Colloquium: Dr. Zhongzhou Chen|University of Central Florida
In-Person PHYS 401
Wednesday, May 8, 2024 · 11 AM - 12 PM
TITLE: "Towards better measurement of student learning outcomes using Data mining and Large Language Models”
ABSTRACT: How much physics did our students actually learn from our physics courses? This seemingly easy question has been one of the major focuses of my research over the past several years, and in this talk I will share my latest efforts in developing assessments that not only are more flexible, accessible and equitable, but also more accurate and reliable. First, I will introduce mastery-based online learning modules, a design that aims to administer rapid pre- and post-tests on student learning at a much higher frequency than conventional assessments. Analysis of data from students interacting with online learning modules revealed rich and diverse patterns of student learning behavior, but also reveals a validity threat to this assessment design likely caused by online answer sharing websites such as Chegg. Next, I will introduce my attempt to overcome this challenge by utilizing Large Language Models to create large isomorphic problem banks, and transform conventional assessment with those banks. Finally, I will share latest developments on how Large Language Models can have the potential to serve as a grading assistance for students' written response to physics questions to enhance the accuracy and reliability of assessments while reducing instructor workload at the same time.
ABSTRACT: How much physics did our students actually learn from our physics courses? This seemingly easy question has been one of the major focuses of my research over the past several years, and in this talk I will share my latest efforts in developing assessments that not only are more flexible, accessible and equitable, but also more accurate and reliable. First, I will introduce mastery-based online learning modules, a design that aims to administer rapid pre- and post-tests on student learning at a much higher frequency than conventional assessments. Analysis of data from students interacting with online learning modules revealed rich and diverse patterns of student learning behavior, but also reveals a validity threat to this assessment design likely caused by online answer sharing websites such as Chegg. Next, I will introduce my attempt to overcome this challenge by utilizing Large Language Models to create large isomorphic problem banks, and transform conventional assessment with those banks. Finally, I will share latest developments on how Large Language Models can have the potential to serve as a grading assistance for students' written response to physics questions to enhance the accuracy and reliability of assessments while reducing instructor workload at the same time.