1 Introduction

Fig. 1
figure 1

Social mediator robots: robots connected via the internet serve as mediators, and humans can exchange messages through the robots [1]

In aging societies, the social isolation of elderly people is becoming a serious issue. Due to a trend toward nuclear families, the population of elderly people who live separately from their family members is increasing. In addition, it is important to note that elderly people do not often reveal their true emotions easily [2, 3]. Even if they are facing serious health or financial problems, it may be difficult for them to self-disclose their troubles to others, including their family members or friends. To prevent the social isolation of the elderly, it is necessary to provide them with the ability to smoothly communicate with their family members or friends while encouraging their self-disclosure.

Meanwhile, the advent of text messaging devices, such as smartphones, made remote communication instant. Currently, these devices are often used by elderly people to keep in touch with their family members and friends in remote places. Furthermore, compared with other classic text-based devices, modern devices with voice recognition capabilities provide a more efficient message creation process for users (such as smart speakers). Although there is still room for improvement, they are particularly welcomed by users who feel stress when texting. Social robots have also been employed as text messaging devices [1, 4, 5]. In this framework, the social robots placed in each house are connected via the internet, and the users can exchange their messages through the (mediator) robots (Fig. 1). More importantly, research evidence suggests that the use of social mediator robots could encourage the elderly to self-disclose to their family members [6]. In addition, the design guidelines of the social mediator robot have been partially identified. For example, personality traits, such as high or low extroversion and high or low neuroticism in the elderly, as well as the type of recipient, such as family or friends, also influence the design of robot behavior [1].

The design guidelines proposed in [1] summarize the requirements for social mediator robots to be used with the elderly (left in Fig. 1). However, there has been no investigation of the behavior of the robot assisting recipients (right in Fig. 1). Robot behaviors, particularly the way messages are conveyed to the recipients, may also affect self-disclosure of the elderly. Specifically, the anxiety felt by the elderly discloser may be exacerbated if the robot’s behavior in delivering the messages is inadequate.

Fig. 2
figure 2

A human-robot interaction used in this study to define behaviors to convey user messages. At first, the robot assisting the elderly speaker offers several messaging options after the elderly speaker’s self-disclosure. In this example, “indirectly” was chosen by the elderly user. According to this information, the recipient-side robot conveyed the message in an indirect way

Anxiety perceived in interpersonal situations is called social anxiety [7]. This anxiety is known to have a strong relationship with the concept of self-presentation, which is defined as “how people attempt to present themselves to control or shape how others view them [8].” According to the self-presentation theory of social anxiety [7], an individual’s confidence in his or her ability to self-present must be increased to decrease the level of their social anxiety. To increase such a confidence of elderly disclosers, the robot should have multiple messaging options (the ways of messaging) to support their self-presentation. Then, each elderly discloser can choose a messaging option based on his or her preference by interacting with the robot. For example, if an elderly discloser wants the mediator robot to convey his or her message to a recipient in a soft and indirect way, the robot can add words or phrases accordingly (Fig. 2).

This paper aims to present knowledge on designing messaging options for social mediator robots and experimental evidence on their effectiveness. To identify effective messaging options, we conducted a large-scale online survey by recruiting 589 elderly participants (Study 1). By asking the participants about their preferences for each of the 15 pre-designed messaging options, we were able to remove some options that were preferred by significantly small numbers of participants. In addition, differences in messaging option preferences according to the situation were identified; they include factors such as the gender and personality traits of elderly disclosers, recipient type (family members/friends), and conversation topic. Then, through an empirical human-robot interaction (HRI) study, the anxiety that the elderly disclosers perceived during self-disclosure was measured (Study 2). The elderly participants (\(N=36\)) were asked to interact with a mediator robot in a situation that required self-disclosing their loss experiences to their own family members who are living in a remote place. There were two types of mediator robots: mediator robot 1: when delivering the message, the robot merely repeats the original message contents of the discloser by using its voice (conventional mediator robot) and mediator robot 2: when delivering the message, the robot applies the messaging option chosen by the elderly discloser. The messaging options developed in Study 1 were implemented in mediator robot 2. Thereby, the effectiveness of the messaging options can be discussed.

The remainder of this paper is organized as follows: Section 2 presents related works. Sections 3 and  4 describe two studies (Study 1 and Study 2, respectively). Section 5 provides general discussions and the limitations of the study, and Sect. 6 presents our conclusions.

