The social network’s efforts with machine learning and artificial intelligence have advanced enough to the point that it can completely take over Facebook’s language translation. Text in posts and comments are now automatically translated so that people all around the world can read them. Given that Facebook now has 2 billion active users, this is an impressive feat.
From SMT to Neural Networks
In order to pull this off on a social network, the machine learning needs to take into account context, slang, typos, abbreviations, and intent all at once. Previously Facebook was using phrase-based machine translation, which is a type of statistical machine translation (SMT).
With phrase-based translation, the goal is to translate entire sequences of words that have different lengths as an effort improve on word-based translation. But now, Facebook has moved on to neural networks with sequence-to-sequence LSTM (long short-term memory).
This is more accurate than phrase-based translation because it doesn’t have to break up sentences into blocks of phrases. The neural network can work with the whole sentence at once and understand the context. The new system increased the accuracy by an average of 11 percent.
Since the neural network can work with abbreviations too, it can successfully translate the English-to-Spanish abbreviation of “tmrw” (tomorrow) into “mañana.” People can still rate translations too, which leads to greater accuracy over time as the system learns which translations were better than others.