De-identifying Student Personally Identifying Information with GPT-4

Academic Article

Academic article assessing GPT-4’s performance in de-identifying personally identifying information from student-generated educational data.

Visit Resource
This link will take you to an external website.
Purpose/Abstract

Education is increasingly taking place in technology-mediated learning environments, making it easier to collect student-generated data such as comments in discussion forums and chats.

Although this data can be valuable to researchers, it often contains sensitive information such as names, locations, social media links, and other personally identifying information that must be carefully redacted before being used for research.

Historically, personally identifying information has been redacted by humans. More recently, researchers have also explored regular expressions and supervised machine-learning methods.

This paper assesses GPT-4’s performance in de-identifying data from discussion forums in nine Massive Open Online Courses.

The results show an average recall of 0.958 for identifying personally identifying information that needs to be redacted, suggesting that GPT-4 is an appropriate tool for this purpose. The tool was also successful at identifying cases missed by humans during redaction.

The findings indicate that GPT-4 can increase the efficiency and enhance the quality of the redaction process. However, precision was considerably lower at 0.526, with the tool over-redacting names and locations that did not represent personally identifying information, showing a need for further improvement.

Citation
Singhal, S., Zambrano, A. F., Pankiewicz, M., Liu, X., Porter, C., & Baker, R. S. (2024). De-identifying student personally identifying information with GPT-4. In B. Paaßen & C. D. Epp (Eds.), Proceedings of the 17th International Conference on Educational Data Mining (pp. 559–565). International Educational Data Mining Society.

Areas researched: Data Use, AI

Previous
Previous

De-identifying Student Personally Identifying Information in Discussion Forum Posts With Large Language Models

Next
Next

Evaluating the Impact of a Mathematics Mastery Learning Platform on Student Achievement: A Large-Scale Longitudinal Analysis