Reward contains positive reward and negative reward.

Positive/negative reward is the positive/negative data which is collected by reinforcement learning procedure.

For text generation task, positive data means we add this data to training dataset, negative data means we remove this data from training dataset.