Prompt-Based Bias Calibration for Better Zero/Few-Shot Learning of Language Models