The Fedora Machine Learning SIG has a proposal on tutorials on how to deploy machine learning applications. One area of application of machine learning is translation. Following are some open source Machine translation engines: http://opennmt.net/, https://github.com/facebookresearch/UnsupervisedMT, https://github.com/moses-smt/mosesdecoder, http://thumt.thunlp.org/, https://marian-nmt.github.io/
Are there preferences on which of these is of most interest or if there are others that should be considered? It is likely the case that the best settings and translation engines will be language dependent. Interfacing these to Weblate will help improve the translation speed and reduce redundant work.
An introduction to machine learning powered translation can be found here: https://www.paddlepaddle.org.cn/documentation/docs/en/1.4/beginners_guide/ba...
Le 2019-08-27 14:14, Benson Muite a écrit :
Are there preferences on which of these is of most interest or if there are others that should be considered?
Hello Benson, sounds really interesting.
I doubt integrating Weblate with any machine translation server would really be an issue.
Question to be asked are: * do other big open-source project use similar tools? * how can we find synergy with other open-source projects? * how easy is it to contribute to the project and maintain the server?
I helped Friedel Wolff to pull content from Zanata for Amagama: https://github.com/translate/amagama-updater It may help us to share the burden of pulling existing translated content to feed our server. Maybe should we also give amagama a try?
I know a contributor told me at Flock Mozilla is doing such thing but have no idea which tools they are using.
Have a nice day
My general suggestion is it could be a dumb TM, but if any sort of AI algorithms involved they should be open source, the only one I know from the list you provided is OpenNMT.
On Tue, Aug 27, 2019, 15:15 Benson Muite benson_muite@emailplus.org wrote:
The Fedora Machine Learning SIG has a proposal on tutorials on how to deploy machine learning applications. One area of application of machine learning is translation. Following are some open source Machine translation engines: http://opennmt.net/, https://github.com/facebookresearch/UnsupervisedMT, https://github.com/moses-smt/mosesdecoder, http://thumt.thunlp.org/, https://marian-nmt.github.io/
Are there preferences on which of these is of most interest or if there are others that should be considered? It is likely the case that the best settings and translation engines will be language dependent. Interfacing these to Weblate will help improve the translation speed and reduce redundant work.
An introduction to machine learning powered translation can be found here:
https://www.paddlepaddle.org.cn/documentation/docs/en/1.4/beginners_guide/ba... _______________________________________________ trans mailing list -- trans@lists.fedoraproject.org To unsubscribe send an email to trans-leave@lists.fedoraproject.org Fedora Code of Conduct: https://docs.fedoraproject.org/en-US/project/code-of-conduct/ List Guidelines: https://fedoraproject.org/wiki/Mailing_list_guidelines List Archives: https://lists.fedoraproject.org/archives/list/trans@lists.fedoraproject.org
Mozilla is using Google Translate in Pontoon right now (they used Bing in the past), so in this case, I don't think we should follow suit :)
Yaron Shahrabani sh.yaron@gmail.com ezt írta (időpont: 2019. aug. 27., K 15:05):
My general suggestion is it could be a dumb TM, but if any sort of AI algorithms involved they should be open source, the only one I know from the list you provided is OpenNMT.
On Tue, Aug 27, 2019, 15:15 Benson Muite benson_muite@emailplus.org wrote:
The Fedora Machine Learning SIG has a proposal on tutorials on how to deploy machine learning applications. One area of application of machine learning is translation. Following are some open source Machine translation engines: http://opennmt.net/, https://github.com/facebookresearch/UnsupervisedMT, https://github.com/moses-smt/mosesdecoder, http://thumt.thunlp.org/, https://marian-nmt.github.io/
Are there preferences on which of these is of most interest or if there are others that should be considered? It is likely the case that the best settings and translation engines will be language dependent. Interfacing these to Weblate will help improve the translation speed and reduce redundant work.
An introduction to machine learning powered translation can be found here:
https://www.paddlepaddle.org.cn/documentation/docs/en/1.4/beginners_guide/ba... _______________________________________________ trans mailing list -- trans@lists.fedoraproject.org To unsubscribe send an email to trans-leave@lists.fedoraproject.org Fedora Code of Conduct: https://docs.fedoraproject.org/en-US/project/code-of-conduct/ List Guidelines: https://fedoraproject.org/wiki/Mailing_list_guidelines List Archives: https://lists.fedoraproject.org/archives/list/trans@lists.fedoraproject.org
trans mailing list -- trans@lists.fedoraproject.org To unsubscribe send an email to trans-leave@lists.fedoraproject.org Fedora Code of Conduct: https://docs.fedoraproject.org/en-US/project/code-of-conduct/ List Guidelines: https://fedoraproject.org/wiki/Mailing_list_guidelines List Archives: https://lists.fedoraproject.org/archives/list/trans@lists.fedoraproject.org
Le 27/08/2019 à 15:04, Yaron Shahrabani a écrit :
My general suggestion is it could be a dumb TM, but if any sort of AI algorithms involved they should be open source, the only one I know from the list you provided is OpenNMT.
I take your comment as an occasion to make a little explanation.
With Weblate, we will have a translation memory server. It's what we abbreviate as TM.
In our hosting plan, I'll ask this translation memory server to be on a second server, on side of Weblate's main server. The user experience will be unchanged, results will still be displayed in Weblate.
My idea is that:
* TM server could have really different requirement in terms of resources,
* we should be able to feed it with content coming from other open-source initiatives,
* we may share it with one or more open-source initiatives,
* we may move it to our infrastructure (I suspect we'll require some automation at some point)
I understand that Benson would like to work on a machine translation server. It's what we abbreviate as MT. It makes it really easy to get confused, I did the mistake myself in the council ticket...
This machine translation server will display suggestions in addition of the translation memory server, and teams will be able to use it if it is relevant for there language.
This is a really welcome initiative, but this won't be part of the hosting plan we asked Council money for. This server will most probably use the community openshift infrastructure, which means Benson and other volunteer in our community will have to administrate it.
Note that both translation memory and machine translation servers only get better with more source content, it's a virtuous circle that benefits everyone.
Jean-Baptiste
-
Balázs Meskó -
Benson Muite -
Jean-Baptiste -
Jean-Baptiste Holcroft
-
Yaron Shahrabani