Do you realize the neural machine learning update of Google Translate?

22 posts / 0 new
Senior Member
Joined: 18.10.2013
Pending moderation

Google Translate has been using neural network learning algorithm for its translations since September. So it's able to learn the languages via shaping neural networks with datasets on Internet. https://research.googleblog.com/2016/09/a-neural-network-for-machine.html

Actually it learns the pure meaning of the sentences. That makes possible cross translations. It seems like the system builds a hidden interlingua. They mentioned that situation:

“Is the system learning a common representation in which sentences with the same meaning are represented in similar ways regardless of language — i.e. an “interlingua”? Using a 3-dimensional representation of internal network data, we were able to take a peek into the system as it translates a set of sentences between all possible pairs of the Japanese, Korean, and English languages.”
https://research.googleblog.com/2016/11/zero-shot-translation-with-googl...

It works for 8 languagues. As far as i see, it works perfectly for English to Turkish translations.

By the way, artificial neural network learning algorithm is a model which is based on biological neural system. With genetic algorithm, it can evolve. This method can be applied to many places. You can see this on that video, neural network learns to play Super Mario:
https://www.youtube.com/watch?v=05rEefXlmhI

Generating music in Bach's style:
https://www.youtube.com/watch?v=A2gyidoFsoI

Learning Van Gogh's Art:
https://www.youtube.com/watch?v=-R9bJGNHltQ

And that's what happens in machine learning:
https://www.youtube.com/watch?v=wvsE8jm1GzE

Super Member
Joined: 02.03.2013

That's awesome.

Editor || ᴋɪssᴇᴅ ʙʏ ғɪʀᴇ . ♔
Joined: 08.09.2014

That's terrifying.

Senior Member
Joined: 12.07.2016

That's Privacy Issue for many people, honest.
Most of these algorithms use "service data" which is, well, anything you type on google/google aided site.
The ethics behind this were questionable even 20 years ago.
Besides that, with these algorithms there is a greater probability of getting a wrong/informal/derogatory/rudimentary translations, which I guess nobody wants.

If you have no problem with it, that's fine. But, I've got a problem with it.

Senior Member
Joined: 18.10.2013
SAKAE wrote:

That's Privacy Issue for many people, honest.
Most of these algorithms use "service data" which is, well, anything you type on google/google aided site.
The ethics behind this were questionable even 20 years ago.
Besides that, with these algorithms there is a greater probability of getting a wrong/informal/derogatory/rudimentary translations, which I guess nobody wants.

If you have no problem with it, that's fine. But, I've got a problem with it.

This update has been announced by Google directly, not a private issue. You can check the Google Blog links that i shared. You can find it even in Wikipedia ( https://en.wikipedia.org/wiki/Google_Neural_Machine_Translation ).

They even have made their neural network libary, Tensorflow, open sourced: https://www.tensorflow.org

Moderator
Joined: 16.01.2013

I noticed already a visible improvement in case of German translations. On one occasion the first half of the reported translation was correct. I was surprised.

On the other hand it shouldn't influence poetry, which cannot be translated literally.

Leader of the Balkan Squad
Joined: 14.10.2016

This may be great for Latin alphabet-based languages, but as far as I'm concerned, it's still pretty bad for Chinese and Greek.

Moderator
Joined: 16.01.2013

I checked now and I noticed one difference: earlier clicking on a translated word resulted in getting a choice of different words. Now these are expressions, still usually they are all wrongly translated.

Super Member
Joined: 11.10.2014

The problem with this sort of thing is that it's based on statistics, and last time I looked there was no way it was going to let syntax get in its way, which meant that it was going to be very wrong very often.
To a great extent, the translation will be hopless if the translated texts database the algorithm uses is too small, and this is of course particularly pernicious when combined with the exclusion of syntactic considerations.
Even for language pairs where there is a very large database things can be horribly wrong. Less than a decade ago you could guarantee that if google translate was asked to translate "He always pants after too much excercise" it would produce some nonsense involving trousers; I checked this out with a few friends so we had plenty of languages to play with, and discovered that the "pants - trousers" connection was enforced even when both languages had the verb completely different from the noun, indicating that translation was done as a two stage trick through some common pseudo-language or perhaps through English, presumably to try to offset a shortage of material direct from language A to language B that ensured statistical methods wouldn't work, but failed because the English-Other Language databases were too small even to get words in OED frequency band 5 right. I checked again a year or so ago, and things had improved a bit - but still the majority of languages were mistranslated more often than not (and "pants" was got wrong in translation between most pair of languages, but not in as many as it was earlier).
There has been some improvement recently. There is now a chance chance that when asked to translate a short (one clause, verb+subject_direct_object) sentence from Scottish Gaelic to English it will get at least two of the subject, the verb, and the direct object right, which is a vast improvement on a couple of years ago, but there's still no chance of it getting a Scottish Gaelic sentence containing subordinate clauses anywhere remotely near correct. It's slightly better with Irish, probably because it has access to a larger database as Irish is an official language of the EU, but still effectively useless unless you have some grasp of the language (my extremely weak Irish is enough to enable me to see google getting it wrong rather too often, sufficiently often that I don't trust anything it produces either from or to Irish unless I actually know it's ok).

whimsical chatterbox
Joined: 14.09.2013

As soon as these machines will reach even a barely adequate level (i.e. enough for a business conversation), these cheap tools will most likely replace costly human translators.