2 Related Works

2.1 Self-Disclosure in Old Age

Self-disclosure was defined by Jourard as “an accurate portrayal of the self to others [9].” In other words, self-disclosure is the spontaneous verbal expression of humans that reveals self-emotion (evaluative intimacy) and self-information (descriptive intimacy) to others.

In developmental psychology, Erikson et al. defined the model of psychological conflict that arises at approximately 65 years old as integrity vs. despair [10, 11]. Both are experienced when people reflect on their lives and think about what they have or have not accomplished. If they cannot integrate during their lives, they will die in despair. Hence, self-disclosure becomes important, particularly for elderly people. Sharing troubles with other people is a trigger for social support [12] and is thus essential for them to enhance their possibility of surviving (social support is classified into four factors- emotional support, material support, informational support, and appraisal support [13]).

Although there are many benefits for disclosers, troubles represented by loss experiences, i.e., loss of physical/mental health, loss of economic base, loss of social networks, and loss of reasons to live, are known to be difficult for elderly people to disclose even with those close to them (e.g., intimate friends or family members) [14]. In general, self-disclosing behavior has been seen as the product of two opposing forces: operating to increase disclosure vs. operating to inhibit disclosure [15, 16]. [15] noted that self-disclosure avoidance occurs due to the perceived harmful consequences of self-disclosure, and it depends on factors such as the gender of the discloser and dialog topic.

In contrast, interpersonal communication often becomes smooth when there is a third person mediator. Anyone, not only a professional with specific expertise such as a broker, can mediate a relationship between two persons (e.g., sometimes a family member can mediate a discussion between other members). If there is intermediation by a third person, the difficulty of self-disclosure could be alleviated, and elderly people might be encouraged to request social support. In fact, the latest findings reported in the HRI field (see next section) imply that if robots to mediate communication for elderly people were designed properly, more intimate self-disclosure could be encouraged.

2.2 Robot-Mediated Communication

In the context of robot-mediated communication, telepresence robots and robotic social mediators that are used in remote/local real-time human communication have been discussed. For remote communication support, telepresence robots have been used as mediators of interpersonal communication and tested for use as elderly support (e.g., healthcare [17, 18], working support [19, 20]). Furthermore, social mediator robots that support local interpersonal communication have also been proposed [21,22,23]. These studies suggest that through interaction with mediator robots, humans may be encouraged to communicate with other people (e.g., elderly people motivated to interact with caregivers during their interactions with a PARO robot [22]).

Although these studies show the potential of mediator robots for elderly users, design knowledge with respect to facilitating self-disclosure is still insufficient. The messaging functions of robots have been studied (e.g. [4, 5]). However, the effects on the communication/self-disclosure of elderly people have not been reported, and effective design guidelines for social mediator robot have not been presented. Furthermore, individual differences are particularly important when using social robotics in mental healthcare [24], mediator robots should be designed in accordance with individual user properties (e.g., personality traits, gender). Mediator robots should be able to adapt to users and change their behaviors according to the topic of each self-disclosure as well as the properties of users.

From this background, effective design guidelines for social mediator robots have been explored. For example, dialog topics in which social mediator robots could well encourage elderly self-disclosure [6, 25] were investigated. The results indicated that in topics concerning loss experience, social mediator robots can be more effective than direct conversations over the phone. Next, the personality factors of the mediator robots and elderly speakers were also discussed, and detailed findings to match their personality traits involving extroversion, neuroticism, and human gender [1] were reported. Additionally, a follow-up study revealed that social mediator robots are particularly useful for male disclosers with high neuroticism; they self-disclosed deeper in using the social mediator robot condition than in the face-to-face condition. However, design guidelines for social mediator robots on the recipient side are still unclear, and further investigation focusing on behavior switching in conveying each of the self-disclosed messages is needed.

3 Study 1: Online Survey

In this study, we created 15 messaging options (shown in Sect. 3.1). By performing an online survey, we explored (1) which messaging options the social mediator robot should have and (2) differences in the preference for each messaging option according to the situation. More specifically, the situation comprises the following elements:

  • Topics of self-disclosure: health, finance, isolation, and reasons to live

  • Recipient type: family or friend

  • Discloser’s gender: male or female

  • Discloser’s personality traits: high neuroticism or low neuroticism

A previous study revealed that the neuroticism traitFootnote 1 of disclosers significantly affect the depth of self-disclosure when using social mediator robots [1]. More specifically, when a social mediator robot was used, male participants with high neuroticism self-disclosed more deeply than the case when in a face-to-face condition for certain topics. Additionally, following the study on computer-mediated communications (CMCs) [30], people with high neuroticism tend to like CMCs more than people with low neuroticism. Therefore, in this study, we also investigated differences in the neuroticism trait of disclosers.

