A Dialogue Tutoring Corpus and Agent for Teaching Programming

Abstract

When it comes to learning, one-to-one tutoring is a very e ective method. However, in our current educational system, this is not always feasible, since one teacher teaches a whole class and has little capacity to tutor each student separately. With the rise of arti cial intelligence, new possibilities emerge. There are already some approaches to building intelligent systems that act like tutors and converse with students. Unfortunately, there are not many large open datasets containing conversational data. In this paper, we introduce a new dataset: the WHJ dataset containing 6K conversations and 18.5K utterances. WHJ is a dataset based on chat messages that were sent between tutors and students on the learning platform White Hat Jr1. On this platform, students learn to code using simple coding projects. Each student has a tutor they can ask questions regarding the project via chat. These chat messages and the corresponding information about the project are the basis of WHJ. In this thesis, we analyse di erent dialogue models whilst working with WHJ and the CIMA dataset by Stasaski et al. [2020].