The document presents a framework for using the semantics of student textbook highlights to predict comprehension and knowledge retention. It uses semantic embeddings to encode highlighted sentences, compares them to questions, and uses the match scores in a model. It finds that augmenting a baseline model with highlighting features improves predictions of question accuracy, especially for held-out students. A semantic encoding of highlights performed better than a positional encoding. The approach works well across different levels of conceptual difficulty as defined by Bloom's taxonomy.