How to prepare the source.att file about cnn-re-tf HOT 1 CLOSED

Hey,

I am not quite sure whether I could just added these other attributes as left, right file, correct me if I am wrong:

Humm, I don't think my context-wise code suites your case. In my code, all contexts (left, right and middle) look up word embeddings, but in your case, you combine text input ("The fox chase a bunny") with numerical features ("23 , 24000") and categorical features ("high school"). And of course numerical and categorical features don't need word embeddings. So you can't reuse my code, you have to re-implement the layer for your non-textual features, I guess.

BUT how abount source.att? How to decide the values between 0.0 and 1.0?

Yeah, this is a problem, how you get the attention. I derived this from some manual annotation (see: data/gold_annotation.tsv, conversion into 0-1 value: score_reliability function).

But actually, attention values should be jointly learned during training, I mean, should not be given in advance. My attention implementation is fixed, just adds extra info to the loss. So, there are two options: (1) use some fixed manual annotation (like my code, which is not so good...) or (2) implement an attention layer to jointly learn the values (I don't have code for that, but it is desired in the actual sense of "attention")

Sorry, I can't help you for this topics...

from cnn-re-tf.