Facebook Aims to Break Down Language Barrier with New Translator AI

WebProNewsSocial Media

Share this Post

True to its goal of connecting people all over the globe, Facebook has been trying to crack the language barrier for its billions of users. On Tuesday, its AI research team revealed a method that will drastically improve the way its 1.8 billion-strong members understand each other on the social media platform.

The Facebook Artificial Intelligence Research (FAIR) team reported a breakthrough in the use of novel convolutional neural network (CNN) as opposed to the recurrent neural networks (RNN) being used by cross platforms to translate a particular language.

Research results found that the new language model is “nine times the speed” of current RNN models. The FAIR team admitted that they have just scratched the surface, as it potentially can be sped up some more using other distillation methods.

For the longest time, the RNN has outperformed the CNN in terms of language interpretation, but it is sorely limited in the way it processes information. As the Facebook engineers explained it, the language works “in a strict left-to-right or right-to-left order, one word at a time.” This means that the program will have to wait for one word to be translated before moving on to the next.

The problem with the RNN is that the GPU, the default hardware that powers modern machine AI, is highly parallel. This is more compatible with the CNN, which interprets data in a hierarchical manner and can make translations simultaneously by factoring in the correlation between each word in a sentence.

Christopher Manning, professor of Computer Science and Linguistics at Stanford University, described Facebook's announcement as an “impressive achievement.” The professor, who works with machine learning, said the breakthrough allows the social media company to speed up existing translation models.

“You can have parallel computation on different parts of a sentence. You don’t have to push things along word by word,” he explained.

Indeed, the research concluded that CNN language is more efficient on account of its multi-hop function, which means it goes back to the original sentence time and again to translate multiple words. It also potentially has the capability to focus on two separate facts at the same time and interpret them within a larger context, which will help break down complex sentences.

Facebook is going to share the source code on GitHub so other engineers can customize the new language translator for further efficiency and accuracy. More importantly, researchers can use the code to break down the language barrier for multiple platforms, not just on Facebook or its affiliated sites.  

WebProNews
WebProNews | Breaking eBusiness News Your source for investigative ebusiness reporting and breaking news.