Leveraging Class Abstraction for Commonsense Reinforcement Learning via Residual Policy Gradient Methods.