We present initial results from a large-scale text mining analysis of student written responses to UM course evaluation questions, based on a dataset comprising millions of comments for thousands of courses over 10 semesters. Our goals are to (1) build on existing analysis of numeric ratings; (2) understand more about which specific positive and negative factors in the learning environment may distinguish courses at similar and different numeric rating levels across different evaluation questions; (3) look at how these positive or negative factors distinguish courses across other facets like sections within a larger course, school/college/department, or instructor demographic. Project contributors: Kevyn Collins-Thompson, Phyllis Ford, Adam Levick, Florence Lee, Mika LaVaque-Manty, Michael Spiegel, Lifan Yang. Supported by CRLT Winter 2014 Learning Analytics Fellows Program.