Explaining Categorical Feature Interactions Using Graph Covariance and LLMs