Meta's influential paper in the authoritative academic journal Nature introduces the "No Language Left Behind" plan. This plan is a highly significant large-scale multilingual model that fully utilizes the powerful advantages of transfer learning.
In this plan, the research team meticulously developed a conditional computational model based on the "Sparsely Gated Mixture of Experts" architecture. To optimize the training of this model, researchers implemented new data mining techniques tailored for low-resource languages.
To prevent overfitting during training on a vast array of tasks, the research team adopted various architectural and training enhancements. They utilized a series of specially designed tools for a comprehensive evaluation of the model's performance. This included the automatic benchmarking tool FLORES-200, the human evaluation metric XSTS, and a toxicity detector covering all languages. With these tools, the research team conducted an in-depth and meticulous assessment of the model's performance across as many as 40,000 translation directions.
The results showed that compared to previous state-of-the-art models, Meta's new model achieved a significant improvement in translation quality, with an average increase of 44% in BLEU scores. This achievement fully demonstrates the team's outstanding innovative capabilities in the field of multilingual translation. Through this project, the research team successfully demonstrated how to extend Neural Machine Translation (NMT) to 200 languages, paving new avenues for global multilingual communication and information dissemination.