The idea seems great but ... does Google translate really needs to learn about this overtime? I don't quite get it. Using the example in the article, Japanese -> English and Korean -> English, and then Japnese -> Korean. Can't the system automatically create that knowlege on its own w/o even human inputs by using the four pairs it already has? It can be called "learning" but probably human inputs aren't required. The learning is really coming from within and all it requires is computing power, time and an algorithm. Maybe I've missed something ....

Wow. This is very cool as the power of being able to see that the languages have a commonality between them. As we add more languages I wonder if the translation will become more or less accurate. If it becomes more accurate we can then "read" some of the forgotten texts. Thank you for sharing, I will need to read up on this project and see how if any training is happening especially in terms of linguistic structure.

Obviously the author doesn't really understand what ANNs are doing.

Hence his misinterpretation of Google staff's "We interpret this as a sign of existence of an interlingua in the network." into THERE IS interlingua created all by the system itself.

Well, I am working a lot with translate.google.com these days, and it sucks: It's arbitrary, imprecise, only usable if you have close knowledge of the context of the text already. But perhaps it's the old technology still being online?

Don't get me wrong, Google Translate is a great tool with A LOT of potential. But before it is justified to compare it to Star Treck's multi-language translator (which I would consider the benchmark) even only being limited to text, there is still at least 5 - 10 years of further R&D to go.

In that sense, putting the "tower of babel" on top of the article is a very apt choice: It is THE symbol of language confusion in the world. But, most likely the author didn't get that as well?

The link below is another article (The Great A.I. Awakening) about Google neuro-translation methodology and several team stories from New York Times. I find it very inspiring. Hope it would be the same to you.

https://mobile.nytimes.com/2016/12/14/magazine/the-great-ai-awakening.html?_r=0&referer=

It is not as impressive as it is made to sound. It simply made different connections between the languages that was not originally encoded. For those not familiar with NN - that is what it does to refine the "best chromosome" for reducing errors.

However, the question here may be broader than the application of an "interlingua" produced by the machines in the translation between languages. Let's look at this from a broader perspective, let us say we are not trying to translate Korean to Japanese, but we are trying to determine the significance of similarities or clashed between concepts that are widely separated from one another. You can provide your own choice about what the widely separated concepts may be. Now let's say that rather than bringing the two concepts or results from each concept and allow them to clash against one another directly, we transfer the concepts and results into an intermediary level and compare the concepts or results at this intermediary level. It might help in bringing some degree of consistency or emphasize inconsistencies between them by using some kind of synthetic intermediary representation of both. This is not something that is currently in use, but it could done. Successfully, who knows?

I hope the application of an "interlingua" will also take care of African languages

Curious on how this compares to the offerings of Microsoft's LUIS and IBM's Watson.