One-on-one human tutoring is remarkably effective. Seminal studies have shown that tutoring is significantly more effective than group instruction and may provide unparalleled opportunities for learning. A central, unanswered research question is, "How do expert tutors provide effective cognitive and motivational support over the course of long-term tutorial interactions to improve learning?" With a curricular focus of college-level computer science education, this project will see the design and evaluation of a computer-based intelligent tutoring system, JavaTutor, which leverages artificial intelligence to provide both cognitive and motivational support. The project will be conducted at North Carolina State University in conjunction with three partner institutions: Meredith College, Shaw University, and St. Augustine's College.
The project has three major thrusts. First, the research team will conduct a semester-long observational study of cognitive and affective tutorial support provided by expert human tutors interacting with students in a fully-instrumented online tutoring environment. The environment will log all tutorial conversations, problem-solving traces, and affective data streams including physiological signals, posture, and facial expressions. Second, the research team will develop an empirically grounded, integrated model of cognitive and affective scaffolding using machine learning techniques including hidden Markov modeling. Third, they will validate the integrated model of cognitive and affective scaffolding in a semester-long experiment with the JavaTutor intelligent tutoring system. Four versions of the JavaTutor system will be deployed and compared. It is hypothesized that over the course of a semester, the version with an integrated model of cognitive and motivational scaffolding will outperform each of the other models on both cognitive and affective student outcomes and yield differential effects across learner groups, accruing particularly significant benefit to low-performing and female students.
The products of this project include findings and technologies that will inform the future development of intelligent tutoring systems. By promoting rich learning interactions through integrated cognitive and motivational scaffolding, the project will create new learning environment technologies that promote high levels of achievement and find broad application in STEM education. It is anticipated that the resulting intelligent tutoring system technologies will serve as a foundation for the next generation of educational software that both complements and expands the impact of classroom teachers. The impact should be significant given the effectiveness of human tutoring and the potential power of these new technologies to support learning.
One-on-one human tutoring is remarkably effective. Seminal studies have shown that it is significantly more effective than group instruction and may provide unparalleled opportunities for affective scaffolding to support emotional states conducive to learning. A central, unanswered question in the field is how do expert tutors provide effective cognitive and affective scaffolding over the course of long-term tutorial interactions to improve learning? This project addressed that question by building on the extensive study of cognitive scaffolding within the intelligent tutoring systems community and leveraging an increasingly active body of research on the role of affect in designing intelligent tutoring systems. The project produced one of the largest corpora of tutorial dialogue ever collected. It includes measures of effectiveness through learning gains and a suite of affective outcomes. The corpus consists of computer-mediated textual dialogue for introductory computer science and is accompanied with logs of fine-grained problem solving, facial videos and Kinect depth videos capturing posture and gesture, as well as skin conductance traces. The corpus also includes pre-measures including demographics, self-efficacy, and attitudes toward learning, and it includes post-measures including engagement and frustration. Learning outcomes were measured with pretests and posttests. The research findings derived from this data set take steps towards (1) improving dialogue act classification with learner characteristics, and (2) modeling learning, engagement, and frustration. First, the findings regarding both supervised and unsupervised machine learning models of dialogue act classification have proved to be the most accurate ever reported in the literature for tutorial dialogue act classification, and hold great promise for developing highly scalable automated approaches in the future. Second, the results demonstrate that nonverbal behaviors at specific moments in the tutoring session are predictive of engagement, frustration, and learning, and reveal previously undiscovered relationships between learning-centered affective states and tutoring. The project’s results have been widely reported. They have been published in two journal articles and presented at thirteen national and international conferences, including the the leading conferences in intelligent tutoring systems, affective computing, and natural language processing: the International Conference on Artificial Intelligence in Education, the International Conference on Educational Data Mining, the Biannual Conference on Affective Computing and Intelligent Interaction, and the Annual SIGdial Meeting on Discourse and Dialogue. In summary, the project has advanced our understanding of how humans learn through tutorial dialogue, and has built foundational technologies for automated understanding and responses that can used for many tutorial dialogue projects in the future. The findings of the effectiveness of tutorial strategies based on problem-solving context, learner characteristics, and real-time affective states influence not only how tutorial dialogue systems will be designed in the future but also increase our understanding of how human tutorial dialogue can best support learners.
