Temperature Based Dialogue in the Duke Programming Tutor

The Duke Programming Tutor is a multi-modal dialogue system designed to help novice Computer Science students write, debug and understand simple programs. In this paper, we describe the architecture of this system at a high level, and present and analyze results from a preliminary test with students. A temperature based dialogue algorithm is proposed in order to improve the modeling of tutorial dialogue and system robustness. We then describe a version of this algorithm that has been integrated into the Duke Programming Tutor. 1 Communicating with Computers If we wish to work with friends or colleagues to carry out a joint task, such as studying or scheduling, we usually prefer to use voice as the primary mode of the collaboration. However, on the computers that we work with every day, few systems attempt to integrate voice as anything but an afterthought to interactions. The most common commercial voice systems are dictation systems and voice command systems for short cuts in windowing environments. Neither of these tasks are particularly collaboratory{the computer is generally a passive participant. There have been several research systems built that have used voice as a primary means of communication. The TRAINS project Allen et al., 1995] studies planning in a train scheduling domain as the topic of human-computer dialogues. The TRAINS system focuses on understanding the details of language used in dialogue, but it is not their primary goal to produce a real-time 1