Feature selection for error detection and recovery in spoken dialogue systems
Item
-
Title
-
Feature selection for error detection and recovery in spoken dialogue systems
-
Identifier
-
d_2009_2013:3dfa86a5157d:10964
-
identifier
-
11327
-
Creator
-
Ligorio, Tiziana,
-
Contributor
-
Susan L. Epstein | Theodore Brown
-
Date
-
2011
-
Language
-
English
-
Publisher
-
City University of New York.
-
Subject
-
Computer science | Error Detection and Recovery | Feature Selection | Human-machine Interaction | Machine Learning | Spoken Dialogue Systems | Wizard of Oz Studies
-
Abstract
-
Spoken dialogue between people and machines is increasingly common, but not as flexible and complex as that between people. Spoken dialogue is susceptible to error---human speech is often loosely structured, people change their mind at mid-sentence, repeat themselves, pause, and produce non-speech sounds. A spoken dialogue system expected to handle complex and flexible dialogue like that between people must be robust to error, and employ strategies for error detection and recovery.;People recover from understanding error during spoken dialogue with little effort, and without disruption of the dialogue flow. To study how people would handle the same kind of errors that a spoken dialogue system must contend with, the work recounted here embedded a person within a spoken dialogue system with noisy speech recognition, that is, the person and the system processed information concurrently. The person was also ablated, that is, had input and output restricted to that of a spoken dialogue system. Run-time system features were collected to train data-driven models of human error detection and recovery. Our results indicate that people have successful dialogue interactions even with very high levels of noise, and they rely heavily on context and questions to disambiguate noisy speech recognition.;Empirical results also demonstrate that data-driven models of human error detection and recovery strategies can be obtained from system features alone. These models can draw from a rich and varied collection of system features across all levels of spoken language processing, and careful selection from among them is essential for effective modeling. To this end, this research developed two novel feature selection algorithms, one general and one knowledge-guided. They effectively and efficiently identify features for a particular machine learning algorithm and data set, and, guided by domain knowledge, they support adequate learning of error detection and recovery models.;The resulting models of human error detection and recovery were then implemented within a spoken dialogue system. Empirical results indicate these strategies support dialogue management to address speech recognition noise, and improve overall task success. Our results demonstrate that successful dialogue with a spoken dialogue system is possible despite high levels of understanding error.
-
Type
-
dissertation
-
Source
-
2009_2013.csv
-
degree
-
Ph.D.
-
Program
-
Computer Science
