Automated Labelling of Dialogue Modes In Tutorial Dialogues

We present in this paper a study whose goal was to automatically label higher level constructs, called dialogue modes, in tutorial dialogues. Each tutorial dialogue is regarded as a sequence of utterances articulated by either the learner or the tutor. The dialogue utterances can be grouped into dialogue modes which correspond to general conversational phases such as dialogue openings, e.g. when the conversational partners greet each other, or serve specific pedagogical purposes, e.g. a scaffolding students’ problem solving process. Detecting dialogue modes is important because they can be used as an instrument to understand what good tutors do at a higher level of abstraction, thus, enabling more general conclusions about good tutoring. We propose an approach to the dialogue mode labeling problem based on Conditional Random Fields, a powerful machine learning technique for sequence labeling which has net advantages over alternatives such as Hidden Markov Models. The downside of the Condition Random Fields approach is that it requires annotated data while the Hidden Markov Models approach is unsupervised. The performance of the approach on a large data set of 1,438 tutoring sessions yielded very good results compared to human generated tags.