Just for fun, I've provided a few attempted translations here. These all come from the front of the training set, which means I'm making predictions on the training dataset, so these translations likely make the model look better than it actually is.

Our first translation gives you a feel for what we're expecting, and the network does a good job:

Input sentence: Help!

Correct translation: À l'aide!

Decoded sentence: À l'aide!

The translation following is more interesting because the learned phrase isn't connected to any of the training phrases. The phrase Vas-tu immédiatement! translates to something like You go immediately which is very similar and perhaps even correct: 

Input sentence: Go on.
Correct translation: Poursuis.
Decoded sentence: Vas-tu immédiatement!

-
Input sentence: Go on.
Correct translation: Continuez.
Decoded sentence: Vas-tu immédiatement!

-
Input sentence: Go on.
Correct translation: Poursuivez.
Decoded sentence: Vas-tu immédiatement!

-

Of course, there are many ways to say the same thing, which makes things more difficult for the network:

Input sentence: Come on!
Correct translation: Allez !
Decoded sentence: Allez!

-
Input sentence: Come on.
Correct translation: Allez!
Decoded sentence: Allez!

-
Input sentence: Come on.
Correct translation: Viens!
Decoded sentence: Allez!

-
Input sentence: Come on.
Correct translation: Venez!
Decoded sentence: Allez!