Meta's NLLB Project Breaks 200-Language Barrier in Machine Translation
A new AI model called No Language Left Behind translates over 200 languages to help close the gap between high and low-resource languages. The system uses a sparsely gated mixture of experts architecture and undergoes rigorous safety evaluations to ensure high-quality, ethical results.
A new initiative called No Language Left Behind (NLLB) takes on the massive challenge of translating over 200 languages to eliminate global communication barriers. The project focuses specifically on low-resource languages that traditional artificial intelligence research historically ignores. To ensure the technology meets actual human needs, the research team conducts exploratory interviews with native speakers before building their translation tools.
The researchers develop a conditional compute model relying on a Sparsely Gated Mixture of Experts architecture to handle this immense linguistic diversity. They train this system on data gathered through novel mining techniques designed specifically for low-resource languages. To prevent the model from overfitting across its thousands of translation tasks, the team introduces several key architectural and training improvements.
Rigorous evaluation remains a core component of this human-centered approach, with the team testing over 40,000 different translation directions using the human-translated Flores-200 benchmark. The project also implements a new toxicity benchmark covering all included languages to thoroughly assess and guarantee translation safety. Ultimately, this comprehensive system achieves a 44 percent improvement in BLEU scores compared to the previous state-of-the-art models.