Improving Efficiency of Natural-Language Text Generation for Automatic Pedagogical Questions

When learning new skills, students may experience problems in understanding the course material and may require a tutor's help in the form of educational scaffolding. That can be organized in the form of learning dialogue consisting of closed-answer questions and explanations for answers to them. However, designing Intelligent Tutoring Systems that support generating pedagogical questions is very labor-intensive. In this paper, we propose a method of improving template-based generation of pedagogical questions by moving the templates from the level of nodes of Thought Process Tree describing the problem-solving reasoning process to the smaller elements like expressions, domain objects, their classes, properties, and relationships. The method allowed to reduce the number of templates used in the intelligent tutoring system for learning programming-language expressions by about half and the number of words used in templates by about 60%.