How to customize a XGBoost objective function for ordinal classification problem?

Question

I am training a model to do the classification on an ordinal response variable with 10 levels.

I have studied a paper called "Squared Earth Mover’s Distance-based Loss for Training Deep Neural Networks" (https://arxiv.org/pdf/1611.05916.pdf) and I would like to customize the XGBoost objective function to work on the "EMD".

However, I have no idea how to write that objective function. Instead of using multi:softmax or multi:softprob, can anyone help me to write a customized objective function catering the EMD?