Meta has developed an artificial intelligence (AI) tool capable of translating 200 different languages, including many that are under-resourced. 

This breakthrough, spearheaded by French and US researchers at Meta, has been detailed in a recent paper published in Nature. 

The innovation marks a significant advancement in the field of neural machine translation.

Traditional AI translation models require vast amounts of online data to train effectively. However, for languages that are not widely documented or easily accessible, known as low-resource languages, gathering sufficient data can be a challenge. 

These languages often lack the extensive digital presence necessary for conventional AI training methods, leading to difficulties in maintaining translation quality.

To address this issue, Meta’s No Language Left Behind (NLLB) team, led by Marta Costa-jussà, developed a cross-language approach. 

This method leverages the AI’s ability to translate high-resource languages to enhance its performance with low-resource languages. 

The resulting model, named NLLB-200, includes three times as many low-resource languages as high-resource ones and performs 44 per cent better than existing systems.

Costa-jussà’s team used a combination of language identification systems and mined bilingual textual data from internet archives to increase the volume of training data for these low-resource languages. 

Despite having access to only 1,000 to 2,000 samples for many of these languages, the team successfully expanded the data set, significantly improving translation quality.

The researchers believe that this tool can greatly benefit speakers of rarely translated languages, providing them with greater access to internet resources and technologies. 

The tool holds particular promise for educational applications, enabling speakers of low-resource languages to access a broader array of books and research articles. However, the team acknowledges that mistranslations may still occur.

The development of NLLB-200 underscores the potential of AI to bridge linguistic gaps, promoting digital inclusivity and expanding access to information across linguistic barriers. 

By making all contributions to this effort freely available for non-commercial use, Meta says it aims to lay the groundwork for a universal translation system that can serve a global audience    .

The full study is accessible here.