The purpose of this proposal is to explore the extent to which timely emotional, cognitive, and metacognitive interventions in tutoring software will have positive effects on students' emotions, attitudes, and achievements in mathematics. The intelligent tutor, Wayang Outpost, a high school mathematics tutoring system, is being enhanced to leverage automatic detection of emotions to guide cognitive, metacognitive, and affective forms of learning support.
The PIs are conducting a set of experiments to understand the interplay of observed emotional states, emotion assessments, student behavior within tutors, and student achievement. In particular, the experiments are testing the effects of the tutoring system when it assesses the emotions of a student and then responds with instructional support appropriate to that student's affect and content knowledge.
In this project, an interdisciplinary team of researchers in learning technologies and mathematics education are working together to investigate issues related to motivation in learning mathematics. They are taking the results of lab-based studies into classrooms. The novel technology and approaches developed in the lab were tested with a small population of learners; in their classroom-based investigations, they are testing feasibility of the approach with a more diverse population and refining the technology for use in a broad range of classroom learning environments. This translational research project will not only make significant contributions to the field of learning technologies, but will also contribute to our understanding of issues related to motivation in mathematics learning.
This project showed that an online tutoring system could identify and respond to student emotion. We made several accomplishments towards modeling and understanding how to address student emotion in smart digital learning environments and tutoring systems. We showed that activating emotions (e.g., math-liking) are highly correlated among each other, but not with deactivating emotions (e.g., boredom, anxious). Deactivating emotions are highly correlated among each other, particularly excitement and interest. Other results focused on two research agendas: 1) validation of models for automatic detection of student emotion and 2) responding to a few emotional states within an intelligent tutoring system for mathematics. The first research question we addressed was whether emotions as enumerated by theorists truly represent our own student emotion student reports. We implemented and verified that a tight relationship exists between our constructs and the more traditional ones in the literature. We conclude that our emotion self-reports capture similar achievement emotion constructs to those of Pekrun and colleagues, and similar math value and self-concept constructs as Eccles and Wigfield. Another research issue that we solved was the concern that deactivating negative emotions (such as boredom, disinterest, lack of excitement) might be impossible to model as a state-based emotional construct. Students are drawn to the level of control provided, and are interested in what the system says about their progress and how they are doing. One peculiar affective construct is confidence, a complex construct that strongly correlates to all of the following: pride, anxiety, frustration, self-efficacy and math liking. Thus, we cannot classify confidence as activating or deactivating. We implemented a mechanism to facilitate a kind of social interaction between students sitting next to each other. We present evidence of learning and progress, as students go through mathematics activities where pairs were collaborating by forcing math problems, and compare these within-tutor outcomes to other students that did not have this kind of emergent behavior. We analyzed whether new variables/features help in the prediction of higher level emotional states. We carried out a process of forward feature selection procedure in attempts to select the most important variables that would help to model and predict student emotions, based on recent behaviors and baseline motivation reports before starting to use the math software. Another issue that we addressed is which emotions exactly to model. Aggregating across emotions (modeling activating negative –frustration and anxiety together, deactivating negative –boredom and lack of excitement together) seemed reasonable, and particularly because we are attempting to intervene without differentiating between frustration and anxiety, for instance. Another research issue that we solved was the concern that deactivating negative emotions (such as boredom, disinterest, lack of excitement) might be impossible to model as a state-based emotional construct. Given floor levels in interest and excitement towards mathematics problem solving, our concern was that they did not fluctuate as much as other emotional states would, and that they would be simply too dependent to the baseline incoming disinterest in math, becoming a mood or a trait instead of a fluctuating emotion. We proved that this is NOT true by finding the contribution of the baseline emotion first (reported by the student at pretest time, e.g. "How boring is it to solve math problems?"), and saw how much more variance is picked up by contextual factors, such as the student not reading the last problem, etc. Basically, even though the baseline was important in the prediction of such emotions, contextual factors (recent history) were important too. The final research question that we addressed was that it is useful and beneficial to students’ affective states to provide interventions just based on student emotion. Student progress pages, which attempt to meta-cognitively support students in a self-regulation process and that are offered at precise moments of student boredom or lack of excitement, revealed improved levels of interest, decreased boredom and lack of excitement, compared to control groups for which the progress page was available but not offered at those key moments.
