The approval of new surveying-and-mapping geographic-information standards still depends largely on manual checks for duplication and novelty, which leads to slow and sometimes inconsistent decisions. We propose an intelligent evaluation framework powered by a large language model to streamline this process. The system combines domain-adaptive pre-training, contrastive learning for semantic similarity, and a dual-tower cross-attention network for novelty assessment, all integrated within a human-in-the-loop feedback loop. Experiments on real-world review data show that the domain-adapted encoder captures specialised terminology more effectively than generic baselines, while the downstream classifier delivers markedly higher precision and recall. Deployed with a FAISS index, the system responds in tens of milliseconds per query and shortens the overall review cycle from weeks to days, providing experts with ranked precedent standards, automated rejection alerts and clause-level explanations. The framework demonstrates the practical value of large language models for modernising standard-governance workflows and can be readily transferred to other regulatory domains.