Decisive Factors in the Annotation of Emotions for Spoken Dialogue Systems

Summary. The recognition of human emotions is a very important task towards im-plementing more natural computer interfaces. A good annotation of the emotional corpora employed by researchers is fundamental to optimize the performance of the emotion recognizers developed. In this paper we discuss several aspects to be consid-ered in order to obtain as much information as possible from this kind of corpora, and propose a novel method to include them automatically during the annotation pro-cedure. The experimental results show that considering information about the user-system interaction context, as well as the neutral speaking style of users, yields a more fine-grained human annotation and can improve machine-learned annotation accuracy by 24.52%, in comparison with the classical annotation based on acoustic features.