Bridging the Digital Divide in Formative Assessment: A Browser-based Speech Recognition and Large Language Model Approach for Real-time Classroom Analysis

This dataset contains the underlying data and analysis code for the study titled "Generative AI feedback loops induce a Digital Hawthorne Effect and enhance student cognitive engagement in low-resource classrooms."Study OverviewThis study proposes a browser-based, privacy-preserving classroom analysis system that utilizes the Web Speech API and the DeepSeek Large Language Model (LLM) to transcribe teacher speech and quantify student engagement in real-time. The empirical research was conducted over 9 weeks involving 49 students in a secondary school setting.Dataset ContentsThe repository includes the following files:1. Classroom_Interaction_Logs.csv- De-identified logs of teacher-student interactions.- Includes interaction frequency and feedback types (e.g., "Strong Affirmation", "Weak Affirmation").- Variables: `Student_ID`, `Week`, `Interaction_Count`, `Feedback_Type`.2. Student_Performance_Data.csv- Academic performance metrics for both Experimental and Control groups.- Variables: `Student_ID`, `Group`, `Midterm_Score`, `Final_Score`, `Progress_Score`, `Interaction_Index`.3. Analysis_Scripts.py- Python scripts used to perform statistical analyses (Pearson correlation, T-tests) and generate the figures presented in the paper.- Dependencies: `pandas`, `matplotlib`, `scipy`, `numpy`.4. Roster_Template.json- 一个结构模板，展示了用于身份消歧义的“名册注入”机制。- 注：不含真实姓名;纯粹用于算法演示。方法论与伦理- 去标识化：所有参与者都被分配随机的字母数字代码（例如S001），以保护隐私。不包含个人身份信息（PII）。- 隐私设计：音频数据在浏览器的易失内存中本地处理后立即丢弃。该数据集中未存储或包含任何原始音频录音。使用说明研究人员可以使用“Analysis_Scripts.py”功能，利用提供的CSV文件重现统计结果（手稿中的图3-6）。