Hi
I ran an empiric test on 3 services, namely Google, ModernMT, and
LibreTranslate, all for the language I translate, pt-BR. The test was
run on untranslated Help strings and when weblate translation memory
gives no match.
I found Google the most accurate. To be honest, it seems Google knows
the string to translate belongs to an office suite spreadsheet Help
string...
ModernMT not good enough, some basic context mistakes.
LibreTranslate quite good actually and promising. My remarks on
LibreTranslate:
* Occasionally, an extra period before a closing tag.
* occasionally, errors on tags: <input>BETA.INV...</input> returned
<input/>BETA.INV...
* the more tags in the string, the more it get confused and renders poorly.
* Admittedly, some strings in English are poorly written. The software
does no miracles, poor English returns poor translation.
* Mistakes in gender determination, if the 2 above are true
* Lousy to detect UI terms (Dialog boxes elements) from normal nouns.
* With confusing English, The usual "word-by-word" translation, returns
English-y Portuguese that demands a manual intervention.
However, I still think translators will benefit to have machine
translation as suggestion alongside memory translation. If
LibreTranslate can "learn" form existing translation memory, I'll expect
improvements.
test strings run on LibreTranslate
* minor intervention
** major intervention
https://translations.documentfoundation.org/translate/libo_help-master/textscalc01/pt_BR/?q=state%3A%3Ctranslated&offset=8
https://translations.documentfoundation.org/translate/libo_help-master/textscalc01/pt_BR/?q=state%3A%3Ctranslated&offset=9
*
https://translations.documentfoundation.org/translate/libo_help-master/textscalc01/pt_BR/?q=state%3A%3Ctranslated&offset=10
https://translations.documentfoundation.org/translate/libo_help-master/textscalc01/pt_BR/?q=state%3A%3Ctranslated&offset=11
*
https://translations.documentfoundation.org/translate/libo_help-master/textscalc01/pt_BR/?q=state%3A%3Ctranslated&offset=15
https://translations.documentfoundation.org/translate/libo_help-master/textscalc01/pt_BR/?q=state%3A%3Ctranslated&offset=16
https://translations.documentfoundation.org/translate/libo_help-master/textscalc01/pt_BR/?q=state%3A%3Ctranslated&offset=17
https://translations.documentfoundation.org/translate/libo_help-master/textscalc01/pt_BR/?q=state%3A%3Ctranslated&offset=24
**
https://translations.documentfoundation.org/translate/libo_help-master/textscalc01/pt_BR/?q=state%3A%3Ctranslated&offset=25
*
https://translations.documentfoundation.org/translate/libo_help-master/textscalc01/pt_BR/?q=state%3A%3Ctranslated&offset=26
*
Regards
Olivier
Em 23/02/2022 11:53, sophi escreveu:
Hi all,
Le 23/02/2022 à 15:46, sophi a écrit :
Hi all,
Following the request made by Stanislav and Olivier on automatic
translation built-in support in Weblate, we discussed the topic
yesterday during our team meeting.
Currently we rely on automatic suggestions from Weblate TM and amaGama
TM, meaning there is no AI added.
The possibility to include Aws, Microsoft, Google, DeepL, Baidu and
more is already there in Weblate, see:
https://docs.weblate.org/en/latest/admin/machine.html
Doing some research on the different translators available, I've seen
some are really high in costs and/or do not cover a large range of
languages. Traductor requests to be added in credits, I don't know for
others.
Yesterday, Olivier proposed Matecat (https://www.matecat.com/) which
is an open source project. He is using it with the Brazilian community
and it gives them very good results. From what I've tested, it
supports a large range of languages, can be used locally. However not
relevant here, I found their support for a large type of file formats
very interesting too.
Could you give a test to Matecat and let me know:
So it seems following Ilmari's researches that my request to test
Matecat is irrelevant. I'm copying his mail in the other thread here:
-----------------------------------------
I looked more deeply into Matecat now - I never before managed to find
information about the way it uses machine translation. It looks like
Matecat itself will not give us any benefits. See these resources:
https://guides.matecat.com/my
"By default, we provide Google Translate API (we pay for it and offer it
for free to Matecat users). If for any reason the Google Translate API
is not available, we automatically switch to the Microsoft Bing API."
https://guides.matecat.com/machine-translation-engines
It looks like ModernMT is the system that is associated with Matecat
creators:
https://guides.matecat.com/modernmt-mmt-plug-in
It is open source, even though it has a higher quality version for
enterprises:
https://github.com/modernmt/modernmt
Weblate does support it, also with a self-hosted version:
https://github.com/WeblateOrg/weblate/blob/main/weblate/machinery/modernmt.py
https://github.com/WeblateOrg/weblate/issues/1627#issuecomment-657171205
However, Weblate also supports LibreTranslate, which uses OpenNMT at its
core:
https://github.com/WeblateOrg/weblate/blob/main/weblate/machinery/libretranslate.py
https://libretranslate.com/
https://github.com/LibreTranslate/LibreTranslate
https://github.com/argosopentech/argos-translate
https://opennmt.net/
Could you please also evaluate LibreTranslate in addition to ModernMT?
--------------------------------------------------
Thanks a lot Ilmari for your feedback. Please follow his request to
evaluate LibreTranslate.
Cheers
Sophie
--
Olivier Hallot
LibreOffice Documentation Coordinator
Rio de Janeiro - Brasil - Local Time: UTC-03:00
LibreOffice – free and open source office suite: https://www.libreoffice.org
Respects your privacy, and gives you back control over your data
http://tdf.io/joinus
--
To unsubscribe e-mail to: l10n+unsubscribe@global.libreoffice.org
Problems? https://www.libreoffice.org/get-help/mailing-lists/how-to-unsubscribe/
Posting guidelines + more: https://wiki.documentfoundation.org/Netiquette
List archive: https://listarchives.libreoffice.org/global/l10n/
Privacy Policy: https://www.documentfoundation.org/privacy
Context
Privacy Policy |
Impressum (Legal Info) |
Copyright information: Unless otherwise specified, all text and images
on this website are licensed under the
Creative Commons Attribution-Share Alike 3.0 License.
This does not include the source code of LibreOffice, which is
licensed under the Mozilla Public License (
MPLv2).
"LibreOffice" and "The Document Foundation" are
registered trademarks of their corresponding registered owners or are
in actual use as trademarks in one or more countries. Their respective
logos and icons are also subject to international copyright laws. Use
thereof is explained in our
trademark policy.