Utterance-Level Modeling of Indicators of Engaging Classroom Discourse

We examine the ability of supervised text classification models to identify several discourse properties from teachers' speech with an eye for providing teachers with meaningful automated feedback about the quality of their classroom discourse. We collected audio recordings from 28 teachers from 10 schools in 164 authentic classroom sessions, which we then automatically transcribed into text utterances and then manually coded to identify whether: (1) the utterance contained a question (as opposed to a statement), (2) the question or statement was Instructional vs. Non-Instructional, and (3) the question or statement was Content-Specific. We experimented with Random Forest classifiers and engineered (linguistic, acoustic-prosodic, and contextual) features vs. open-vocabulary n-grams as features to discriminate these discourse variables at the utterance level in a teacher-independent fashion. We achieved AUC scores ranging from 0.71 to 0.77 using open-vocabulary language modeling, which were well above chance (AUC = 0.5), an important step towards our predominant goal of constructing of an automated feedback system for teacher reflection and learning. [For the full proceedings, see ED599096.]