Facebook makes complete transition to neural machine translation
For some time, Google, Microsoft and Facebook have been making a transition to new machine translation engines- moving from phrase based statistical machine translation to neural machine translation. Facebook is the latest of these companies to have completed the process. With its new machine translation engine, it hopes to improve the accuracy of its translations and how natural they sound. Facebook has around 2 billion users, scattered over a wide range of countries and speaking many different languages, which generates a huge volume of status updates, posts, notes and other text every day. According to the social network, this amounts to more than 4.5 billion translations in over 2000 directions daily. Particular difficulties are caused by the fact that text written on social media tends to contain features such as slang, spelling errors and abbreviations.
Previously, Facebook had made use of phrase-based statistical machine translation engine. Such systems start off by breaking up text into words or small groups of words. Calculations of probability are then used to select a suitable translation from a large corpus of translated texts. Using this technology means that only a few words are used at a time, creating difficulties when there are significant differences between the word order of two different languages. Neural machine translation engines are better able to take the context of a sentence into account, alongside everything translated so far, which Facebook claims allows it to create “more accurate and fluent translations”. There are also improvements with how the system handles instances where there is no “direct” translation in the target vocabulary for a particular word.
Rather than just using an RNN (recurrent neural network) approach, Facebook’s engine also utilises a convolutional neural network (CNN) approach. Whereas RNNs read words in a left-to-right or -right-to-left order, taking one word at a time, CNNs can compute all element simultaneously. The Facebook Artificial Intelligence Research (FAIR) team has recorded several improvements in translation quality, both over phrase-based statistical translations and neural network systems that only use RNN. Facebook therefore believes there to be great potential for this type of system and its future development.