Verbalized Representation Learning for Interpretable Few-Shot Generalization