How reparametrization trick broke differentially-private text representation learning