The Influence of Context Knowledge for Multi-modal Affective Annotation

摘要

To provide successful human-computer interaction, automatic emotion recognition from speech experienced greater attention, also increasing the demand for valid data material. Additionally, the difficulty to find appropriate labels is increasing. Therefore, labels, which are manageable by evaluators and cover nearly all occurring emotions, have to be found. An important question is how context influences the annotators' decisions. In this paper, we present our investigations of emotional affective labelling on natural multi-modal data investigating different contextual aspects. We will explore different types of contextual information and their influence on the annotation process. In this paper we investigate two specific contextual factors, observable channels and knowledge about the interaction course. We discover, that the knowledge about the previous interaction course is needed to assess the affective state, but that the presence of acoustic and video channel can partially replace the lack of discourse knowledge.