Facebook will be able to translate up to 100 languages ​​without using English

Facebook developed the first model of automatic translation multilingual, capable of translate up to 100 languages ​​without using English as an intermediary. The system, called the M2M-100, uses artificial intelligence.

According to Facebook research assistant Angela Fan, this is an important step towards a universal model that understands all languages ​​in different tasks. The company has not yet released information on when the model will be implemented. So far, the technology is just a research project.

How the study was conducted

a Facebook/Disclosure

Initially, the research team collected 7.5 billion sentence pairs from the internet in 100 different languages, giving priority to the translations most requested by Internet users.

Then, the languages ​​were separated into 14 groups, based on linguistic, geographic and cultural similarities. Um of these groups, for example, includes common languages ​​of the India, como hindi, bengali e marata. To make it easier for people to understand, the team decided to create translation bridges.

In the case of Indian languages, Hindi, Bengali and Tamil served as intermediaries for the Indo-Aryans. With this technique, the company says that outperformed English-centric systems by 10 points in the BLEU metric, which evaluates machine translations, it reached the mark of 20.1.

Comparison between the new translation model, with 20.1 points in the BLEU metric;  and the current model, with only 16.7 points.
Comparison between the new translation model, with 20.1 points in the BLEU metric; and the current model, with only 16.7 points.Source: Facebook/Disclosure

“When translating, say, Chinese to French, most English-centric multilingual models train from Chinese to English and from English to French, because English training data is widely available,” explained Angela Fan. “Our model trains directly on Chinese to French data to better preserve meaning.”

Although it has not yet been incorporated into the Facebook, where users post content in over 160 languages, tests performed by the team indicate that the model can support a wide range of translations.

Previous post 10 tech news to start the day (10/14)
Next post Snap Layout: How to Resize and Organize Windows in Windows 11