3.1 Design of the Messaging Option

According to the self-presentation theory of social anxiety, perceived anxiety increases when people are unsure whether they are giving others the impression they want to [7]. Therefore, if the kind of impression they want to give or not give to others is known, it would allow us to design the robot’s behavior accordingly. Additionally, in the field of psychology, certain feelings are known to be associated with a reluctance to self-discloseFootnote 2. Therefore, we designed the robot’s behavior to encourage self-disclosure of the elderly based on the knowledge of the role of these feelings.

In psychology, two types of aversion to self-disclosure are discussed depending on the direction of the aversion (intrapersonal aversion and interpersonal aversion) [31]. Regarding intrapersonal aversion, negative feelings tend to be promoted when disclosing negative emotions [32], and self-evaluation decreases due to the discloser self-perceiving negative aspects of themselves [33]. Regarding interpersonal aversion, disclosing negative topics can weaken relationships with recipients [34], and can annoy even close relations. To mitigate these aversions, the robot may need to have messaging options that limit the information conveyed to the recipient. As an example of these behaviors, we created the following three messaging options. DISCLOSER is to be replaced by the name of the elderly discloser, and MESSAGE is to be replaced by the elderly’s original message to the robot. XX is to be replaced by short terms representing the topics of MESSAGE.

  • (Option 1) Deliver the message to the recipient as a speculation of the robot while concealing that the elderly user disclosed his or her troubles; when it speaks, the script is “Maybe I’m thinking too much, but I believe DISCLOSER is troubled about XX recently.”

  • (Option 2) Deliver the message to the recipient as a common trouble of elderly people while concealing that the elderly user disclosed his or her troubles; when it speaks, the script is “By the way, it seems there are many elderly people who are troubled about XX. However, I heard that it is hard for them to say this for themselves.”

  • (Option 3) Deliver the message to the recipient without revealing the true emotions of the elderly discloser; when it speaks, the script is “DISCLOSER said that there is nothing troublesome and he or she is doing well, but if he or she was forced to say something, DISCLOSER would say MESSAGE.”

On the other hand, sometimes people feel hopeless (have no positive expectations) about self-disclosure [31]. Usually, self-disclosure can facilitate social support from recipients [12]. However, if disclosers give up on receiving social support from recipients, they could stop their self-disclosures. In social support studies on elderly people, the emotional and material support provided by caregivers have been discussed [13, 35, 36]. Therefore, robot behaviors that could facilitate mental/material social support from recipients may be useful, particularly for elderly users who feel hopeless about their self-disclosures. As an example of these behaviors, the following three messaging options were created:

  • (Option 4) Ask the recipient to help the discloser, not just deliver the message; when it speaks, the script is “DISCLOSER said that MESSAGE. I know you are very busy, but would you please help DISCLOSER? I hope you help him or her.”

  • (Option 5) Tell the recipient that other elderly people are being helped by their family members and friends, not just deliver the message; when it speaks, the script is “DISCLOSER said that MESSAGE. There are a lot of elderly people who have such troubles, but it seems that the number of family members and friends who support them are increasing.”

  • (Option 6) Emphasize the discloser’s current feeling that he or she is suffering when it delivers the message to the recipient; when it speaks, the script is “I heard that DISCLOSER had (discloser’s state, such as sad) events. Then, he or she said that MESSAGE.”

In addition, six messaging options were also created based on the personality traits of the robot at the recipient side. This may also influence the self-disclosures of elderly people. In designing these options, we referred to three personality traits (extroversion, conscientiousness, and neuroticism) chosen from the Big Five personality modelFootnote 3. Based on those personality traits, the following options were created and can be used with the verbal behaviors introduced above.

  • (Option 7) Lively; when speaking, the robot uses a lively tone of voice.

  • (Option 8) Calmly; when speaking, the robot uses a low tone of voice.

  • (Option 9) Seriously, when delivering the message, the robot behaves seriously.

  • (Option 10) Informally; when delivering the message, the robot behaves informally.

