Abstract

Classroom observation methods are fundamental tools for improving the quality of education and students' academic achievement. However, they traditionally require participation of trained observers, making them expensive, prone to rater bias and time consuming. Hence, to address these challenges we present a cost-effective and non-intrusive method that automatically detects different teaching practices. In particular, we extracted acoustic features and transcriptions from teachers' talk recordings to train a multimodal learning model called Teacher Activity Recognizer from Transcriptions and Audio (TARTA), which detects three categories derived from the Classroom Observation Protocol for Undergraduate STEM (COPUS), namely Presenting, Administration, and Guiding. We found that by combining acoustic features and transcriptions, our model outperforms separate acoustic- and transcriptionbased models at the task of predicting teachers' activities along the lessons. In fact, TARTA can predict with high accuracy and discriminative power the presence of these teaching practices, achieving over 88% of accuracy and 92% AUC for all three categories. Our work presents improvements with respect to previous studies since (1) we focus on classifying what teachers do according to a validated protocol instead of discerning whether they or their students are speaking and (2) our model does not rely on expensive or third party equipment, making it easier to scale to large volumes of lessons. This approach represents a useful tool for stakeholders and researchers who intend to analyze teaching practices on a large scale, but also for teachers to receive effective and continuous feedback.