Think I’m not gettting exactly what you’re trying to do, but I can at least say what shape input and output you would get from your network.
Input would be (batch_size, length). Make sure your input is a single numpy array with that shape (not a list of arrays).
Conv output would be (batch_size, length-2)
Final output will be (batch_size, vocab_size). Make sure your labels match this shape for each batch.
You may want soft max instead of sigmoid on that last layer if you want the total output to sum to 1 (otherwise can add up to anything from 0 to vocab_size).