  • (Option 11) Anxiously, when delivering the message, the robot tells the recipient about the discloser’s problem in a way that expresses the robot’s worry.

  • (Option 12) Optimistically; when delivering the message, the robot tells the recipient about the discloser’s problem in a way that expresses the robot’s optimistic thinking on that the topic.

Last, the number of nonverbal cues may also affect self-presentation, especially for computer-mediated communication [37]. Therefore, three messaging options were created to represent the messaging format, which contains various amounts of nonverbal information.

  • (Option 13) Deliver the message as it is; when delivering the message, the robot reads the original message of the discloser by using its voice.

  • (Option 14) Deliver as a voice message; the robot records a speech of the elderly discloser and plays it to the recipient.

  • (Option 15) Deliver as a video message; the robot records a video of the elderly discloser and plays it to the recipient.

3.2 Participants

Participants for this study were recruited through MACROMILL, Inc., which is a leading market research company in Japan. Initially, we obtained 720 participants over 65 years of age (50% female, 50% male; age: \(M = 69.8\) years, \(SD = 4.11\)). However, we excluded data obtained from 131 participants who were not confident they understood the entire survey procedure (measured by a question item); therefore, data from 589 participants were used in the analyses. The fieldwork was conducted from December 17th to December 19th, 2018 by MACROMILL, Inc.

The target profiles of the participants were (1) persons who had communicated with their own children and intimate friends by phone or face-to-face at least once in the past three years; (2) persons who were living alone or with only their spouses; and (3) persons who were concerned about any of the loss experience topics (health, financial, isolation, and reasons to liveFootnote 4) and had never disclosed these concerns to other people. To exclude participants who did not meet these requirements, a screening process was carried out before proceeding to the main survey.

Additionally, we assumed two types of recipients (family/friend). The family recipient was defined as the family member from a young generation who communicated with the participant the most. The friend recipient was defined as the intimate friend of over ten years who communicated with the participant the most. The reasons for considering these two types of recipients were as follows: (1) relationships between disclosers and recipients are known to be an important factor when disclosers determine the breadth and the depth of topics they discuss [38], (2) the social support requested from family members can differ from that requested from friends [35].

Each participant imagined one type of recipient (family or friend, between-participants factor). The family recipients selected by the participants included daughters (50.0%), sons (29.5%), sons’ wives (4.5%), and others (< 1.0%). Fifty-two percent of the family recipients were the same gender as the disclosers. Friend recipients had an average age of 68.3 years (\(SD = 6.20\)), and 93% of the recipients selected were the same gender as the disclosers. To have a better awareness of each recipient during the survey, all participants were asked to complete forms with the initials of the recipients.

3.3 Measurement

3.3.1 Preference for Messaging Options

To ask the participants about their preferences for each messaging option, we created 15 questionnaire items (Table 1). We asked the participants to indicate the degree of their agreement toward using each messaging option on a scale ranging from “1: strongly disagree” to “7: strongly agree.” Higher numbers indicate more preferable robot behavior as evaluated by the participants. When asking the participants about their preferences for the use of options 1–6, the dialog scripts shown in Sect. 3.1 were also presented as examples to help the participants understand.

Table 1 Question items used to measure the preference of elderly disclosers for each of the messaging options

3.3.2 Personality Traits

A Big Five scale [39] was used to evaluate the neuroticism of each participant, which was represented by five items: easily feel uneasy, anxious mind, get discouraged, easily feel nervous, and depressive. A seven-point scale ranging from “1: strongly disagree” to “7: strongly agree” was used in answering question items concerning those scales.

3.3.3 Participant’s Comprehension

As explained in Sect. 3.2, the participant’s comprehension of this survey was checked by using two questions: “I could not understand some video contents” (the video contents will be explained in the next section) and “I could not understand how the mediator robot works.” The participant could assess his or her comprehension based on a seven-point scale ranging from “1: strongly no” to “7: strongly yes.” The results data obtained from participants who marked a 5–7 on either/both questions were excluded from analyses.

3.4 Procedure

At the beginning, an experimenter explained the basic concept of social mediator robots and the purpose of this survey to each participant. Then, the participants were instructed to watch a short video clip that shows a scene in which a recipient receives a message from a robotFootnote 5 This video was used for introducing the participants to how the robot conveys a message to recipients with a simple example. The video clip was taken from the first-person perspective and simulated human-robot conversations (Table 2). Although the voices of the robots were recorded, the utterances of a recipient were presented using only subtitles (no voice). The length of the video clip was approximately 30 s. Then, the participants were asked about their preference for each messaging option using the questionnaire and then completed the personality questionnaire. Finally, the participants reported their overall comprehension of the survey.

