For spoken dialog systems, PARADISE [Walker et al. 1997] provides a framework to train a user satisfaction prediction model on given data. The approach weights and sums interaction parameters to predict a satisfaction metric calculated from a questionnaire. In this paper, we try to tackle a major problem of these models, namely their weak generalizability. We show, that the weights associated with interaction parameters in the model change in dependence of the system's major problems by examining correlations under different quantities of understanding errors in the dialogs.
展开▼
