Deep Visual Semantic Embedding with Text Data Augmentation and Word Embedding Initialization

Table 1

Notations used in this paper.

Notation	Description
l	The length of sentence
n	Number of times doing augmentation operation
p	The probability to remove every word in the sentence
	The percent of words to be changed in the sentence
	The triplet loss
	The loss of proposed model
	The similarity of anchor xa and positive input xp
xa	Anchor input
xp	Positive input
xn	Negative input
	The margin that let the negative pairs away from each other
	The similarity of image i and text t
i	Paired image
t	Paired text
	Not paired image
	Not paired text