Table 2 Simulated conversations between a human recipient and a robot shown in the video

3.5 Experimental Design and Statistical Analysis

The participants were divided into 2 groups based on their neuroticism traits. The median value obtained from the personality scale divided participants into high/low neuroticism groups (a boundary value was 4.00). The mean value was 4.84 (\(SD = .56\)) in the HN (high neuroticism) group and 3.28 (\(SD = .62\)) in the LN (low neuroticism) group. There was a significant difference between them (\(t = 32.1\), \(p < .001\), \(df = 587\)). The structure of the participants is summarized in Table 3.

Table 3 Participants groups divided by personality traits, type of recipients and topics chosen by them at the time of their recruitment

A three-step analysis was conducted. As the first step, we categorized all messaging options into groups by using an exploratory factor analysis (EFA) to establish the categorization of behaviors that a social mediator robot should possess. After that, we analyzed more closely the participants’ preferences for each messaging option. The aim was to identify messaging options that will yield a statistically equal distribution of participants who prefer or do not prefer these options. For each option, the number of participants who scored more than 4.0 and those who scored less than 4.0 was compared through a chi-squire test (in the comparison, the participants who scored 4.0 (neutral) were excluded). In this way, each option can be classified as a major option in that it was preferred by many of the participants, a minor option in that it was preferred by a few of the participants, or a neutral option in that it did not receive a significant different level of support from the participants. Lastly, relationships between preferred neutral options and the situations were investigated. A \(2 \times 2 \times 2 \) randomized factorial ANOVA was performed on each of the neutral options, with the factors of recipient type, discloser’s gender, and the neuroticism traits of the participants, for each of the four topics.

3.6 Results

3.6.1 Initial Categorization of Messaging Options

EFA using the maximum likelihood method with a promax rotation was performed for all 15 items. The factor analysis revealed three factors with eigenvalues greater than one. In addition, these three factors accounted for 44.7% of the overall variance. Table 4 listed the factor loading of each item loaded onto the three factors (factor loading \(>.40\)), and no cross loading items were foundFootnote 6.

Table 4 EFA (exploratory factor analysis) results

The first factor was labeled requesting-support type. This factor consisted of five messaging options; for example, (Option 4) the robot asks the recipient to help the discloser and does not just deliver the message, or (Option 11) the robot tells the recipient about the discloser’s problem in a way that expresses the robot’s worry (Cronbach’s \(\alpha = .79\)).

The items loaded onto the second factor were related to behaviors of the robot that conceal some part of the self-disclosed contents; thus, we labeled this factor concealing type. This factor consisted of five messaging options; for example, (Option 2) the robot delivers the message to the recipient as a common trouble of elderly people while concealing that the elderly user disclosed his or her troubles, or (Option 3) the robot delivers the message to the recipient without conveying the true emotions of the elderly discloser (Cronbach’s \(\alpha = .74\)).

Items loaded onto the third factor were related to the message formats that contain more nonverbal information of the elderly discloser than is allowed by the mediator robot reading the message. This factor consisted of two messaging options; the robot delivers the message as a voice message (Option 14) or as a video message (Option 15). Thus, we labeled this factor recording type (Cronbach’s \(\alpha = .69\)).

In addition to all of the factors having relatively high or moderate levels of scale reliability, the items of each factor were semantically consistent with each other. We conclude that these items can be used as a valid scale to measure preferences for messaging options.

3.6.2 Three Option Labels (Major, Neutral and Minor Options)

Table 4 shows the chi-square values of each comparison. The number of participants who preferred Options 9 and 11 was greater than that of those who did not. Therefore, these items could be regarded as major options. As there were no significant differences in the number of participants who preferred and who did not prefer Options 1, 2, 4, 5, and 14, these items were regarded as neutral options. However, the number of participants who did not prefer Options 3, 6, 10, 12, and 15 was larger than that who preferred this item; thus, they were classified as minor options.

3.6.3 Differences in the Preferences by Situation

We analyzed how the preference for messaging options classified into neutral options varied according to the situation. Significant (\(p < .05\)) and marginal (\(p < .06\)) results (main effects and interactions) obtained through ANOVA as described in Sect. 3.5 are summarized in Table 5.

