Mentioned earlier seq2seq model, we start from here.
Seq2seq model he used the encoder-decoder architecture, this time we should point to map the paper it!
In the past, the language model in practice, will be put into a word RNN years, it will immediately spit out a RNN corresponding word out.
Like into a English word, it will spit out a corresponding French word, then one of these predictions to bring the next level.
While doing very intuitive, but he did not complete translation of a sentence.
Grammar of the language is different, it is difficult to make the words eleven pairs of enantiomers large column 26 seq2seq model translation.
This seq2seq model takes a different approach, a set of LSTM as encoder, responsible for the sentence to be translated is converted into a vector of fixed length, then this vector to another LSTM converted into target sentence, LSTM behind this is the role of the decoder.
Under such a model framework will be split into two parts encoder with the decoder, so that each part can accept two or produce sentences of different lengths, and get a good score.
In practice, the two LSTM will indeed pick up, so nothing to talk about details.
But this model does solve the problems of different length of sentences to sentences of different lengths.