It is very common for Facebook to perform around 4.5 billion automatic translations on daily basis. Previously, the models applied for translation were simple phrase-based models. Recently the platform has shifted its focus towards using more advanced models. According to Facebook, now these translations are all processed by using the neural network.
The purpose of the use of the neural network is to create seamless and highly accurate experiences for translation. For 2 billion users on Facebook this isn’t an easy task but with neural network makes it possible to take account of the content, abbreviations, typos, and the intent. The major difference between the previous system and the new one is the attention span. In the phrase based system, the translation was done word by word or by looking at the short sentences whereas the neural network studies the complete sentence at a time. This is done with the help of a specific kind of machine learning component known as LSTM or long short-term memory network.
The use of a neural network for translation has many advantages. By comparing the two phrases from a Turkish to English translation on Facebook reveals how considering the whole sentence for translation comes up with a more appropriate result.
According to the company when a word in a sentence does not have a direct corresponding translation in a particular language, a placeholder for the unknown word is generated by the neural network. the translation for that specific work is searched in a kind of an in-house dictionary that is built with Facebook’s training data. this helps to replace the unknown word. With the help of this method, the abbreviations such as “tmrw” can be translated into the intended meaning as tomorrow. The company states that many future developments paths are opened by the neural networks for creating a much-improved translation experience. The company is also focusing on the multilingual models for translating many language directions.