Just as search engines tend to discard pages containing uncommon opinions, these translators will probably drop the subtelties of languages as excess luggage. Less common meanings of words will be discarded or treated as parts of idiomatic constructs, i.e. deprived of some or all of their expressiveness. Less common languages will be disregarded as having too small a training set, too low a return on investment, etc.

As a consequence, understanding between non-speakers will be restricted to a lowest common denominator of factual statements and paradigms, in the most widely spoken languages. The world will be flooded by cheap automatic translations that will filter out whatever subtelty these machines won't be able to handle, and yet this will give the illusion of accessing other people's thoughts.

As if there were not enough other issues with it, this could ultimately damage our capacity to think. A kind of nuspeak forced upon mankind by machines of loving grace.

Editor || ᴋɪssᴇᴅ ʙʏ ғɪʀᴇ . ♔
Joined: 08.09.2014

This video pretty much sums up all that I would like to say: https://www.youtube.com/watch?v=SPAmbUZ9UKk

Super Member
Joined: 02.03.2013

AI learning how to translate = Good
AI with nuclear launch codes = Bad

whimsical chatterbox
Joined: 14.09.2013

We, as a species, don't seem to have learned that all red buttons are not worth pushing. I don't follow this amusing young guy's conclusions about super-AIs, but I sure think AI is really not the best thing to invent in a world where Google is more powerful than most, if not all, countries in the world.

Even without turning the solar system into heaps of fake handwritten flyers, these machines could bring the disatrous effects of "the winner takes all" principle to a whole new level.

Look at how Google came to be in undisputed control of our basic means of accessing information. After two decades, only a handful of competitors manage to get some cash out of cloud computing, and none could challenge Google's monopoly on Internet search.
As for sovereign states, they probably tried to control them politically but, as far as I know, no public authority is able to duplicate or even emulate their technology.

Though I don't believe they will achieve whatever consciousness any time soon, the AIs will certainly learn very quickly to perform a huge lot of tasks far better than humans.
Something comparable to industrial revolution, where peasants shrunk from a vast majority of the population to a few percents in the span of a few generations.
Only this is likely to happen much, much faster, and I wonder what kind of new jobs we are supposed to whip up to keep tens or hundreds of millions of obsolete blue and white collars busy, while a handful of little geniuses and smart guys will reap the benefits of all this automated manpower.

Thinking of the power these machines could grant to the first companies achieving breakthroughs is quite enough to keep me awake at night.

Guest
petit élève wrote:

...could challenge Google's monopoly on Internet search

There's DuckDuckGo, to give you at least a sense of privacy Teeth smile But you probably already knew that.

whimsical chatterbox
Joined: 14.09.2013

Yes, but it works not nearly as well as Google, and they are based on the Yahoo-Bing search engine anyway.

At the core you need hundreds of thousands of computers to index the Internet, and there are two, maybe three conglomerates in the world trying to do this : Google, the Microsoft-Yahoo alliance and the mysterious Baidu in China.

Decentralized cooperative search like YaCy is a joke in comparison. Microsoft and Yahoo are used ten times less than Google. The winner takes all...

Editor || ᴋɪssᴇᴅ ʙʏ ғɪʀᴇ . ♔
Joined: 08.09.2014

Even with DuckDuckGo, it is fruitless. There is no such thing as Internet security. Everything you do is tracked, you leave footprints everywhere. That is the reality of technology, but we are bound to it, regardless.

whimsical chatterbox
Joined: 14.09.2013

Exactly. Guys like DuckDuckGo are just filtering a small portion of our poisoned food while we live in a poisoned atmosphere.

Super Member
Joined: 06.02.2016

i use google TL very much and think its great that you can edit stuff now more easily. also realized this for the main google page, when you look for historic characters with missing birthdates or pictures.... TBH i don't really care how it works, long as it does the job.

whimsical chatterbox
Joined: 14.09.2013

These gifts might prove to come at too high a price one day. But then we won't be able to live without them, of course.

Super Member
Joined: 11.10.2014

My worry with that is that we'll be tricked into believing that it's doing the job when it isn't. For example if you ask it to translate "free the slaves" and it gives you "make the slaves' price zero" and you believe it you may end up supporting something you never intended to support. Every time I've asked it to translate anything involving the word "free" into a Gaelic language it has come up with "saor in aisce" (Irish: free of charge) or "saor an asgaidh" (Scottish: free of charge), and that hasn't improved with the recent changes.

Guest

So Orwellian.

I find it terrifying, too. The world already has enough problems.

Guest
vargkustaa wrote:

AI learning how to translate = Good
AI with nuclear launch codes = Bad

Too bad questions of ethics are the least of their worries. Sad smile