The creation of crowd-sourced content in learning systems is a powerful method for adapting learning systems to the needs of a range of teachers in a range of domains, but the quality of this content can vary. This study explores linguistic differences in teacher-created problem content in ASSISTments using a combination of discovery with models and correlation mining. Specifically, we find correlations between semantic features of mathematics problems and indicators of learning and engagement, suggesting promising areas for future work on problem design. We also discuss limitations of semantic tagging tools within mathematics domains and ways of addressing these limitations.