RI: SMALL: LexE: Using Two-part Lexical Entrainment for More Efficient and Reliable Spoken Dialogue Systems

Carnegie Mellon University, Pittsburgh PA

Investigators

Abstract

When humans speak to each other and want the dialogue to go well, they adapt to each other?s manner of speaking, using the same words, grammatical constructions and expressions. In order to make fundamental improvements in the performance of spoken dialogue systems, LexE is using subtle techniques to model this adaptation, which is called lexical entrainment. LexE is getting users to adapt their speech to the system, as well as getting the system?s speech to adapt to what the user says. To do this, LexE studies human-human dialogues to find the words and constructions (the ?primes?) that are often adopted by dialogue participants. The spoken dialogue system then uses these primes in its output. The system also detects the expressions that its user employs to refer to objects uses them in its synthetic speech. Two real-user spoken dialogue systems are being used as test platforms for LexE. The first is a bus information system for the Port Authority of Allegheny County; the second is the City of Pittsburgh 311 non-emergency service. By making these publicly-available spoken dialogue systems easier to use, LexE makes them (and other spoken dialogue systems) more accessible to a large part of our population, many of whom, the elderly, for example, get much of their information over the telephone. The techniques developed in this project also provide insights for the education of non-native speakers and for speech therapy, where tutoring systems can imitate the way humans implicitly correct errors in what their interlocutors say.