Abstract

The way that content-related keywords co-occur along a lesson seems to play an important role in concept understanding and, therefore, in students' performance. Thus, network-like structures have been used to represent and summarize conceptual knowledge, particularly in science areas. Previous work has automated the process of producing concept networks, computed different properties of these networks, and studied the correlation of these properties with students' achievement. This work presents an automated analysis of teachers' concept graphs, the distribution of relevance amongst content-related keywords and how this affects students' achievement. Particularly, we automatically extracted concept networks from transcriptions of 25 physics classes with 327 students and compute three centrality measures (CMs): PageRank, Diffusion centrality, and Katz centrality. Next, we study the relation between CMs and students' performance using multilevel analysis. Results show that PageRank and Katz centrality significantly explain around 75% of the variance between different classes. Furthermore, the overall explained variance increased from 16% to 22% when including keyword centralities of teacher's discourse as class-level variables. This paper shows a useful, low-cost tool for teachers to analyze and learn about how they orchestrate content-related keywords along with their speech.