With more than 2 billion users, Facebook Inc. has to deal with dozens of different languages on its social network, which poses a bit of a barrier to the company’s new mission to “bring the world closer together.” Facebook hopes that artificial intelligence will be the answer to this problem, and today the company announced that its translations now rely entirely on cutting-edge neural machine learning.

In a blog post published today, Facebook researchers Juan Miguel Pino, Alexander Sidorov and Necip Fazil Ayan explained just how hard it is to deal with so many languages.

“Creating seamless, highly accurate translation experiences for the 2 billion people who use Facebook is difficult,” they said. “We need to account for context, slang, typos, abbreviations, and intent simultaneously.”

According to the team, Facebook recently switched its backend translation systems entirely to neural networks, which handle more than 2,000 translation directions and 4.5 billion translations every day. They say that these translations are more accurate than Facebook’s previous system, which used phrase-based machine translation models.

Facebook is not the only company to use phrase-based translation, nor is it the first to begin moving away from it. Google Translate has also used phrase-based translation, and like Facebook, Google has also been shifting to neural machine translation over the last year with its aptly named Google Neural Machine Translation.

Phrase-based translation breaks sentences down into pieces before translating them individually and stitching them back together for the final product. This system is fairly fast, but it is also inaccurate and results in translations that may be technically correct by phrase but are not accurate overall.

Meanwhile, neural machine translation considers the entire content of a message together, which is more resource intensive but usually results in a more fluent translation. In today’s announcement, Facebook also said that neural machine translation is better at handling unknown or misspelled words, since it is able to look at context clues to figure out the word’s intended meaning.

While both Facebook and Google are moving to neural networks, the two companies are taking different approaches to the transition. Google’s system uses recurrent neural networks, or RNNs, which tend to be better at reading data in sequences, for example reading words left to right. Meanwhile, Facebook has been working to implement convolutional neural networks, or CNNs, which are normally used for visual tasks like image recognition.

RNNs have traditionally been better at translation, but Facebook believes CNNs can achieve the same accuracy at significantly greater speeds. In May, the company said its CNN-based translation system was nine times faster than existing networks, and a team of Facebook researchers explained that CNNs are a better fit for the latest machine learning hardware.

“RNNs operate in a strict left-to-right or right-to-left order, one word at a time,” the Facebook team said at the time. “This is a less natural fit to the highly parallel GPU hardware that powers modern machine learning.”

While Facebook seems to believe in the future of CNNs for translation, most of the company’s system still relies on RNNs. At the moment, Facebook has only used CNNs for English-to-French and English-to-German translation, but the research team said that CNNs are “an exciting new development path, and we will continue our work to utilize CNNs for more translation systems.”

Photo: Facebook