Scholarlink

Artificial intelligence is increasingly used to provide automated feedback to teachers, offering scalable and personalized professional development on effective teaching practices. Machine learning models power automated feedback tools by detecting and measuring these practices. One such high-leverage practice, opportunities to respond (OTRs), is a teaching practice widely recognized for increasing student engagement and participation. In this study, we explored the use of GPT-4o to identify OTRs within classroom transcripts generated using AssemblyAI’s automatic speech recognition (ASR) model and a fine-tuned diarization model. We annotated transcripts to provide ground-truth labels and compared the performance of sentence-level and turn-level prediction approaches. Our fine-tuned GPT-4o model achieved an F1 score of 0.75 at the sentence level and 0.88 at the turn level. Lesson-level analyses revealed that both sentence- and turn-level aggregations correlated highly with ground-truth OTR counts, with turn-level predictions demonstrating fewer false positives. These findings demonstrate that GPT-4o effectively detects OTRs from noisy classroom recordings, enabling scalable, accurate automated feedback on a teaching practice to support teacher growth and student engagement.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

From Noisy Classroom Transcripts to Actionable Feedback: Fine-Tuning GPT-4o to Detect Teachers’ Opportunities to Respond

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Jessica Vitale,
  • Kemal Berk Kocabagli,
  • Shyamoli Sanghi,
  • Alyssa Van Camp,
  • Sarah Miller,
  • Berk Coker

摘要

Artificial intelligence is increasingly used to provide automated feedback to teachers, offering scalable and personalized professional development on effective teaching practices. Machine learning models power automated feedback tools by detecting and measuring these practices. One such high-leverage practice, opportunities to respond (OTRs), is a teaching practice widely recognized for increasing student engagement and participation. In this study, we explored the use of GPT-4o to identify OTRs within classroom transcripts generated using AssemblyAI’s automatic speech recognition (ASR) model and a fine-tuned diarization model. We annotated transcripts to provide ground-truth labels and compared the performance of sentence-level and turn-level prediction approaches. Our fine-tuned GPT-4o model achieved an F1 score of 0.75 at the sentence level and 0.88 at the turn level. Lesson-level analyses revealed that both sentence- and turn-level aggregations correlated highly with ground-truth OTR counts, with turn-level predictions demonstrating fewer false positives. These findings demonstrate that GPT-4o effectively detects OTRs from noisy classroom recordings, enabling scalable, accurate automated feedback on a teaching practice to support teacher growth and student engagement.