Text this: Representation discovery in non-parametric reinforcement learning