Uniclass and Multiclass Connectionist Classification of Dialogue Acts

Classification problems are traditionally focused on uniclass samples, that is, each sample of the training and test sets has one unique label, which is the target of the classification. In many real life applications, however, this is only a rough simplification and one must consider some techniques for the more general multiclass classification problem, where each sample can have more than one label, as it happens in our task. In the understanding module of a domain-specific dialogue system for answering telephone queries about train information in Spanish which we are developing, a user turn can belong to more than one type of frame. In this paper, we discuss general approaches to the multiclass classification problem and show how these techniques can be applied by using connectionist classifiers. Experimentation with the data of the dialogue system shows the inherent difficulty of the problem and the effectiveness of the different methods are compared.