Masked IRL: LLM-guided reward disambiguation from demonstrations and language