Is it possible to batch train an Adaptive MT engine?
Thread poster: Panopticon
Panopticon
Panopticon
Spain
Local time: 10:26
Member (2012)
Spanish to English
Jan 15, 2019

Howdy,

I have just bought a subscription to the Adaptive MT engine thingy and the initial output is pretty rough and unusable to be honest. I was wondering if anybody knows of a way to batch train the MT all at once rather than having to do it piecemeal over many segments/documents, which will take hours/days/weeks. Surely it should be possible to upload the source and target versions of a previous document that you have translated and have it do its calculations all at once and the
... See more
Howdy,

I have just bought a subscription to the Adaptive MT engine thingy and the initial output is pretty rough and unusable to be honest. I was wondering if anybody knows of a way to batch train the MT all at once rather than having to do it piecemeal over many segments/documents, which will take hours/days/weeks. Surely it should be possible to upload the source and target versions of a previous document that you have translated and have it do its calculations all at once and then use them in future translations rather than having to do it segment by segment? I am currently sitting on a gold mine of previous translations stretching back years that could be used. Maybe there is an add-on on the SDL website that can do it? Somebody must have thought of this before...

There is definitely a manual work around that could be used but it would still be pretty time-consuming, but I'm hoping there is an automatic way.

Anybody?
Collapse


 
Michael Beijer
Michael Beijer  Identity Verified
United Kingdom
Local time: 09:26
Member (2009)
Dutch to English
+ ...
Here's the app (but I recommend using a better adaptive MT engine to start with) Jan 15, 2019

Try this: https://appstore.sdl.com/language/app/adaptivemt-trainer/863/

However, I find the base quality level of their engines so bad that you would probably be better off using something good like DeepL, Google Translate (neural) or ModernMT (a relatively new player). I'm currently using ModernMT inside MateCat, and it is pretty darned good! It's also adaptive,
... See more
Try this: https://appstore.sdl.com/language/app/adaptivemt-trainer/863/

However, I find the base quality level of their engines so bad that you would probably be better off using something good like DeepL, Google Translate (neural) or ModernMT (a relatively new player). I'm currently using ModernMT inside MateCat, and it is pretty darned good! It's also adaptive, but the initial output quality is MUCH better than anything SDL has to offer.

info:

https://www.matecat.com/adaptive-mt-modernmt/
https://www.matecat.com/support/advanced-features/modernmt-mmt-plug/
https://www.modernmt.eu/

Michael

PS: For the very technical among us (not me!), ModernMT is also available as open-source software on GitHub (https://github.com/ModernMT/MMT ), so you can use the technology to create your own engine if you want

[Edited at 2019-01-15 23:03 GMT]
Collapse


Rex Su
 
Panopticon
Panopticon
Spain
Local time: 10:26
Member (2012)
Spanish to English
TOPIC STARTER
Thanks! Jan 16, 2019

Excellent, that is just what I was looking for! You are right, the initial output is definitely a bit shoddy. But here's hoping that it will improve with enough training, training which I have to pay for myself

Can the other Adaptive MTs you mentioned be used in Studio too?


 
Michael Beijer
Michael Beijer  Identity Verified
United Kingdom
Local time: 09:26
Member (2009)
Dutch to English
+ ...
some info Jan 16, 2019

Yes, ModernMT can, to a certain extent. I think the engine used in the 'MyMemory Pro' SDL App (which can be used inside Studio) is either the same or similar to the ModernMT (i.e. adaptive) one. I am currently in contact with the developers of ModernMT, but am not yet 100% sure how they relate/differ. Will post more info if and when I get some!

DeepL (which is one of the absolute best in terms of quality, at least in my pairs: Dutch-English) can be used in Studio, via an SDL app, b
... See more
Yes, ModernMT can, to a certain extent. I think the engine used in the 'MyMemory Pro' SDL App (which can be used inside Studio) is either the same or similar to the ModernMT (i.e. adaptive) one. I am currently in contact with the developers of ModernMT, but am not yet 100% sure how they relate/differ. Will post more info if and when I get some!

DeepL (which is one of the absolute best in terms of quality, at least in my pairs: Dutch-English) can be used in Studio, via an SDL app, but is not adaptive. The same applies to Google, although Google has also developed something adaptive, if I remember right, called 'AutoML' (https://slator.com/features/inside-googles-custom-neural-machine-translation-automl-translate/ ).

I think Microsoft also has an adaptive MT system (you can also use in Studio via an SDL app), and train with your own TMXs.

That is, lots to play with!

Michael
Collapse


 
RWS Community
RWS Community
United Kingdom
Local time: 10:26
English
Worth rephrasing I think... Jan 16, 2019

Michael Beijer wrote:
However, I find the base quality level of their engines so bad that you would probably be better off using something good like DeepL, Google Translate (neural) or ModernMT (a relatively new player). I'm currently using ModernMT inside MateCat, and it is pretty darned good! It's also adaptive, but the initial output quality is MUCH better than anything SDL has to offer.


