Decomposing language models into understandable components