Don’t worry, Google’s AI hasn’t created its own language to hide evil intentions from its human masters, but it has started to teach itself how to translate without a human directly teaching it.

Jamie Davies

November 28, 2016

4 Min Read
Google AI translation tool starts to teach itself

Don’t worry, Google’s AI hasn’t created its own language to hide evil intentions from its human masters, but it has started to teach itself how to translate without a human directly teaching it.

This is another one of those incremental advancements to a technology which has the potential to impact the industry is a hugely significant manner, but Google has come up trumps again with another machine learning win. The news is focused around what is has described as ‘Zero-Shot Translation’.

In short, it basically means the translation tool can translate from one language to another without being taught how to do so. Let’s take an example. The team taught the tool how to translate English to Japanese (and vice-versa), as well as English to Korean (and vice-versa once again), but not how to translate between Japanese and Korean.

The AI can now draw conclusions that by translating a phrase between English/Japanese, and pairing with the translation of the same phrase between English/Korean, it can therefore logically translate the same phrase between Japanese and Korean (and vice-versa).

That is the simple explanation of ‘Zero-Shot Translation’, the AI can now translate without having to be taught the translation itself. This isn’t an advancement which is going to shake the earth, but it is a nice indication that AI is starting to get to a point where it can have a meaningful (if only minor) impact on daily life. It’s a nice, clever bit of work from the Google team.

image01.gif

At the heart of the advancement is a Neural Machine Translation (GNMT) system, which the Google Translation team switched to in September. Until this point the team were reliant on Phrase-Based Machine Translation as the key algorithm behind the translation service, and while it has been effective, it was limited in its ability to adapt. The GNMT system allows the team to incorporate more speech and image recognition capabilities, as well as broader machine learning concepts.

What is very interesting here is that the system would appear to have created its own language where it understands and comprehends the meaning and context of words. The Google Translate team believes it has created an ‘interlingua’ understanding, by where it recognises common representation in which sentences with the same meaning are represented in similar ways regardless of language.

The below image represents a 3-dimensional representation of internal network data which shows the system as it translates a set of sentences between all possible pairs of the Japanese, Korean, and English languages.

Translation.png

“Part (a) from the figure above shows an overall geometry of these translations,” the Google team said on its blog. “The points in this view are coloured by the meaning; a sentence translated from English to Korean with the same meaning as a sentence translated from Japanese to English share the same colour. From this view we can see distinct groupings of points, each with their own colour.

“Part (b) zooms in to one of the groups, and part (c) colours by the source language. Within a single group, we see a sentence with the same meaning but from three different languages. This means the network must be encoding something about the semantics of the sentence rather than simply memorizing phrase-to-phrase translations. We interpret this as a sign of existence of an interlingua in the network.”

It’s early days for the team, but this is a good sign of progress. It shows the team are beginning to develop logic and problem solving skills in the system, as opposed to simply memorizing phrase-to-phrase translations.

The next test will focus on whether it can adapt to human laziness. A surprising number of us don’t write exactly how we think, therefore another individual has to understand and interpret because subconsciously they know about this oversight. Correcting the slight meaning divergences (as well as basic human error) will be the next stage for the team, as well as overcoming cultural differences in basic definitions of the same word in different languages. There’s still some way to go before human translators are replaced, but it’s getting there.

And is it just your correspondent’s imagination or have there been some notable improvements in the Google Translate tool in recent months…

You May Also Like