Machine Learning Thesis Proposal
- Gates&Hillman Centers
- ROBERT FISHER
- Ph.D. Student
- Machine Learning Department
- Carnegie Mellon University
Context Awareness and Personalization in Dialogue Planning
In this thesis proposal, we present techniques to improve the effectiveness of a conversational intelligent agent by incorporating rich, contextual information into the dialogue planning process. First, we propose the usage of linguistic features derived from semantic and discourse parsing to better understand the structure of a dialogue with a user. Discourse parsing gives us the relational structure of a dialogue and can be used to identify inter-sentential references and connections in a conversation. We have developed a novel, spectral algorithm for discourse parsing that is statistically consistent and computationally efficient. We also suggest the incorporation of non-linguistic contextual factors, such as a user's physical engagement or the aural characteristics of the environment. Modeling the diverse array of social factors we wish to include requires a very expressive state space representation.
To account for this, we propose dialogue planning using contextual information stored in a contextual knowledge base that can be combined with Predictive State Representations and imitation learning techniques. A conversational agent with access to a contextual knowledge base will be able to respond to environmental and linguistic conditions in ways that a simple question answering agent is not capable of. Our work is motivated by the education domain, and we intend for our methods to be included in a conversational intelligent tutoring system. By incorporating information about the student, the environment, and the state of the dialogue into a planning system, we hope to reduce the collaborative effort required to teach new concepts and to improve the overall effectiveness of conversational tutoring systems.
Reid Simmons (Chair)
Carolyn Rose ́
Dieter Fox (University of Washington)