I am currently studying Deep Learning based Machine Translation systems but I'm not sure in my understanding the logic of the process. I understand that the source and destination language translation sentence pairs must be represented as word2vec vectors, but why is it necessary to apply two (encoder-decoder) Recurrent Neural Networks? My first idea would be applying only one RNN, where the input is the source language examples (in the form of word2vec vectors) and the output is simply a word2vec sequence of the destination language. Why is it necessary to use another RNN?
My additinal question is if this system is flexible enough to cope with synonimes, word order variations and other disambiguities? Is it capable of approximate the correct meaning of a new to-be-translated source language sentence?
And last but not least: how could one evaluate such a model where many translations can be correct at the same time?