Social interactions can support the treatment of emotion-related problems by aiding a person's emotional process. A number of studies have showed a consistent inclination of humans to talk about and socially share their emotional experiences, especially for an intense and/or negative emotion exposure. Although we have seen encouraging progress in affective human-computer interaction, the potential benefits for users by incorporating emotion in computer interaction are not yet studied in depth. For example, emotion elicitation looks at the change of emotion in dialogue, however its application to for emotional improvements is not yet well researched. Furthermore, although there exist technologies that address clinical emotional disturbances, such as depression and distress, there is a lack of research on emotion improvement from negative emotional exposures commonly encountered in everyday life.
The goal of this thesis is to diminish these gaps. In particular, I aim for chat-based dialogue systems with an implicit goal of eliciting emotion improvements though dialogue interactions. I examine the concept of eliciting emotional improvements through two perspectives: short-term and long-term. Assuming positive emotional state as the goal, short-term elicitation of emotional improvement is reformulated into turn-based positive emotion elicitation. As the effort matures, complexity of the problem is incrementally increased by incorporating various dialogue aspects that relate to the elicitation goal. Long-term emotional improvement is then inspected by extending the positive emotion elicitation scope to the entire dialogue.
First, I study emotion processing and negative emotion recovery in human communication through corpus construction and analysis. Second, to endow dialogue systems with an attunement of emotional context in dialogue, I propose novel neural network architectures that allows a dialogue system to track emotion and incorporate this information in the response generation process. Third, novel methods to learn dialogue strategies for short-term positive emotion elicitation in chat-based dialogue systems are proposed. Aspects that contribute to elicitation success are inspected: emotion, dialogue action, and emotional impact. Lastly I present the result of preliminary study on long-term emotion improvement elicitation. The dialogue structure allowing long-term emotion improvement is identified, and simulations using language modeling techniques are examined. The efforts presented in this thesis should serve as the basis for future efforts in supporting emotion improvement through human-computer interactions.