Could the encoder/decoder cells of NMT be "mix and matched"? The goal of the encoder LSTM is to produce a meaning vector that totally encompasses the content of the source sentence.
The goal of the decoder LSTM is to use the meaning vector is to produce an accurate representation of the meaning vector.
Therefore, couldn't the encoder and decoder LSTMs be used seperately once disconnected? 
Train one NMT model on English -> French.
Train a second NMT model on Spanish -> Italian.
Could the encoder cell for the former model be used with the decoder cell from the second cell? If not (because they might develop in tandem), could training with arbitrary pairings of encoder/decoder overcome this?
 A: Translation whose model scales linearly with the number of languages is an open research area. A complication is that modern translation systems use attention heads, which couples the language representations more tightly perhaps. Bengio's team has created a system whose parameters scale linearly with the number of languages:
https://arxiv.org/pdf/1601.01073.pdf
Multi-way, multi-lingual neural machine translation with a shared attention system
Firat, Cho, Bengio, 2016
Abstract
"We  propose  multi-way,  multilingual  neural
machine translation.  The proposed approach
enables  a  single  neural  translation  model  to
translate between multiple languages,  with a
number  of  parameters  that  grows  only  lin-
early  with  the  number  of  languages.
This
is  made  possible  by  having  a  single  attention mechanism that is shared across all language  pairs.    We  train  the  proposed  multiway, multilingual model on ten language pairs
from  WMT’15  simultaneously  and  observe
clear performance improvements over models
trained on only one language pair.  In particular, we observe that the proposed model significantly improves the translation quality of
low-resource language pairs."
