Anton Leuski and David DeVault: “A Study in How NLU Performance Can Affect the Choice of Dialogue System Architecture”

July 5, 2012 | Seoul, Korea

Speaker: Anton Leuski and David DeVault
Host: SIGdial 2012

Abstract: This paper presents an analysis of how the level of performance achievable by an NLU module can affect the optimal modular design of a dialogue system. We present an evalua- tion that shows how NLU accuracy levels im- pact the overall performance of a system that includes an NLU module and a rule-based di- alogue policy. We contrast these performance levels with the performance of a direct classifi- cation design that omits a separate NLU mod- ule. We conclude with a discussion of the po- tential for a hybrid architecture incorporating the strengths of both approaches.