META Engineers Build a Universal Translator of 200 Languages in Open Source Software

IBL News | New York

META, the social media conglomerate META, will open source a newly engineered AI model that can translate into 200 different languages.

This tool, called No Language Left Behind, is part of Meta’s R&D project to create a universal speech translator, which would help the growth of Facebook, Instagram, virtual reality destinations, and other of its social media properties. The open source approach — with the code available on GitHub — is taken to attract more technicians to the development work.

In an abstract, Meta explains that this universal translator is “driven by the goal of eradicating language barriers on a global scale” with the help of machine learning technologies.

From a technical perspective, “it’s a conditional compute model based on Sparsely Gated Mixture of Experts that is trained on data obtained with novel and effective data mining techniques tailored for low-resource languages.”

The model, that lays the groundwork for a universal translation system, achieves an improvement of 44% BLEU (which stands for Bilingual Evaluation Understudy).

Technical Description of the Project (PDF)