Which is the training label in the "incremental step" ?

[andongBlue]

Hello,

When I read the paper and the code, a question always confuse me. Which is the training label in the "incremental step" ? In the paper, previous target hidden state is the input of MLP and relative distance is the output of MLP. However, what is the label for training this MLP? Dose this label need to be constructed manually or is it obtained in other way?