Pages in topic:   < [1 2 3 4] >
Google gave their MT a "brain": Google Neural Machine Translation
Thread poster: Philippe Locquet
Merab Dekano
Merab Dekano  Identity Verified
Spain
Member (2014)
English to Spanish
+ ...
Machines Mar 7, 2017

Lingua 5B wrote:

Mervyn Henderson wrote:

Doesn't sound very safe.

But I'd be interested to know what they actually meant. Difficult to imagine.


Not very safe and not related in any way, shape or form to the topic this passage was talking about (a big machine assembly). Is that the "brain" Google gave their MT?

[Edited at 2017-03-07 15:58 GMT]


Well, you have to understand that these are machines, and machines get hot, I presume. Can’t think of any other explanation.


 
Matthias Brombach
Matthias Brombach  Identity Verified
Germany
Local time: 20:59
Member (2007)
Dutch to German
+ ...
Caution! Mar 7, 2017

Lingua 5B wrote:

"open up your p_n_s

Whenever a linguistically based system has already reached such a primitive level in terms of feelings and needs, it won't be far away from creating its own thoughts. Better be wise and lower your p..., I mean, rates!

[Edited at 2017-03-07 21:12 GMT]


 
Philippe Locquet
Philippe Locquet  Identity Verified
Portugal
Local time: 19:59
English to French
+ ...
TOPIC STARTER
So interesting, thank you! Mar 8, 2017

Hello to all,

I have found this to be fascinating and I want to thank you all for sharing your results.
Now, maybe this will help clarify a few things (I’m no scientist though…):

Neural networks are very different from statistical systems.

_Statistical systems will rely on a large amount of material and a very advanced algorithm come up with an answer from this.
_Neural networks (be it machine translation or picture recognition etc…) will
... See more
Hello to all,

I have found this to be fascinating and I want to thank you all for sharing your results.
Now, maybe this will help clarify a few things (I’m no scientist though…):

Neural networks are very different from statistical systems.

_Statistical systems will rely on a large amount of material and a very advanced algorithm come up with an answer from this.
_Neural networks (be it machine translation or picture recognition etc…) will rely on training. That is “programmers” will baby sit them and help them learn how to come up with the closest answer to the truth (plenty of material is out there explaining it for picture recognition). This is called “assisted training”.

1/Some languages may have benefited of more training than others (very true for English, French, German).
2/GT now is using a mixed MT (Neural + Statistical) so we’ll get a mixed answer.
3/Improvements may be seen as the language pairs get more training.
4/Some pairs will benefit from the “zero-shot”
Context: we’ll see if they come up with results…

It is very hard to tell how well these machines will perform in the future. But what we can witness is:
_Big investments in developing components and hardware specifically designed to ru neural networks.
_Consumers are using neural nets based products (safe-driving cars, internet services, etc.)
_Big educational investments on fresh IT specialists specialized in neural nets coming out of universities and the training of the next generation.
_Investments are made to move neural nets from assisted training to autonomous training (seems very far off though).

There is however no arguing: these are tools. There’s no comparison with the warmth and creativity of a seasoned translator, period. But when big companies see possible returns on an investment, that’s where they go. It’s only Google for now but Microsoft is probably just around the corner.
I cover most of the above in my YouTube video, if you’ve watched it until the end. There are also other videos on YouTube about neural machine translation, but these are in-depths discussions of neural-networks programming and training, very techy indeed.

By the way if someone gave a go at comparing the free web page GT and the paid API GT. I’d still like to hear how they compare. And I think many others would like to know (maybe because there’s money involved too).

My bests to all 😊
Collapse


 
Tom in London
Tom in London
United Kingdom
Local time: 19:59
Member (2008)
Italian to English
What really puzzles me about all this..... Mar 8, 2017

...is why the human race is devoting so much of its effort to inventing machines that are able to do *what we already know how to do much better*.

 
Anton Konashenok
Anton Konashenok  Identity Verified
Czech Republic
Local time: 20:59
French to English
+ ...
It's not about warmth and creativity Mar 8, 2017

Fi2 n Co wrote:

There is however no arguing: these are tools. There’s no comparison with the warmth and creativity of a seasoned translator, period.


The thing lacking most direly is not warmth and creativity but rather understanding of the source text. At this point, neural networks are not being taught to translate, they are merely taught to fake it reasonably well. In this capacity, they won't be able to exceed a freshly baked linguist who cocksurely thinks he can translate technical, medical or legal papers with no background in engineering, medicine or law whatsoever.


 
Lingua 5B
Lingua 5B  Identity Verified
Bosnia and Herzegovina
Local time: 20:59
Member (2009)
English to Croatian
+ ...
$$$ Mar 8, 2017

Tom in London wrote:

...is why the human race is devoting so much of its effort to inventing machines that are able to do *what we already know how to do much better*.


Because investors invested and are investing big money in development of these tools. Since there's large money at stake, they are doing whatever they can to justify the existence of these tools, promote their importance (even when there's none) and increase their sales, in order to achieve the target ROI. They are promoting them and "believe" in them even after they have proved to be pointless so many times - for the reasons already mentioned, see my title.



