Enhancing Large Language Model Efficiencyvia Symbolic Compression: A Formal Approach Towards Interpretability