Neural language processing is a type of A.I. that is designed to process language — but bringing features powered by the tool to a new language takes almost as much work as building a completely new program. To solve this, Facebook engineers created multilingual embeddings. This technique classifies words that have similar meanings — such as soccer and fútbol — together, placing them with other similar words inside the embed.
By grouping words with similar meanings together, a single program can be trained in multiple languages. Facebook developers say that this works even if the language wasn’t used while training the A.I. For example, if soccer and fútbol are classified together, the program will recognize fútbol even when the program was only trained using the English soccer. Facebook says the update increases translation accuracy to 95 percent, with 20 to 30 times more speed for processes using natural language processing.
Before the multilingual embeddings, Facebook took one of two approaches to bringing features to different languages. One, they could create a separate set of training data for each language, but re-training for every language would mean delays in bringing the feature to a global audience. The second option previously used by Facebook was to run the text through a translation, then plug the English text into the English-trained A.I. Translation errors, however, reduced accuracy, and again, speed was also a drawback.
With the changes, Facebook users worldwide will see features in their language faster, including M Suggestions inside Messenger and Recommendations. Overall, using Facebook across multiple languages will improve, Facebook says.
But the update has implications even for English users that aren’t bilingual. The programs that detect posts that are against Facebook’s policies will gain more accuracy across more languages, Facebook says, supporting the platform’s ongoing work against hate speech, terrorism, and fake news.
Facebook’s engineers said that while the program shows improvements in English, German, French, and Spanish, there is more work to do to add data from additional languages. An expansion could help the software better understand cultural phrases that may not have literal translations, like “raining cats and dogs.”
