Login
Computer Science @ Rochester
Thursday, April 26, 2012
9:00 AM
CBS Room 703
Ph.D. Thesis Proposal
Katherine Lang
University of Rochester
AID: An Intelligent Dialogue System for Food Relief Assessment Interviews
Growing demand for food aid has provided motivation for research into how to better handle these needs without causing more damage to the affected area. MIFIRA, a method of analysis for food aid distribution in food insecure areas, is currently being used in some African countries. However, the tree-based implementation of the data collection process on which MIFIRA is based is neither time nor resource efficient and would benefit greatly from the introduction of electronic help.

Our research aims to provide that assistance with the AID natural language dialogue system. The AID system, or Artificially Intelligent Interview Dialogue system, is a guidance system that will assist interviewers, not familiar with the reasoning behind the MIFIRA interviews, in successfully collecting consistent and accurate information in the field. Most programs used to collect data via interviews, such as the MIFIRA assessment interviews that are available to the public and not off-the-shelf tend to utilize ridged state-based systems with predefined responses for information collection. However, these systems lack the flexibility that is necessary for interviews containing open-ended questions, such as “Why do you prefer one market over the others?”, that cannot be simply separated into a finite number of categories or exact phrases.

AID's dialogue manager utilizes a finite-state automaton and extends the framework for increased flexibility. These extensions include logical form matching to interpret utterances provided by the robust parser, a variant of the TRIPS parser, and the incorporation of mixed-initiative dialogue. Semantic parsing done by the TRIPS parser adds robustness to the system including the use of semantic rules, which the dialogue manager can use to cover a broad selection of answers.

Interestingly, our state-based dialogue control is far more flexible than McTear's dialogue control classification predicts. Moreover, our variant of the TRIPS parser can handle partial parses and extra information as well as natural language input from the user, unlike most state-based NLDSs [Ferguson and Allen 1998]. Finally, AID strives for a conversational feel to put less pressure on the user to know exactly what he/she should say to collect the interview data, including the familiar AIM chat window as the GUI so the user can interact with the program more naturally.

After defining the domain for our research, high level background information will be given about MIFIRA and natural language dialogue systems, including more in-depth discussions of the natural language understanding component and the dialogue manager. Our focus will then turn to the design of the AID system and how it handles the given domain, comparing the advantages and disadvantages of the different types of dialogue control. We will conclude by discussing the future of this project and what improvements will be explored for the AID system.