Predicting the quality and usability of spoken dialogue services

In this paper, we compare different approaches for predicting the quality and usability of spoken dialogue systems. The respective models provide estimations of user judgments on perceived quality, based on parameters which can be extracted from interaction logs. Different types of input parameters and different modeling algorithms have been compared using three spoken dialogue databases obtained with two different systems. The results show that both linear regression models and classification trees are able to cover around 50% of the variance in the training data, and neural networks even more. When applied to independent test data, in particular to data obtained with different systems and/or user groups, the prediction accuracy decreases significantly. The underlying reasons for the limited predictive power are discussed. It is shown that – although an accurate prediction of individual ratings is not yet possible with such models – they may still be used for taking decisions on component optimization, and are thus helpful tools for the system developer.