The implementation of BiDAF is not exactly same as the one in the original paper
chunchiehy opened this issue · comments
For the calculation of similarity matrix,
sim_matrix = tf.matmul(passage_encodes, question_encodes, transpose_b=True)
This is a simple dot product. But the original paper choose
where h
is the passage representation and u
is the question representation.
This maybe not a big problem, but it may still cause some difference.
Thanks for your comment! And we have to clarify that our implementation is not strictly same as the original one. And for the similarity matrix, we indeed select a simplified implementation. It may cause some differences, but we do not investigate significant performance difference for now.