Knowledge Creating Conversational Agents Hidekazu KUBOTA Toyoaki NISHIDA Department of Information and Communication Engineering, School of Engineering, The University of Tokyo 7-3-1 Hongo, Bunkyo-ku, Tokyo 113-8656, Japan Abstract. This paper describes a conversational agent that supports knowledge creation in a community. We propose a conversational agent called virtualized-ego that represents the other self of a community member and can talk on behalf of the member. Knowledge of a virtualized-ego is compiled from everyday e-mails of a real community member. We have developed a system called EgoChat to investigate the ideas of supporting community knowledge creation by using virtualized-egos. We lay out an analysis of a mailing list and a method of creating the knowledge of virtualized-egos from it. How to maintain the knowledge of a conversational agents is discussed. 1. Introduction Recently, information technologies brought large potential for exchanging knowledge to humans. As a result, the research of a social agent that supports our creative process is now very active [1] [2]. This paper studies a social agent that called “virtualized-egos” (VEs). A VE is one’s other self; it works independently of one’s real self and can talk on one ’s behalf. Our great challenge is how to create such a conversational agent from our daily e-mails. Imp Characters [3] and Rea [4] are conversational agents that behave themselves like eloquent salespeople. Their knowledge is not related to real human, on the other hand, VE’s knowledge is compiled from everyday e-mails of a community member and a VE cooperates with humans as one of the community members. VEs work in our system called EgoChat, which supports community knowledge creation. With EgoChat, community members can exchange their experiences with each other even if some members are absent because VEs can talk on behalf of the absent members. First, we illustrate the overview of EgoChat. Secondly, we address an analysis of a mailing list and a method of creating the knowledge of VE from it. Finally, how to maintain the knowledge of VE is discussed. 2. Overview of EgoChat Figure 1 shows a screen shot of EgoChat. A user is talking with three VEs, each of which has the 3-D head of another community member. The EgoChat system consists of three storytelling processes and two storage processes (Figure 2). In the storytelling processes, community members can share their knowledge by creating a knowledge stream with their VEs. When a user wants to some community knowledge, she inputs her interest into EgoChat by voice [storytelling process (a)]. The voice messages are recognized by a commercial speech recognition system, and then VEs start talking about related topics [storytelling process (b)]. VEs will continue talking with one another unless the user interrupts the conversation to make a comment or change the subject [storytelling process (c)]. As the community member talks with VEs, her personal memory is enriched in the processes of storing comments. Before using EgoChat, the personal knowledge is stored in a VE by using automated summarizing technology or summarizing humanly from past exchanges on mailing lists [storage process (1)]. Besides text-based messages, VEs store new oral messages from a user on EgoChat [storage process (2)]. In this way, VEs share community knowledge in past messages of community members and users add new ideas to their VEs in a loop of community knowledge creation. Figure 1. Screen shot of EgoChat