Electronic Theses and Dissertations



Document Type


Degree Name

Doctor of Philosophy


Computer Science

Committee Chair

Vasile Rus

Committee Member

Arthur Graesser

Committee Member

Deepak Venugopal

Committee Member

Scott Fleming


Historically, Collaborative Problem Solving (CPS) systems were more focused on Human Computer Interaction (HCI) issues, such as providing good experience of communication among the participants. Whereas, Intelligent Tutoring Systems (ITS) focus both on HCI issues as well as leveraging Artificial Intelligence (AI) techniques in their intelligent agents. This dissertation seeks to minimize the gap between CPS systems and ITS by adopting the methods used in ITS researches. To move towards this goal, we focus on analyzing interactions with textual inputs in online learning systems such as DeepTutor and Virtual Internships (VI) to understand their semantics and underlying intents. In order to address the problem of assessing the student generated short text, this research explores firstly data driven machine learning models coupled with expert generated as well as general text analysis features. Secondly it explores method to utilize knowledge graph embedding for assessing student answer in ITS. Finally, it also explores a method using only standard reference examples generated by human teacher. Such method is useful when a new system has been deployed and no student data were available.To handle negation in tutorial dialogue, this research explored a Long Short Term Memory (LSTM) based method. The advantage of this method is that it requires no human engineered features and performs comparably well with other models using human engineered features.Another important analysis done in this research is to find speech acts in conversation utterances of multiple players in VI. Among various models, a noise label trained neural network model performed better in categorizing the speech acts of the utterances.The learners' professional skill development in VI is characterized by the distribution of SKIVE elements, the components of epistemic frames. Inferring the population distribution of these elements could help to assess the learners' skill development. This research sought a Markov method to infer the population distribution of SKIVE elements, namely the stationary distribution of the elements.While studying various aspects of interactions in our targeted learning systems, we motivate our research to replace the human mentor or tutor with intelligent agent. Introducing intelligent agent in place of human helps to reduce the cost as well as scale up the system.


Data is provided by the student.

Library Comment

Dissertation or thesis originally submitted to ProQuest