Facebook M2M-100 Language Translation Model Now Open Sourced

[engine]

Facebook has a new AI multilingual, 100 language translation model, M2M-100, which it releasing to open source. The company says it doesn't use English-centric translation, which makes it more accurate and better preserves the meaning by translating direct from language to language.

When translating, say, Chinese to French, most English-centric multilingual models train on Chinese to English and English to French, because English training data is the most widely available. Our model directly trains on Chinese to French data to better preserve meaning. It outperforms English-centric systems by 10 points on the widely used BLEU metric for evaluating machine translations.

[about.fb.com...]

Developers can find this on GitHub. [github.com...]

[JorgeV]

Hello,

Good. Like that, instead of having our content republished by scrappers, they'll republish it tens of times, translated in different languages...

[tangor]

Chuckles. One problem with BLEU scores is they favor short translations ... as in 3-5 words. Anything above that you still won't get a native speaker (lang 1) to (lang 2) by a practiced translator of both languages.

Is this an advance? YES! Ten points above a C- is an advance.

Star Trek's universal translator is not here yet.