Mitigating Word Bias in Zero-shot Prompt-based Classifiers