Decades of research has demonstrated the importance of teachers on both student short-term learning and long-term outcomes including college enrollment, earnings, and health (Chetty, Friedman, Rockoff 2014; Rivkin, Hanushek, and Kain 2005). However, researchers know far less about what teaching strategies constitute high-quality teaching, impeding effective school-level decision-making and education policy making more generally. A key challenge of this knowledge gap lies in the difficulty of measuring the complex classroom interaction process. Classroom observations by expert raters following structured observational protocols have been proven to be costly and subject to raters’ bias (Bell et al. 2014).

The recent development of computational linguistic analysis provides an unprecedented opportunity of measuring aspects of teaching that are difficult for human raters to capture with higher precision and lower cost. For example, Kelly et al. (2018) used both automatic speech recognition and machine learning to detect teachers’ use of authentic questions, an important dimension of classroom discourse. Relatedly, Wang, Miller, and Cortina (2013) used an automated speech recognition tool to precisely classify the interaction patterns between teachers and students and provide timely feedback to teachers that could help them monitor students’ active participation in classroom discussion. While both studies demonstrate the potential of computational techniques in measuring teaching practices in some ways, they focus on either a discrete measure of teaching or the mechanical turn-taking between teachers and students without factoring in the complex interactive process of classroom discourse. Neither of them corroborates the computer-generated measures with respect to student outcomes.

This paper tackles the challenge of measuring teaching by focusing on a complex classroom dialogic feature—uptake. Uptake happens in a classroom when a teacher builds on the previous student contribution, by revoicing it and using it to set the subsequent dialog direction (Collins 1982; Nystrand 1997; Herbel-Eisenmann et al. 2009). We build an unsupervised linguistic model of uptake and apply it to nearly 2,500 classroom videos collected for the National Center for Teacher Effectiveness study (2010-2011 through 2012-2013). Preliminary results suggest that our automated measure highly correlates with similar constructs captured by human raters under the Mathematical Quality of Instruction (Hill et al. 2008) and the Classroom Assessment Scoring System (La Paro, Hamre & Pianta, 2012). Our future analysis will further validate this measure using human annotations and student achievement such as test scores. In sum, this study demonstrates that the use of rich textual information and computational methods can help unpack quality teaching and inform critical education policy discussions more generally.