Chatbot-like dialogue agents typically use character-anchored question-answer pairs for response generation. However, creating such pairs requires manual effort and is very costly. We propose using role play to efficiently collect such character-anchored question-answer pairs. In our role-play-based scheme, users play the roles of certain characters and respond to questions. Since multiple users can play the role of the same character, questions addressed to one character can be quickly answered by any of the role-players of that character. This leads to the efficient collection of character-anchored question-answer pairs. In addition, role play has a certain entertaining nature, which could encourage the active participation of users. From the analysis of a trial run of a website based on our scheme, we found that users were highly motivated to provide data and that appropriate question-answer pairs were collected, which suggests the effectiveness of role play for dialogue data collection.
Bibliographic reference. Higashinaka, Ryuichiro / Dohsaka, Kohji / Isozaki, Hideki (2013): "Using role play for collecting question-answer pairs for dialogue agents", In INTERSPEECH-2013, 1097-1100.