The quality of the AI ​​text translation system has improved by 44%, using over 50 billion parameters to translate 200 languages

Meta Platforms today opened the system code of NLLB-200, an artificial intelligence system developed internally by Meta that can translate text in 200 languages.

Meta also announced a set of tools designed to help researchers more easily apply NLLB-200 to software projects.

According to Meta, many of the 200 languages ​​NLLB-200 can understand are not well supported by other AI translation systems. While currently widely used translation tools support fewer than 25 African languages, the NLLB-200 supports up to 55 African languages.

Translation accuracy is another area where NLLB-200 outperforms other tools, Meta said. Meta uses the accuracy standard of the BLEU evaluation system, an algorithm used to measure the quality of machine-translated text. According to Meta, the BLEU score of NLLB-200 is 44% higher on average than before.

Meta CEO Mark Zuckerberg said: "We just open sourced a home-grown AI model that can translate 200 different languages ​​- many of which are not supported by current translation systems. We put This project is called No Language Left Behind, and the artificial intelligence modeling technology we use is producing high-quality translations for languages ​​spoken by billions of people around the world.”

NLLB-200 has more than 50 billion Parameters, these configurations determine how the AI ​​system processes data. The more parameters an AI system has, the higher its accuracy will be.

The NLLB-200’s large number of parameters is not the only factor in its ability to support 200 languages ​​with high accuracy, as the NLLB-200 system also draws on many other AI innovations developed by Meta engineers.

Meta uses the in-house developed LASER toolkit to provide support for machine learning related research. Using the toolkit, researchers can train a neural network to perform a specific task in one language and then relatively easily adapt the neural network to other languages, which would be useful for translation purposes. Meta has developed a new NLLB-200 system that supports an improved version of LASER - LASER3.

The original version of LASER included a neural network called an LSTM, a specialized component that converts text into a mathematical representation that the AI ​​system can understand. This mathematical representation helps produce more accurate translations. In LASER3, Meta replaces the LSTM neural network with Transformer, an advanced natural language processing model that can perform the same task more efficiently.

Meta also used several other methods to improve the capabilities of the NLLB-200, such as Meta upgrading the system used to collect training data and making changes to the AI ​​training workflow.

Meta uses the in-house developed Research SuperCluster supercomputer (pictured) to train NLLB-200. When Meta first introduced Research SuperCluster in January this year, it said that the system was equipped with 6,080 Nvidia's latest A100 data center GPUs and would eventually be upgraded to 16,000 GPUs.

Meta plans to use NLLB-200 to provide better automatic translation capabilities on Facebook, Instagram and other platforms, and the system is expected to support more than 25 billion translations per day.

While Meta is working to promote NLLB-200 internally, it also plans to help other enterprise organizations implement the system into their own software projects.

In addition to NLLB-200, Meta has open sourced code that can be used to train AI, as well as a dataset called FLORES-200 for evaluating translation accuracy. Meta will provide up to $200,000 in funding to help nonprofits adopt NLLB-200. In addition, Meta will also cooperate with the Wikimedia Foundation to apply automatic translation technology to Wikipedia articles.

The above is the detailed content of The quality of the AI ​​text translation system has improved by 44%, using over 50 billion parameters to translate 200 languages. For more information, please follow other related articles on the PHP Chinese website!