Two different approaches to connectionist classifi- cation are studied in this work, in the context of a dialogue system. Classification of the user turns in terms of dialogue acts is accomplished with multilayer perceptrons. As we face a multiclass classification problem (there are 10 classes of dialogue acts), two approaches to classification are compared: binary multilayer perceptrons (one neural network for each class) and a unique neural network for all classes. In both cases, threshold values in order to accept or reject a given uterance as belonging to a particular class are needed for the neural classifiers. We use ROC graphs in order to select the threshold values of the classifiers to obtain the best tradeoff between accuracy and missclassification.
Cite as: Diaz, W., Castro, M.J., Ferri, F.J. (2004) Parameter setting of connectionist classifiers in a dialogue system. Proc. 9th Conference on Speech and Computer (SPECOM 2004), 474-480
@inproceedings{diaz04_specom,
author={Wladimiro Diaz and Maria Jose Castro and Francesc J. Ferri},
title={{Parameter setting of connectionist classifiers in a dialogue system}},
year=2004,
booktitle={Proc. 9th Conference on Speech and Computer (SPECOM 2004)},
pages={474--480}
}