Reinforcement Learning for Few-Shot Text Generation Adaptation