Download Full PDF

胡月宝

贺禹婷

新加坡南洋理工大学国立教育学院，新加坡

 

摘要

本文针对人工智能教育中教师端工具不足的问题，提出一套智能课堂话语分析系统（ AI-based Classroom Discourse Analysis System）。研究利用新加坡 12 名教师 24 节课逐字稿，结合自动语音转写与 OpenAI 的 GPT-4 模型话语功能分类，对比人工与 AI 在处理效率与分类表现上的差异。结果显示，智能课堂话语分析系统在转写与初步分类阶段将整体耗时压缩至纯人力流程的一小部分，但在话语功能分类上与人工仅呈中等偏低的一致水平，并出现稳定的“双峰”偏差 ：对课堂管理、社交互动等形式特征突出的话语高估，对引入、知识讲解与总结等依赖语篇结构的话语低估。研究据此将 AI 定位为预标注与趋势分析工具，由教师针对高风险类别进行复核，并指出未来需在模型中引入教学阶段与语篇结构等特征，以提升课堂话语分析的有效性。

 

关键词

智能课堂话语分析，人工智能教育，人机协同，教师专业发展

 

Classroom Discourse Analysis in Human–AI Collaboration: Evidence from a Pilot Study in Singapore Primary Chinese Classrooms

 

Guat Poh Aw

Yuting He

Nanyang Technological University, Singapore

 

Abstract

Addressing the underdevelopment of teacher-oriented tools in Artificial Intelligence in Education (AIED), this study proposes an AI-based Classroom Discourse Analysis System (AI-CDAS) designed to support instructional diagnosis and classroom interaction analysis. The system was developed and evaluated using verbatim transcripts from 24 lessons taught by 12 primary Chinese language teachers in Singapore. Automated speech recognition was employed for transcription, and OpenAI’s GPT-4 model was applied to classify discourse functions. System performance was examined by comparing AI-assisted and fully human workflows in terms of processing efficiency and classification consistency.

 

Results indicate that the AI-CDAS substantially reduced overall processing time at the transcription and preliminary coding stages, compressing the workflow to a fraction of the time required by manual analysis. However, agreement between AI and human coders on discourse function classification remained at a moderate-to-low level. A stable bimodal bias was observed: the model tended to overestimate utterances with salient formal features (e.g., classroom management and social interaction) while underestimating discourse functions dependent on broader pedagogical sequencing and discourse structure (e.g., lesson introduction, knowledge explanation, and summarization).

 

Based on these findings, the study positions AI not as a substitute for human judgment but as a pre-annotation and trend-detection tool, with teachers responsible for reviewing high-risk categories. The findings further suggest that future iterations should incorporate instructional phase indicators and discourse-structural features to enhance classification validity in classroom discourse analysis.

 

Keywords

Classroom discourse analysis, AI in education, human–AI collaboration, teacher professional development