What's In My Human Feedback? Learning Interpretable Descriptions of Preference Data