Closed matt-gardner closed 7 years ago
Ah i've just looked at your next PR and it seems like some of the issues here might have been fixed there. Sorry if that's the case, i'll review that tomorrow.
Yeah, I was thinking this could be a self-contained change, but after finishing the next one and seeing your comments, I don't think I succeeded with that. I'll close this one in favor of the next one.
This lets you have two separate embedding matrices, one for words and one for characters (or any other token namespace that you want). This will make it a lot easier to have an
Encoder
that has both a word vector and a CNN character encoder.