Modeling and Inspecting the Question-Asking Process in Educational Dialogues

Опубликовано 12 августа 2016, 0:50

While many studies have demonstrated that dialogue-based tutoring systems have a positive effect on learning, the significant amount of human effort required to author, design, and tune system behaviors still provides a major barrier towards widespread deployment and adoption of these systems. Machine learning presents a path towards reduced human effort, however the custom-built nature of these systems means that any learned behavior is strictly tied to a single implementation. Ideally these behaviors should be able to extend to a variety of materials and concepts. To enable this kind of generalization will require a meta-level model of the dialogue that abstracts utterances to their action, function, and content. In this talk, I describe the DISCUSS dialogue move taxonomy, an intermediate representation that allows for lesson-independent modeling of dialogue behavior. To demonstrate the utility of this representation, I explore how DISCUSS-based features assist in the process of ranking and selecting follow-up questions within the context of the My Science Tutor (MyST) intelligent tutoring system. Moreover, I show how DISCUSS enables us to model and identify the factors driving the decisions made by experienced human tutors when teaching.