PIs: Shiting Lan, University of Massachusetts Amherst (Award Details)
Neil Heffernan, Worcester Polytechnic Institute (Award Details)
Ryan Baker, University of Pennsylvania (Award Details)
In recent years, there has been increasing effort to integrate modern artificial intelligence technologies into adaptive learning systems to enhance student learning. One key emerging area is in the use of models that can recognize student emotion in context, referred to as affective states. These models typically take the form of machine learning classifiers that recognize affect from the student’s interaction with an online learning system. In this project, the investigators will develop adaptive learning systems that actively enlist the help of teachers to develop better student affect detection methods. In return, the system will support the work of teachers by providing them reports on the affective state of each student in real-time. The system will then learn to mimic teachers’ choices of intervention methods for disengaged students in order to deliver interventions automatically. Overall, this project is anticipated to lead to i) better understanding of how to leverage and align to teachers’ perspectives in detecting and responding to affect, and ii) enhanced intervention by both teachers and automated software that re-engages students and improves learning outcomes.
This project will be organized into three phases. First, the investigators will employ active machine learning methods to ask teachers to observe specific students when they have a break in classroom activity; these methods can improve the quality of the affect detectors by providing data on the students whose affective states are most informative to improve the classifier, rather than the standard method of developing these detectors by observing students in round-robin fashion. Second, the investigators will incorporate richer data types (specifically, self-reported confidence ratings of affect labels) into the detectors to improve their quality. These self-reported confidence ratings reflect how uncertain humans are about specific affect judgements, which will be compared to the uncertainty of classifiers, to possibly reveal insights into student affect, such as what the properties are of situations where affect is ambiguous. Third, the investigators will use crowdsourcing to solicit ideas from teachers as to when specific affect interventions will be appropriate for specific students, and will develop automated intervention methods using reinforcement learning. These automated intervention methods are highly scalable since they can enable the system to take the actions the teacher would take to intervene to support different students experiencing negative affect at the same time. This intervention system will be tested in real classrooms as students learn within ASSISTments, a free web-based learning platform used by over 60,000 students a year. If successful, this project will lead to new scientific discoveries on the dynamics of affect and new technology for scalable student affect detection and intervention.