Papers

We propose a framework combining acoustic, textual and temporal features to recover interaction structure in authentic classrooms. Across 412 lessons it segments teacher and student speech at 0.91 accuracy and reliably separates four questioning behaviours.

Knowledge tracing models that excel on public benchmarks degrade sharply on single-class data. Comparing five model families, we find degradation tracks the sparsity of practice sequences rather than sample size.

From more than two thousand hours of classroom audio we derive a nine-category taxonomy of teacher questioning and report how its distribution shifts across subjects and grade levels.