Switching to Discriminative Image Captioning by Relieving a Bottleneck of Reinforcement Learning