Semantic Anomaly Detection with Large Language Models