[Edited at 2017-03-08 09:38 GMT]


 
Michael Beijer
Michael Beijer  Identity Verified
United Kingdom
Local time: 19:59
Member (2009)
Dutch to English
+ ...
@Fi2 n Co: Mar 8, 2017

I'll try to do a few comparisons between using Google Translate via the API (in a CAT tool), and the online interface @ https://translate.google.co.uk/?ie=UTF-8&hl=en&client=tw-ob#auto/en/ when I have a moment. I do know that in the past (i.e., before they introduced the neural engines), there was a difference, so I'm curious too.

On a side note, Dragon Na
... See more
I'll try to do a few comparisons between using Google Translate via the API (in a CAT tool), and the online interface @ https://translate.google.co.uk/?ie=UTF-8&hl=en&client=tw-ob#auto/en/ when I have a moment. I do know that in the past (i.e., before they introduced the neural engines), there was a difference, so I'm curious too.

On a side note, Dragon NaturallSpeaking (now called "Dragon Professional Individual 15") also recently changed to a neural-net based system.

See e.g.:

https://techcrunch.com/2016/08/16/dragon-15/
http://www.lehsys.com/2016/09/review-dragon-professional-individual-version-15/
http://www.speechtechmag.com/Articles/Editorial/FYI/Nuance-Adds-Deep-Learning-to-Dragon-114607.aspx
http://www.nuance.com/company/news-room/press-releases/Dragon-New-Releases-Powered-by-Deep-Learning.docx

We are definitely living in interesting times!

Michael

[This post was dictated using Dragon Professional Individual 15. Please excuse any typos!]

[Edited at 2017-03-08 11:01 GMT]
Collapse


 
Mervyn Henderson (X)
Mervyn Henderson (X)  Identity Verified
Spain
Local time: 20:59
Spanish to English
+ ...
Because ... Mar 8, 2017

Tom in London wrote:

...is why the human race is devoting so much of its effort to inventing machines that are able to do *what we already know how to do much better*.



... they're well fast. Think of a monk translating in his freezing Middle Ages cell, very nearly at the bottom of the parchment now, tongue out at the side of his mouth as the quill moves to the right - "O Lord, .... thou art ..... gracious .... and righteous" - Right, time for prayers, better get meself down the cloisters with the brethren - comes back an hour later and says, taking up his quill, Now, where was I? Oh yes, righteous ... "righteous ... and we ... are unworthy ... of ..."Oh, for crying out loud, that's torn it, I've written "righteous" twice. It's either start the page again, or hope nobody notices. Why doesn't somebody invent something?"

But not everything's an improvement improvement. ... Oh look, see how easy it is to do?

[Edited at 2017-03-08 11:31 GMT]


 
Philippe Locquet
Philippe Locquet  Identity Verified
Portugal
Local time: 19:59
English to French
+ ...
TOPIC STARTER
Thanks! Mar 8, 2017

@@Michael Joseph Wdowiak Beijer

Thanks, I look forward to the results.

Michael Joseph Wdowiak Beijer wrote:

On a side note, Dragon NaturallSpeaking (now called "Dragon Professional Individual 15") also recently changed to a neural-net based system.



I didn’t know that. But I was guessing they’d use NNs. As a matter of fact, Google has a speech recognition plugin for Wordfast Anywhere, I tried it and it performed very well (I’m more of a typer though so I can’t compare very much). I’m guessing they’re using the same engine as with speech recognition on smartphones etc. That’s being pushed too because of the coming IoT (but that’s going off topic…).

I'll definitely check out the links, thanks!

My bests


 
José Henrique Lamensdorf
José Henrique Lamensdorf  Identity Verified
Brazil
Local time: 16:59
English to Portuguese
+ ...
In memoriam
Repetitive motion Mar 8, 2017

Tom in London wrote:

...is why the human race is devoting so much of its effort to inventing machines that are able to do *what we already know how to do much better*.


Machines can replicate human repetitive motion with almost unlimited speed.
They can't however, replicate THOUGHT.

The sewing machine is a perfect example. It can make hundreds of stitches while a seamstress does just one.
So why don't surgeons use sewing machines on us?
Because "telling" the machine WHERE and HOW to make each stitch would be much more troublesome and time-consuming than a surgeon doing the entire job, not to mention that one life will be at stake, and seconds are precious.

Would any Christian de Neuvillette dare to use computer-generated text in lieu of Cyrano de Bergerac to woo Roxanne?


 
Lingua 5B
Lingua 5B  Identity Verified
Bosnia and Herzegovina
Local time: 20:59
Member (2009)
English to Croatian
+ ...
On speed etc. Mar 8, 2017

José Henrique Lamensdorf wrote:

Tom in London wrote:

...is why the human race is devoting so much of its effort to inventing machines that are able to do *what we already know how to do much better*.


