Correctable Landmark Discovery via Large Models for Vision-Language Navigation