Parsing the Language of Expression: Enhancing Symbolic Regression with Domain-Aware Symbolic Priors