Learning Invariable Semantical Representation from Language for Extensible Policy Generalization