Rapidly Developing Dialogue Systems that Support Learning Studies Pamela Jordan, Michael Ringenberg and Brian Hall Learning Research and Development Center University of Pittsburgh {pjordan,mringenb,mosesh}@pitt.edu Abstract: We describe a dialogue system construction tool that supports the rapid development of dialogue systems for learning applications. Our goals in developing this tool were to provide 1) a plug-and-play type of system that facilitates the integration of new modules and experimentation with different core modules 2) configuration options that effect the behavior of the modules so that the system can be flexibly fine-tuned for a number of learning studies and 3) an authoring language for setting up the domain knowledge and resources needed by the system modules. Keywords: dialogue system, learning environments INTRODUCTION In this paper we introduce, TuTalk 1 , a dialogue system shell and content scripting language that supports the rapid development of dialogue systems to be used in learning studies. We intend to make TuTalk publically available once development is completed. TuTalk was strongly influenced by our past tutorial dialogue research. As part of this work we created a dialogue system and authoring tools to support our studies involving knowledge construction dialogues (KCDs). A KCD is a main line of reasoning that the tutor tries to elicit from the student by a series of questions. Typically at the end the tutor summarizes the main line of reasoning. This style of dialogue was inspired by CIRCSIM-Tutor’s directed lines of reasoning [Evens and Michael, 2006]. The KCD dialogue system was used in the Why2-Atlas [VanLehn et al., 2002, Jordan et al., 2006], Andes [Ros´ e et al., 2001] and ITSpoke [Litman and Forbes-Riley, 2004] physics tutoring systems. In addition, it was incorporated into the ProPL computer science tutoring system [Lane and VanLehn, 2005] and it was also used for a separate study of when during physics training dialogues are useful [Katz et al., 2005]. All of this work successfully engaged students in natural language dialogues and students were able to learn from the interactions. The goal for TuTalk was to redesign this software to provide 1) a modular architecture 2) easily modifiable system behavior to support a larger variety of learning studies and 3) easily authorable content. We anticipated addressing three classes of users; 1) those who intend to test complex dialogue hypothe- ses (e.g. is a speech-act pattern of inform->justify better than justify->inform for learning a concept) 2) those whose central hypothesis may not be about the fine points of dialogue but may nonetheless need a natural language dialogue capability to support their learning experiment (e.g. when will a natural lan- guage dialogue-like capability increase learning) 3) those who wish to measure performance differences of alternative natural language processing (NLP) modules or techniques (e.g. what effect does improved merging of sentences during language generation have on student comprehension). Two problems with experimenting with dialogue systems for the first two classes of users are 1) getting the desired configuration and system behavior necessary to support an experiment that tests a particular hypothesis and 2) setting up the domain specific resources required by the dialogue system to support the experiment. No one set of NLP modules or instantiation of an NLP module will be perfect for providing all the dialogue behaviors that an experimenter may need. But by adding some finer-grained control over the behaviors of modules, one set of dialogue system modules may support many experiments before other modules must be swapped in to provide the desired dialogue system behavior. While our goals of a plug-and-play architecture and easy configuration are not new to the NLP dialogue community, a tool that is also tailored to learning applications to ease the authoring process 1 An acronym for Tutorial Talk.