교사 발화 재구성을 위한 TMSSR 프레임워크 기반 Lesson Play AI 피드백 시스템 설계 및 적용

초록

This study designed and implemented an AI-based feedback system for Lesson Play, grounded in the Teacher Moves for Supporting Student Reasoning (TMSSR) framework by Ellis et al. (2019), aiming to support teachers in producing utterances that promote students' mathematical reasoning. The system's reliability and the effects of the use of the system were examined. The web-based platform constructed the TMSSR framework as a knowledge base (vector store) and referenced it through the File Search function of the OpenAI Assistants API, thereby classifying teacher utterances at the line level, generating feedback, and supporting a cyclical writing–feedback–revision process. The system was applied over the course of one semester with 34 preservice and inservice teachers, and analyses were conducted across three dimensions. First, the system's classification reliability showed very high agreement with human experts, yielding Cohen's Kappa values of .922 for category classification and .892 for potential-level classification. Second, a comparison of preservice teachers' Lesson Plays between the beginning and end of the semester revealed that the proportion of high-potential moves increased from 61.1% to 75.8%, with notable increases in the responding and extending categories. Individual-level analyses likewise confirmed consistent improvement across all participants. Third, a single-case qualitative analysis demonstrated that iterative AI feedback can extend beyond category transitions of individual utterances to reconfigure teachers' perspectives on the structure of classroom discourse and the scope of mathematical inquiry. This study proposes design principles for AI feedback systems that are grounded in theoretical frameworks, integrated with teacher educators' practice, and supportive of learners' cyclical revision experiences.

키워드