Meta's AI can translate dozens of resource-scarce languages

The technology behind Meta's artificial intelligence model, which can translate between 200 different languages, is NatureThis model increases the number of languages ​​that can be translated using machine translation.

Neural machine translation models utilize artificial neural networks to translate languages. These models typically require large amounts of online accessible data for training, but for some languages ​​(called “low-resource languages”), this data may not be public, cheap, or commonly available. Increasing the model's language output in terms of the number of languages ​​it translates can negatively impact the model's translation quality.

Marta Costa-jussà and the No Language Left Behind (NLLB) team developed a cross-lingual approach that enables neural machine translation models to learn how to translate low-resource languages ​​using their existing ability to translate high-resource languages.

As a result, the researchers developed an online multilingual translation tool called NLLB-200 that covers 200 languages, includes three times as many low-resource languages ​​as high-resource languages, and performs 44% better than existing systems.

Because the researchers only had access to 1,000–2,000 examples for many low-resource languages, they used a language identification system to identify more instances of specific dialects to increase the amount of training data for NLLB-200. The team also mined bilingual text data from the Internet Archive to help improve the quality of the translations provided by NLLB-200.

The authors note that the tool could help speakers of languages ​​with little translation access the Internet and other technologies. They further emphasize that the model has particular application in the education sector, as it could help speakers of low-resource languages ​​access more books and research papers. However, Costa-Giusa and his co-authors acknowledge that mistranslations can still occur.