Machine translation is the automatic translation from one language to another. ## statistical machine translation Statistical machine translation was the standard approach for machine translation from the early 1990s to the early 2010s, when it was replaced by neural machine translation. These consisted of a translation model to map the words in the source language to words in the target language and an alignment model which aligns the translated sequence. ## neural machine translation Neural machine translation uses a single neural network to translate from source to target with an encoder-decoder architecture. The encoder encodes the source sentence into a vector or matrix. The decoder generates the target sequence. One early model by Sutskever et al. (2014) called `seq2seq` used sentence embeddings as the initialization of the RNN. LSTM with four layers were used with the input sequence reversed. In more advanced neural machine translation, [[attention]] is used to find alignment between the two sequences, an idea missing from `seq2seq`, first proposed by Bahdanau et al. (2015). ## context vector