Kyle Gorman

What would you do about this? My thought was that it's hard to imagine (except for debugging I guess) why you'd want to decode anything but the target.

It occurs to me that clashes can't happen anyways because of the special treatment we give to features. Does that help at all?

Yeah, you could just add features to the source vocabulary. (Actually I'd do it regardless of whether concatenation is happening since it's harmless and it's annoying to pass information about...

@Adamits I believe has long been addressed. Can you confirm?

A test involving "decoding source" (since you thought this was useful as an operation) would be welcome.

Sure, I have been wondering that too. Is there any way to absolutely forbid them from being copied into the target sequence though? (And wouldn't that give rise to a...

"To conatenate or not to concatenate": great SIGMORPHON short paper idea.

> > My proposal is that we should, in attentive LSTM and transformer, have separate (LSTM and transformer, respectively) encoders for features and these encoders should then be lengthwise concatenated....

> 1. It looks like this is most cleanly managed by having a `features.py` module that just maintains encoders to be imported by individual architectures. Hi Travis. This sounds good...

I don't see how this design would work with pointer-generators, so that needs to be clarified. (You may want to chat with @Adamits about this later---there's a subtlety we discussed...