Dynamic Bayesian Network Models for Peer Tutoring Interactions

The ability to automatically distinguish between effective and ineffective patterns in collaborative learning sessions opens doors to improved opportunity for learning in pairs or groups even when a teacher might not be available to facilitate. In this chapter, data from one-time computer-based peer tutoring sessions are modeled using hidden Markov models (HMMs) in two ways. The first model uses an input–output HMM to compare the assistance value of different tutor inputs in helping the tutee correct a mistaken step in solution. This model uses only automatically generated codes based on context and cognitive content of the tutor chat. The second model predicts tutee normalized gains from pre- to posttest in the experimental condition. Both cognitive and affective labels to tutor chats (human coded) were included as well as tutee (in)correctness, undos, and chats back to the tutor. Performance of the HMM is favorable compared to a “static” logistic regression model using aggregated totals of the same observables. Some of the hidden states are readily interpretable, though deeper comparison between high- and low-gain groups is part of ongoing work.