Facebook announced that its translation systems are now fully powered by neural networks – computing systems that are loosely modeled after the neurons in the human brain, moving away from the traditional phrase translation model.
Prior to using neural networks, Facebook used the phrase-based machine translation models. In these old translation models, sentences were broken down into individual words or phrases. One of the drawbacks of these old translation models, according to Facebook, is that they can only take into account several words at a time, leading to poor translations between languages with different word orderings.
The neural networks, on the other hand, consider the entire context, typos, slang and abbreviations of the source sentence, creating more fluent and accurate translations. For instance, Facebook’s English-to-Spanish neural network can translate “tmrw” – short for tomorrow – into “mañana.”
Below is an example of the translation produced by the Turkish-to-English phrase-based machine translation model:
Below is the translation produced by the Turkish-to-English neural network:
According to Facebook, the new neural networks now powered all of the tech giant’s backend translation systems, which produce 4.5 billion translations each day.
“Completing the transition from phrase-based to neural machine translation is a milestone on our path to providing Facebook experiences to everyone in their preferred language,” Facebook said in a statement. “We will continue to push the boundaries of neural machine translation technology, with the aim of providing humanlike translations to everyone on Facebook.”