Publication Date

8-2019

Date of Final Oral Examination (Defense)

8-3-2019

Type of Culminating Activity

Thesis

Degree Title

Master of Science in Computer Science

Department

Computer Science

Supervisory Committee Chair

Dr. Casey Kennington, Ph.D.

Supervisory Committee Member

Dr. Jerry Alan Fails, Ph.D.

Supervisory Committee Member

Dr. Maria Soledad Pera, Ph.D.

Abstract

Current Digital Personal Assistants can be quite efficient while performing routine tasks like setting up reminders and looking up information. However, they do not attempt to establish common ground–the process of establishing and building mutual understanding–and require a significant amount of initial data to learn how to understand user intent. In this thesis, an incremental processing framework is leveraged through a chatbot interface which updates its understanding state at each inputted word, asks the user to clarify input when the system is unsure and prompts user to give feedback several times during an interaction, all of which are instrumental in establishing conversational grounding between them and enable the system to begin with little or no training data. User interactions can be utilized as labeled data for retraining the model and improving it. We evaluated our model with users on Amazon Mechanical Turk and with each iteration–retraining the model with the labeled data from previous interaction and opening it for new users–this conversational grounding model learns a mapping between the users' words and the actions performed by the system to improve the chatbots natural dialogue. Hence, demonstrating that since dialog processing involves language, it should be seen as a type of joint activity that requires coordination of both participants to establish common ground in order to communicate successfully and systems like these that have provisions for conversational grounding can work with little or no training data. Moreover, our data shows that with each update of the model the user affinity towards the system increased and the users prefer a system that asks for multiple clarifications over the course of interaction than a system that assumes understanding of utterances without giving any explicit feedback.

DOI

10.18122/td/1577/boisestate

Share

COinS