Towards Computationally Verifiable Semantic Grounding for Language Models