Language translation has typically been done by recurrent neural networks (RNN), which process language one word at a time in a linear order, either right-to-left or left-to-right, depending on the language. This CNN-based architecture pays attention to words farther along in a sentence to help understand the meaning from context farther along the string of words, much like humans do. While the older RNN method has been typically fine for end users in regards to speed and accuracy, there’s a functional limit to the tech, one which the parallel processing model of CNNs can address. This is the first time a CNN has outperformed the more traditional RNN techniques. Facebook hopes to use the new methodology to scale its translation efforts to cover “more of the world’s 6,500 languages.”
Now that the popular social network has chosen CNN translation processing architecture, it will be interesting to see what comes next. Fast, accurate language translation might make our world feel a little smaller and more connected without the barrier of language in the way. The impact of this new technology will likely be felt globally, especially across the many Facebook-owned apps that help connect us all, like Messenger, WhatsApp and Instagram.