Statistical spoken dialogue systems (SDSs) use reinforcement learning to learn a dialogue policy that decides what to do based on the dialogue context (also called dialogue state). Previous work on this problem has mainly addressed slot-filling dialogue, in which the user presents a complex request (e.g. an appointment booking), and the system tries to fill a set of slots (e.g. date and time) to satisfy the user's request. This project significantly extends and generalizes prior work by allowing automated dialogue policy creation for other genres of dialogue including question-answering and negotiation. The following open research issues are investigated: (1) the extent to which the three very different genres of dialogue (slot-filling, question-answering, and negotiation) can be represented using the same kind of dialogue policy representation; (2) whether state-of-the-art learning techniques, that work well for small state spaces and simple interactions, can scale to the needs of more complex dialogues and larger state spaces; (3) methods for compactly representing the dialogue state and for combining learned and hand-crafted policies; (4) development of automated metrics for measuring the quality of simulated users and learned policies; (5) validation of those metrics with respect to how well they correlate with human evaluations.

Statistical SDSs facilitate easier creation of dialogue systems (less hand-crafting by dialogue system experts) that are more tuned in to user behavior (learning policies from data and simulation). This project broadens the types of systems that can be developed with this kind of approach (not just slot-filling, but also simple question-answering and more complex negotiation). The advances made in the project are encoded in a toolkit (to be publicly distributed) specifically designed for statistical dialogue management. This toolkit allows broader access to this technology, by providing the potential to attract more researchers from academia and industry to the field of SDS, and make the use of statistical techniques available to non-experts; it can also be an excellent resource for teaching statistical dialogue management to students.

Agency
National Science Foundation (NSF)
Institute
Division of Information and Intelligent Systems (IIS)
Type
Standard Grant (Standard)
Application #
1117313
Program Officer
Tatiana D. Korelsky
Project Start
Project End
Budget Start
2011-08-01
Budget End
2014-07-31
Support Year
Fiscal Year
2011
Total Cost
$449,988
Indirect Cost
Name
University of Southern California
Department
Type
DUNS #
City
Los Angeles
State
CA
Country
United States
Zip Code
90089