Long short term memory () is a recurrent neural network () architecture that remembers values over arbitrary intervals. Originally invented to encode and predict time series it has been recently successfully used in Natural Language Processing, for encoding sequences of words, i.e. phrases.