Predicting student outcomes from unstructured data

We investigated the validity of applying topic modeling to unstructured student text data from online class discussion forums to predict students’ final grades. Using only student discussion data from introductory courses in biology and economics, both probabilistic latent semantic analysis (pLSA) and hierarchical latent Dirichlet allocation (hLDA) produced significantly better than chance predictions which improved with additional data collected over the duration of the course. Predictions were more accurate from hLDA than from pLSA, suggesting the feasibility and value of deriving conceptual hierarchies relevant to actual student data. Results indicate that topic modeling of studentgenerated text may provide a useful source of formative assessment to support learning and instruction.