For dialogue system to be accepted into people's lives, it is important to give people sense of afinity to them. In order to do that, we introduce face expression into dialogue system and suggest the architecture of face expression generation. Face expression express emotion, personality, and wiling. It can reduce inorganic and mechanical impression of systems and make afinity. In this architecture, emotion which is estimated or specified in scenario file is converted to face expression. Emotion is expressed two dimensional number based on Russell's circumplex model. Therefore, this is easily applied to other emotional actions such as voice tone. Finally we implemented the system and integrated to the dialog robot "WANCO". We evaluated that the user's afinity is enhanced by adding facial image.
pressions, the developer can either write the corresponFor AI such as dialogue system to be accepted as a hu- dence between facial expressions and speech in advance man partner, it is essential to make people feel familiar or automatically estimate emotions from text of speech. with AI. Especially, in areas such as health care, it is im- In this study, we examine two methods: one is for the portant to motivate users to use the system frequently. developer to specify facial expressions in advance in ruleOne way to give users a sense of afinity to systems is based dialogues, and the other is to create an emotion to express emotions. This would visually impress users dictionary to estimate rough emotions in order to generthe personality of the system, and motivates them to talk ate facial expressions from words in the text. to the system. In this study, we will improve the sense of afinity felt by users by implementing facial graphics that change face expressions in the interface of the 2. Socially Responsible AI for spoken dialogue system. Furthermore, we will suggest Well-being architecture of face expression generation. To connect dialogue manager system to face expression, we also con- Familiarity with AI integrates it into human society and sider a markup language to write emotional action. The gives it a role as a component of society. This is very architecture of the dialogue system is figure1. This is a important for the success of AI as more than just a tool. rule-based system in which the content of dialogue is described in advance as a scenario file.
antness, and the vertical axis represents arousal level. The red dots show examples of typical emotions. In the proposed method, emotions are specified by 2D coordinates in Russell’s circumplex model.
Face expressions express emotions, and therefore emotions can be converted to them. It should be possible to Table 1 estimate suitable face from conversation text. However, Number of responses for each option in the survey using emotion to decide face expression make some advantages. First, the emotional state of system becomes An experiment was conducted to determine whether clear. Furthermore, the emotions are able to be expressed the addition of facial expressions to the interaction with other means, such as tone of voice, gestures, and wording. the system could create a sense of afinity. Two faces The implementation of face model can be changed to oth- with and without facial expressions are displayed on the ers. When we decide the mapping between emotions and left and right sides of the screen, and several users are them, the system performs multiple emotional expression asked to talk with the system. After the dialogue, the consistently. On the other hand, the mapping must be users were asked whether their facial expressions were obtained by other means. In the current implementation, natural and they felt a sense of afinity with the system. emotional coordinates are mapped to facial movements, The results in table1 show that both were improved with e.g., eyebrow and mouth angle positions, eyelid openings, facial expressions. However, none of them received a low etc., for continuous facial expression generation. rating. This may be because of not strongly emotional scenario, so the presence or absence of facial expressions did not make much unnaturalness.
As figure3, emotions estimated or specified in scenario ifles is holded in emotion manager. And then it is passed to face-expression interface as EmojiML. EmojiML describes emotion and other emotional actions. The example of a tag embedded EmojiML to make system to say "Thank you" with pleasure is as below. <f-exp type="pleasure">Hello</f-exp>
Examples of face expression generation is figure4.
users a positive impression. The consideration of correspondence between facial expressions and emotions is insuficient. In order to make natural face, find suitable method of generating face asking users opinion.