Hello to all!
I’m making a little post to introduce you to a technology that I find interesting and that I don’t hear much about. You may know ModernNMT, a company that offers translation services, focused on translators. I tested their online translator and I found it quite powerful, and by digging a bit I saw that their translation system is open source (based on fairseq) and quite easy to use. I think it can be interesting for those who want to have multilingual translation models without too much training difficulty. Of course there is always the problem of resources, the models you can use on the website are probably trained on a huge amount of data and so few people can reproduce this level of training, but it can still be useful.
Their website: ModernMT | Translate
Their open source engine: GitHub - modernmt/modernmt: Neural Adaptive Machine Translation that adapts to context and learns from corrections.
note: I didn’t know where to share this information so don’t hesitate to tell me if this is not the right place to^^
I personally tried it, few months back. While it’s really promising, I believe it would be nice if we could finetune their algorithm that choose which of your data is relevant or not for training.
When I did my test, often some terms where not translated properly, because their “cosine search” didn’t find a segment that was matching their threshold. So unless the current segment you are translating is really similar to one in your data, the training part will not take it. But you might just have one terme in the segment you want to translate that you have all over the place and that is always translate in a certain way and then it’s randomly not translated properly has the training is different per segment. So it’s render an inconstant translation quality.
I had much better results with Google Auto ML which the translation is really “constant”, but much more pricey.
Yes, I personally lean more towards OpenNmt but the advantage of this tool is that it is extremely easy to use. After that it doesn’t receive updates very regularly.
In any case, thank you for your answer