Connector that delivers the highest quality

[Nick]

Dear all,

Out of the different Connectors listed in Okapi, which one would you say provides the highest quality of Machine translation?

I imagine it depends on certain factors. In my case it is for translation of an e-learning within the social sciences domain, but the language is not that technical. It’s for translation from English to French and Spanish.

I’m looking for the Connector that delivers the highest quality translation using machine translation (not looking at price). But I have difficulty seeing the trees for the forest. Which one(s) would you suggest to use?

Best,
Nick

[Stephen Holmes]

Hi Nick

It’s totally dependent on the source content in terms of style, terminology and syntactic factors, such as formatting.  English to French and Spanish generally translated well in most of the popular public systems  (Spanish and Portuguese normally feature at the top of the quality outputs).

If you have a glossary for your eLearning content, then either custom systems or terminology adapting engines can be created for Google AutoML, Amazon, DeepL and  Microsoft MT and would give improved good output.  If you don’t have a glossary, then we’ve found Google AutoML still performs well in generalized language output, but DeepL and ModernMT are worth a look and can be superior or at least equivalent in many scenarios.

MT systems can be religious turf war of opinions, so this is just mine ;)

The simple summary here is that the output, without knowing the source, is non-determinate and generally speaking, it’d be wise to conduct a test using a scoring methodology such as TER and a Human In The Loop validation to help you make a selection.  There’s so much more that could be discussed from your original question, but I hope this helps!

Cheers

Stephen

--
You received this message because you are subscribed to the Google Groups "okapi-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to okapi-users...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/okapi-users/435f3ef5-4977-4bbf-9947-76ea783105e4n%40googlegroups.com.

[danielhug]

Hi Nick and Stephen,

I agree with you, Stephen, you can make this into a fine art. Generally speaking, if you have a lot of custom terminology AND translation memories/glossaries then looking into a customizable MT system is worth it. I am using Opus CAT MT, which is open source, you run it on your own machine. It's customizable, and it performs very well if you have sufficient TUs to train it. The out-of-the-box performance is medium. I can't see a connector for Okapi though. I used to use Modern MT, which is also customizable and very user-friendly.

You will have to see what works best for your language pairs and topics. I guess my main point is that customization can be a very time-saving feature if you have custom terminology. Okapi supports several customizable MT systems like Kantan (expensive) and ModernMT (much cheaper). Highly customizable systems often require very large TMs to really work well, so you have to see whether that applies to you.

Daniel