Representation learning for computational sociopragmatics