Peer-Based Intelligent Tutoring Systems: A Corpus-Oriented Approach

Abstract

In this thesis, we present an artificial intelligence approach for tutoring students in environments where there is a large repository of possible learning objects (e.g. texts, videos). In particular, we advocate that students learn on the basis of past experiences of peers. This aligns with McCalla's proposed ecological approach for intelligent tutoring, where a learning object's history-of-use is retained and leveraged to instruct future students. We offer three distinct models that serve to deliver the required intelligent tutoring: (i) a curriculum sequencing algorithm selecting which learning objects to present to students based on benefits to knowledge obtained by similar peers (ii) a framework for peers to provide commentary on the learning objects they've experienced (annotations) together with an algorithm for reasoning about which annotations to present to students that incorporates modeling trust in annotators (i.e. their reputation) and ratings provided by students (votes for and against) for the annotations they have been shown (iii) an opportunity for peers to guide the growth of the corpus by proposing divisions of current objects, together with an algorithm for reasoning about which of these new objects should be offered to students in order to enhance their learning. All three components are validated as beneficial in improving the learning of students. This is first of all achieved through a novel approach of simulated student learning, designed to enable the tracking of the experiences of a very large number of peers with an extensive repository of objects, through the effective modeling of knowledge gains. This is also coupled with a preliminary study with human participants that confirms the value of our framework. In all, we offer a rich and varied role for peers in guiding the learning of students in intelligent tutoring environments, made possible by careful modeling of the students who are being taught and of the potential benefits to learning that would be derived with the selection of appropriate tutorial content.iii Acknowledgements First and foremost I would like to thank my supervisor, Professor Robin Cohen, for guiding me through the completion of this dissertation and my PhD studies. Robin overwhelmed me with the amount of time and effort she put in to helping me complete my doctoral work. My committee members, Professor Edward Lank and Professor Peter van Beek, have provided detailed and helpful feedback throughout the process, which I greatly appreciate. Thanks as well to Professor Gord McCalla for his helpful advice. My appreciation goes out to the anonymous reviewers who considered earlier versions of this work and provided feedback during conference and journal submissions over the last 4 years.I would like to acknowledge Ahmed Ataullah, Simina Brânzei, John Doucette and Lachlan Dufton for being wonderful officemates: always willing to talk through issues I was working on, go for coffee and share a laugh. I greatly enjoyed working in the AI Group at the University of ...