Automating Meso-Level Classroom Analysis: An LLM-Based Approach Using the PTA Coding Framework

• In Event: Virtual Poster Session
  In Virtual iPresentation Only: Division C - Learning and Instruction / Division C - Section 3b: Technology-Based Environments Virtual Poster Session

Abstract

Automatically identifying meso-level instructional activities from classroom discourse is a significant challenge for advancing educational research. This study addresses this gap by constructing a novel Procedure-Task-Action (PTA) coding framework to enable LLM-based analysis of teaching. We developed an effective training strategy, integrating prompt engineering with adapter fine-tuning on a 7B-parameter model. Using primary school Chinese lesson transcripts, we found that curating a training set of well-structured lessons was critical for model convergence and performance. The resulting specialized model achieved 85.4% accuracy in recognizing instructional activities. This validated methodology offers a scalable, cost-effective approach, helping to construct a new vision for the future of automated, equity-focused instructional analysis.

Authors

• Xiangfang Song, East China Normal University

• Jingliang Kong, East China Normal University