**1. Introduction**

In recent years, communication robots have become popular in the real world [1]. To promote their use in the current society, the capability of communication in robots should be improved so that users can interact with robots more easily. One approach to realizing easy communication with a user is using humanoid robots. For example, robots have been used as communication media that provide information autonomously to visitors in public spaces such as a science museum [2] and a shopping mall [3]. Humanoid robots that resemble the appearance of human beings are called android robots. Owing to their appearance, an observer who sees android robots considers android robots as human for a few seconds [4]. Android robots can present a stronger sense of existence than other media such as video, or a speakerphone [5]. Android robots can effectively represent human-like mental states, including subtle emotion, by using nonverbal communication [6]. Therefore, it is argued that android robots can become influential conversation media that can effectively convey information [7].

In previous work on the conversation of agents, voice has been widely adopted as one of the most accessible and most universal modalities in many applications, such as an application for weather forecasts [8], for the guiding of buses [9], and for providing information to visitors in a science museum [10]. However, voice recognition in the real world has still been a formidable challenge [11]. Furthermore, especially in robot applications, it becomes more difficult because a human would often speak in less formally correct ways, such as in dialects, toward the human-like agent. Adopting a very human-like appearance in the android robot also causes a risk called the adaptation gap [12]: humans tend to be easily disappointed by the robot's utterance because they have heightened expectations for a

human-like, contextually natural response due to the appearance of human-like organs corresponding to voice production, i.e., mouth and throat.

To adopt verbal communication in robot applications without suffering from potential failure and disappointments, a specific form of communication called multi-party conversation has been developed in the field of human–robot interaction [13]. There is a study of robots as a passive social medium, where robots provide information for people with dialogue between two robots while not interacting with people directly [14]. It was reported that passive social robots succeeded in getting more attention from observers than a single robot receives alone. People who observe a scene in which two robots communicate with each other are more likely to treat those robots as if they are human and evaluate their dialogue as more understandable than they would for robots without communication between them [15]. However, without listening to a human response, it is difficult for robots to keep a human paying attention to the conversation for an extended period. Arimoto et al. showed that inter-robot turn-taking triggered by a human response contributes to maintaining the sense of conversation, even though the robots produced pre-scripted utterances and ignored the content of the human response [13]. However, the extent to which their conversation was successfully conveyed to humans has not been examined, especially in the case of affective content. Therefore, we constructed and examined a scenario-based, semi-passive conversation system using two androids that basically talked to each other and sometimes actively listened for human answers to their questions, including affective ones. To validate the usefulness of semi-passive social androids as a medium to provide information for people, a comparative experiment with passive social androids was conducted. Specifically, we evaluated the feasibility of semi-passive social androids to transmit the feelings of androids (subjective information) and the contents of the conversation (objective information) to people. To confirm the functionality of communicating subjective information, we prepared a script that had some statements for making participants feel empathic concern with androids and asked for their impressions of the androids through a questionnaire. Additionally, to confirm the functionality of communicating objective information, we added general information about androids in the script and measured how much information the participants could remember with a recall test. The between-subjects design experiment was conducted, where the subjects talked with either semi-passive social androids or passive social androids for ten minutes. After the talk, participants evaluated the androids and the conversations with a recall test and a questionnaire. From the result, it was verified that the semi-passive social condition was assessed higher than the passive social condition based on the information memorized by the subjects and the emotions conveyed and evoked by the androids.
