2019 年 18 巻 1 号 p. 17-25
Recently there is a growing need for affect-awareness in computer games. In-game emotion recognition, however, requires applying costly feature extraction methods and/or labor-demanding annotation of large datasets. To make emotion recognition cost-efficient, this study proposes (A) a social relations-directing, “emotion-sensitive” dialogue act model consisting of social acts, and (B) an approach for emotion recognition, utilizing the association strength ratio of emotion types and the proposed acts. In the study, five Japanese in-game dialogues were tagged with labels of emotion types and of social acts. Emotion type-social act co-occurrence was analyzed, and the corresponding association strength ratios were computed. In a validation experiment, the proposed approach was tested, enhancing a baseline emotion classifier's recognition accuracy by 8%.