There is no English translation for this web page.
Modeling user judgments provided while interacting with a spoken dialog system using hidden Markov models
LOCATION: Auditorium 1 , TEL, Ernst-Reuter-Platz 7, 20th floor
Date/Time: 19.05.2014, 14:15-15:00
SPEAKER: Philipp Heß
For the evaluation of spoken dialog systems, the quality of dialogs as perceived by the user plays a central role. "Direct” measurements of quality, however, can be costly or even infeasible. Models of user judgments thus serve as practical alternatives which, on top of serving evaluation purposes, can even permit real-time adaption during the interaction with a user. We modeled user judgments using hidden Markov models (HMM). Partially labeled data were taken from an experiment where users could judge dialogs with a spoken dialog system at arbitrary points during the interaction. Using HMMs with GMMs, we were able to correctly classify more than 99% of the user judgments. This out-performed the approach of using quantized data (44%) and the baseline (37%).
HOST: Klaus-Peter Engelbrecht