Let the Rule Speak: Enhancing In-context Learning Debiasing with Interpretability