Post-problem reflective tutorial dialogues between human tutors and students are examined to predict when the tutor changed the level of abstraction from the student's preceding turn (i.e., used more general terms or more specific terms); such changes correlate with learning. Prior work examined lexical changes in abstraction. In this work, we consider semantic changes. Since we are interested in developing a fully-automatic computer-based tutor, we use only automatically-extractable features (e.g., percent of domain words in student turn) or features available in a tutoring system (e.g., correctness). We find patterns that predict tutor changes in abstraction better than a majority class baseline. Generalisation is best-predicted using student and reflection features. Specification is best-predicted using student and problem features.