".... MUCH better than anything SDL has to offer that I am prepared to pay for." If you are using the free SDL Language Cloud then I'd agree. But if you were using the NMT engines then I would not, but these are not free.


 
Michael Beijer
Michael Beijer  Identity Verified
United Kingdom
Local time: 09:26
Member (2009)
Dutch to English
+ ...
You might have a point Jan 16, 2019

SDL Community wrote:

Michael Beijer wrote:
However, I find the base quality level of their engines so bad that you would probably be better off using something good like DeepL, Google Translate (neural) or ModernMT (a relatively new player). I'm currently using ModernMT inside MateCat, and it is pretty darned good! It's also adaptive, but the initial output quality is MUCH better than anything SDL has to offer.


".... MUCH better than anything SDL has to offer that I am prepared to pay for." If you are using the free SDL Language Cloud then I'd agree. But if you were using the NMT engines then I would not, but these are not free.


Also, ModernMT definitely isn't cheap.

Will have to try out SDL's NMT offering one of these days and report back!

Michael


 
Panopticon
Panopticon
Spain
Local time: 10:26
Member (2012)
Spanish to English
TOPIC STARTER
Do I have NMT if I have subscribed to the Language Cloud? Jan 16, 2019

SDL Community wrote:

Michael Beijer wrote:
However, I find the base quality level of their engines so bad that you would probably be better off using something good like DeepL, Google Translate (neural) or ModernMT (a relatively new player). I'm currently using ModernMT inside MateCat, and it is pretty darned good! It's also adaptive, but the initial output quality is MUCH better than anything SDL has to offer.


".... MUCH better than anything SDL has to offer that I am prepared to pay for." If you are using the free SDL Language Cloud then I'd agree. But if you were using the NMT engines then I would not, but these are not free.


Do I have NMT (I'm guessing that means Neural Machine Translation?) if I have subscribed to Language Cloud? Or is this to be subscribed elsewhere?

Also is SDL's Adaptive MT confidential? Nobody will have access to it other than me? This is an important issue for me since I must respect my client's Confidentiality Agreements. A non-confidential service is a no-go for me, which obviously rules out Google Translate.


 
Panopticon
Panopticon
Spain
Local time: 10:26
Member (2012)
Spanish to English
TOPIC STARTER
DeepL confidential? Jan 16, 2019

Michael Beijer wrote:

Yes, ModernMT can, to a certain extent. I think the engine used in the 'MyMemory Pro' SDL App (which can be used inside Studio) is either the same or similar to the ModernMT (i.e. adaptive) one. I am currently in contact with the developers of ModernMT, but am not yet 100% sure how they relate/differ. Will post more info if and when I get some!

DeepL (which is one of the absolute best in terms of quality, at least in my pairs: Dutch-English) can be used in Studio, via an SDL app, but is not adaptive. The same applies to Google, although Google has also developed something adaptive, if I remember right, called 'AutoML' (https://slator.com/features/inside-googles-custom-neural-machine-translation-automl-translate/ ).

I think Microsoft also has an adaptive MT system (you can also use in Studio via an SDL app), and train with your own TMXs.

That is, lots to play with!

Michael


Thanks for the info Michael. Do you know if DeepL is confidential? I wouldn't be breaking any client confidentiality agreements by using it?


 
Panopticon
Panopticon
Spain
Local time: 10:26
Member (2012)
Spanish to English
TOPIC STARTER
Update: deactivated Language Cloud subscription after seeing no improvements from batch training Jan 17, 2019

So, I tried to batch train the so-called "Adaptive" MT using the add-on but after about 40,000 of training I saw absolutely no difference apart from maybe 5 or 6 words in a 20,000 word document. Since the baseline output is next to useless and the training seems to do nothing after 40,000 words (how many words training does it take to start to show up?) I think any "training" needs to done for free. If the baseline output were good to start with I might consider it but, being as poor as it is, I... See more
So, I tried to batch train the so-called "Adaptive" MT using the add-on but after about 40,000 of training I saw absolutely no difference apart from maybe 5 or 6 words in a 20,000 word document. Since the baseline output is next to useless and the training seems to do nothing after 40,000 words (how many words training does it take to start to show up?) I think any "training" needs to done for free. If the baseline output were good to start with I might consider it but, being as poor as it is, I decided to deactivate the subscription instead.

DeepL on the other hand, now that is something...
Collapse


 
Colin Smith
Colin Smith
United Kingdom
Local time: 09:26
French to English
+ ...
SDL Adaptive MT May 24, 2019

I cancelled my Adaptive MT subscription eventually as it really wasn't improving very much with use (from a low base). I now use Google NMT for financial and legal documents (once I worked out how to get the API key), which it handles quite well.

 


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


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

Is it possible to batch train an Adaptive MT engine?







CafeTran Espresso
You've never met a CAT tool this clever!

Translate faster & easier, using a sophisticated CAT tool built by a translator / developer. Accept jobs from clients who use Trados, MemoQ, Wordfast & major CAT tools. Download and start using CafeTran Espresso -- for free

Buy now! »
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! »