There is a problematically small percentage of underrepresented populations working as STEM professionals. In order to address this issue, educators need to find techniques to better help underrepresented populations learn. In this project, Neil Heffernan of WPI, Dongwon Lee of Penn State, and Xintao Wu of University of Arkansas will use ASSISTments to learn 1) which existing educational YouTube videos cause the most learning and 2) if viewing educational videos enhances learning more than not viewing educational videos.
We will run experiments that will assess the helpfulness of a video by a student's next problem correctness. This experimentation will be applied across content for three Open Educational Resource (OER) textbooks that are openly licensed and free to schools. In order to accomplish this task, we will first search YouTube for videos that address the skill addressed by an OER problem. Second, we will extract features of these videos such as the complexity of language, speaking rate and ethnicity of the speaker. We will then assess the video quality using MTURKS and content experts. In the final year of the project, we will run a week-long study to compare the effects of helpful videos to a business-as-usual condition where students can "just Google it" to find homework assistance.
This work was made possible by NSF grants 1940236, 1940076, and 1940093.
Precision Learning: Personalizing Video Viewing
We will use AI techniques to find videos that enhance student learning on average and for specific underrepresented populations. We will randomize students to view Videos A, B and C to determine the helpfulness of each video for student learning. We will study this data both on average and for specific demographics. It is possible that Video A will be most effective for Hispanic males, while Video C may be most effective for Asian females. We will investigate which features of videos are effective for which students using TETRAD causal modeling tools.
The proposed research will advance theoretical understanding of fundamental issues related to personalized learning and enable data-driven experimentation of learning theories. The causal modeling we will do will allow us to learn the video features that are correlated with learning effectiveness, thus assisting learning scientists. We hope to learn which video features can help close achievement gaps, allowing teacher training programs to address these concerns and better equip teachers to address achievement gaps.