A Hybrid Neural Network BERT-Cap Based on Pre-Trained Language Model and Capsule Network for User Intent Classification

Algorithm 1

The dynamic routing algorithm.

	Input: all lower-level capsule U = [u1, u2, …, um]
	Output: the result of the squash activation function a
(1)	procedure Routing (ui, r)
(2)	for all lower-level capsules ui: bi = 0
(3)	for r iterations do:
(4)	for all lower-level capsules ui: ci = softmax (bi)
(5)	The higher-level capsule: s =
(6)	The higher-level capsule: a = squash (s)
(7)	for all lower-level capsules ui: bi = aui
(8)	return a