Construction of a dialog system with emotions for elderly persons by neural networks

We have developed an interactive system with a dialog function that can analyze the emotions of a user and respond to him/her with facial expressions. The dialog system is based on our "emotion generating calculations" (EGC) method, which is based on C.D. Elliot's (1992) emotion-eliciting condition theory. The proposed method can judge whether an event is pleasant or not and obtain each attribute value for 20 various emotions under that paricular situation. Successfully, we consider adjectival relations and adverbial phrases, and we input their attributes into the EGC, such that we can evaluate the strength of each emotion. This system can depict the face according to the context of the conversation. Although each attribute value in the EGC shows one aspect of feeling, we must integrate those attributes to express human-like emotions. We applied a fuzzy reasoning model for selecting the face, as proposed by T. Yamashita (2000). In this method, nine faces with three levels of eye and eyebrow deflection and three levels of mouth deflection were used to construct the fuzzy reasoning model. In this study, we extended Yamashita's system of eight emotions to a sophisticated system with 20 emotions by means of a neural network.