Hard to sound beautiful when the piano is broken!

Mongolian vs GT– Artificial Intelligence and Machine Translation

A serious bug was recently discovered in the AI translations from Mongolian to English. This bug suggests the Artificial Intelligence software may be filling blanks to make sense of just random combinations of letters. We did tests in other languages and found the same issue there as well. You can see examples of Punjabi, Icelandic and Telugu below.

To some, this might not seem like a big problem. Isn’t the software just trying to make sense out of what you are writing? Yes, it is. However, this is a big problem because it seems like the software can’t tell the difference between a mistake or completely incoherent Mongolian sentence and a coherent sentence. It’s attempts to fill in the blanks do not clue the user into an issue. Instead, it leaves the user unaware that one side of the translation is a complete mess.

See some of the samples below and compare their back translations.

Gibberish Mongolian to English

English to Mongolian

In the example above, you can see with the back translation that the Mongolian in the two different cases is  not even close to the same.  This should signal to the technology that there is an issue. However, the only way a user would know would be by doing a back translation.  Easy to do in this case because the first Mongolian phrase is clearly different.  However, sometimes the differences between back translations would be almost unnoticeable and completely meaningless without a native Mongolian.

And, last but not least, I wouldn’t be doing a good job if I left out the fact that in the sample above GT has a poor grasp of the English language by using “has got” and doesn’t resemble clean professional native English.  In fact, it looks like a non-Native English translation.  Few would disagree that they would prefer “The company has a foothold” over “The company has got a foothold”.  This could be caused by many different issues with the technology, but the end result is the same.  A poor translation that would lead to embarrassment and errors.

We look forward to seeing this technology improve.  As previously stated, we have high hopes for machine translation in the future.  In the meantime, please contact a professional language service for anything important.

Here are a few more funny translations that you might enjoy.  Take care and good luck.

And this one seems like the machine translation is agreeing with our assessment as it seems disappointed in its own programming.




By Allison Baillieston | January 21, 2018 | Categories: ITC | No Comments

About the Author: Allison Baillieston