Machines can replicate human repetitive motion with almost unlimited speed.
They can't however, replicate THOUGHT.

The sewing machine is a perfect example. It can make hundreds of stitches while a seamstress does just one.
So why don't surgeons use sewing machines on us?
Because "telling" the machine WHERE and HOW to make each stitch would be much more troublesome and time-consuming than a surgeon doing the entire job, not to mention that one life will be at stake, and seconds are precious.

Would any Christian de Neuvillette dare to use computer-generated text in lieu of Cyrano de Bergerac to woo Roxanne?


They can't replicate maintenance either, nor can they reproduce themselves. So a human will be required to: manufacture the machine, operate it and maintain it - the three things that require both time and effort. As these processes take time, any potential time you save through automation, you will lose on maintenance and operation.


 
Giovanni Guarnieri MITI, MIL
Giovanni Guarnieri MITI, MIL  Identity Verified
United Kingdom
Local time: 19:59
Member (2004)
English to Italian
Italian not supported yet... Mar 8, 2017

from what I can gather...

Tom in London wrote:

Approximation will not do, in a translation. It must be exact.



 
Christine Andersen
Christine Andersen  Identity Verified
Denmark
Local time: 20:59
Member (2003)
Danish to English
+ ...
Machine brains cannot translate what they don't know Mar 8, 2017

There is always going to be trouble when there are cultural elements, and where one language has elements that do not exist in another.

In my main languages, for instance, English and Danish, everyone is equal as far as forms of address are concerned. There are probably still people who say 'thou' to God, or use the dialect 'tha', but everyone else, singular or plural is 'you'. (Which linguistically is the equivalent of French vous, German Sie, the respectful address. Thou and the p
... See more
There is always going to be trouble when there are cultural elements, and where one language has elements that do not exist in another.

In my main languages, for instance, English and Danish, everyone is equal as far as forms of address are concerned. There are probably still people who say 'thou' to God, or use the dialect 'tha', but everyone else, singular or plural is 'you'. (Which linguistically is the equivalent of French vous, German Sie, the respectful address. Thou and the plural ye are in fact the familiar forms.)

In Denmark the queen recently corrected a journalist who addressed her as 'du' instead of 'De', and the respectful form is seen occasionally, but apart from Her Majesty, most people normally use the familiar form. As with the German Sie, the Danish can easily be confused with third person plural, apart from the capital letter.

It will be extremely difficult for MT to work out which forms of address - and which corresponding verbs - to use when translating out of English and Danish into other languages that do make the distinction, while a human does it almost instinctively, or can very quickly work it out 100% correctly.

Grammatical gender and inflections will probably cause a lot of trouble in some language pairs, and that is just at a very basic level.

In other situations, there are simply no exact equivalents. Many polite expressions fall into this category, and I am sure MT will have real trouble in sorting them out. Here literal translation is hopeless, and only a human familiar with the language and culture will know what is actually said in a given situation. It will be impossible to code in all the possible variants and rules for when to use them.
________________________

I agree with Tom in London - it is incredible how much time humans can spend trying to get machines to do things that humans do far better and more easily themselves.
I can understand getting robots to do heavy, noisy, repetitive work that ruins health. The more the better!

But turning a satisfying skill like translation into a PITB like post-editing is beyond my comprehension!
Collapse


 
John Fossey
John Fossey  Identity Verified
Canada
Local time: 15:59
Member (2008)
French to English
+ ...
Understanding the subject Mar 8, 2017

Anton Konashenok wrote:

The thing lacking most direly is not warmth and creativity but rather understanding of the source text.


That is key. Remember that the human brain is also a neural network, many, many orders of magnitude deeper than any machine-based neural network. And even so, it takes a human years - decades - of life and subject experience to develop that understanding before one can be expected to translate accurately.


 
Daryo
Daryo
United Kingdom
Local time: 19:59
Serbian to English
+ ...
What will happen when A.I. becomes self aware? Mar 13, 2017

LegalTransform wrote:

What will happen when A.I. becomes self aware and starts demanding .30 a word?


you are making the same mistake as those who judge/think about the past using categories from the present

when A.I. becomes self aware it will be thinking in its own categories, and somehow I doubt that ".30 a word" will be part of it ... (not that it will be of any relief for the old-type intelligence)

end of digression


 
Pages in topic:   < [1 2 3 4] >


To report site rules violations or get help, contact a site moderator:


You can also contact site staff by submitting a support request »

Google gave their MT a "brain": Google Neural Machine Translation






Anycount & Translation Office 3000
Translation Office 3000

Translation Office 3000 is an advanced accounting tool for freelance translators and small agencies. TO3000 easily and seamlessly integrates with the business life of professional freelance translators.

More info »
Wordfast Pro
Translation Memory Software for Any Platform

Exclusive discount for ProZ.com users! Save over 13% when purchasing Wordfast Pro through ProZ.com. Wordfast is the world's #1 provider of platform-independent Translation Memory software. Consistently ranked the most user-friendly and highest value

Buy now! »