According to the tech giant, most translation models currently being used rely heavily on English when performing translations. For example, when translating Chinese to French, these AI will first translate Chinese into English, and then English to French, as English training data is the most available and easy to obtain. However, this antiquated process decreases translation accuracy, which has now been significantly improved thanks to Facebook’s new multilingual machine translation model.
Unlike its predecessors, the new AI — dubbed the M2M-100 — is capable of translating Chinese to French (or any of the other 98 languages) directly without that middle stage. To achieve this feat, Facebook trained the model on a total of 2,200 language directions, which it claims is 10 times more than the previous best, English-centric multilingual models out there. 7.5 billion sentences across 100 languages were used to build a massive data set, from with another 15 billion parameters were created.
“For years, AI researchers have been working toward building a single universal model that can understand all languages across different tasks,” says Facebook. “A single model that supports all languages, dialects, and modalities will help us better serve more people, keep translations up to date, and create new experiences for billions of people equally. This work brings us closer to this goal.”
Source: Hypebeast Kr