Text this: Adaptive representations for reinforcement learning