Refining Language Models with Compositional Explanations