. I will watch the video with interest.

  9. pallav says:


    About 40 minutes into the video, Norvig starts describing some of the algorithms behind Google Translate. The first 40 minutes are important to set context though. Also, side note, the blog title is somewhat misleading – the fact that AI has taken us so far (whatever that means) yet does not limit how far it may take us in the future.

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s