Table 5 Significant and marginal results for the participants’ preferences for each messaging option classified as a neutral option. A \(2 \times 2 \times 2 \) ANOVA was performed on each messaging option, with the factors of recipient type, discloser’s gender, and the neuroticism traits of the participants, for each of four topics (RS type, C type, R type denote requesting-support type, concealing type, and recording type, respectively)

In the topic concerning health, messaging options that were included in the requesting-support type showed significant effects. A significant main effect of the discloser’s neuroticism was found for Option 4. The participants with higher neuroticism (HN) traits preferred this messaging option more than those with lower neuroticism (LN) (\(M_{HN}=4.14\), \(M_{LN}=3.67\)). Additionally, in this topic, a marginal interaction between the discloser’s gender and the recipient type was found for Option 5. The female participants preferred to use this messaging option for their friend recipients rather than family recipients (\(t = 2.27\), \(p = .025\), \(df = 143\)). In addition, when a family member was assumed to be the recipient, the female participants’ preference for Option 5 was significantly smaller than that of the male participants (\(t = 2.20\), \(p = .030\), \(df = 143\)) (Fig. 3 left). Therefore, these results suggest that the preference of the female disclosers for Option 5 may be uniquely decreased when messaging a family recipient.

Fig. 3
figure 3

Preference for Options 4 and 5, which were classified into the requesting-support type. Significant interactions were found for the topics concerning health and finance

In the topic concerning finance, significant/marginal interactions between the discloser’s gender and the recipient type were found for every type of messaging option. Regarding the requesting-support type, a significant effect was found for Option 4. The female participants preferred to use this messaging option for family recipients rather than friend recipients (\(t = 2.05\), \(p = .042\), \(df = 131\)). In addition, when the recipient was a friend, the female participants’ preference for Option 4 was significantly smaller than that of male participants (\(t = 1.93\), \(p = .056\), \(df = 131\); Fig. 3 right). Therefore, these results suggest that the preference of the female disclosers for Option 4 may be uniquely decreased when messaging a friend recipient. Regarding the concealing type, the male participants preferred to use Option 1 and 2 for friend recipients rather than family recipients (Option 1: \(t = 1.79\), \(p = .076\), \(df = 131\); Option 2: \(t = 2.30\), \(p = .023\), \(df = 131\); Fig. 4 left). For Option 14 included in the recording type, there was a gender difference, particularly in the case of friend recipients. When the recipient was a friend, the female participants’ preference for Option 14 was significantly smaller than that of male participants (\(t = 2.72\), \(p = .008\), \(df = 131\); Fig. 5 left).

Fig. 4
figure 4

Preference for Options 1 and 2, which were classified into the concealing type. Significant interactions were found in the topics concerning finance, isolation, and reasons to live

In the topic concerning isolation, a significant interaction between the recipient type and the discloser’s neuroticism was found for Option 1, which is included in the concealing type. A post hoc comparison showed that the participants with high neuroticism preferred to use this messaging option for friend recipients rather than family recipients (\(t = 1.71\), \(p = .089\), \(df = 143\); Fig. 4 center). Furthermore, another significant interaction was observed for Option 14, included in the recording type, which involves the factors of the discloser’s gender and the discloser’s neuroticism. More specifically, the male participants with low neuroticism preferred this messaging option more than those with high neuroticism (\(t = 2.36\), \(p = .019\), \(df = 143\)). Additionally, a significant gender difference was found in the participant group with low neuroticism. Among the participants with low neuroticism, male participants preferred Option 14 more than female participants (\(t = 2.75\), \(p = .007\), \(df = 143\); Fig. 5 right). Therefore, the preference of the male disclosers with low neuroticism for Option 14 may be uniquely increased when messaging a family member.

In the topic concerning reasons to live, significant results were found only in the concealing type. A significant main effect of the discloser’s gender was found for Option 2. The male participants preferred this messaging option more than the female participants (\(M_{male}=4.05\), \(M_{female}=3.61\)). Additionally, significant interactions between the recipient type and the discloser’s neuroticism were found for Option 1 and 2, which indicate that the participants with high neuroticism preferred to use these messaging options for family recipients more than for friend recipients (Option 1: \(t = 2.00\), \(p = .047\), \(df = 140\); Option 2: \(t = 2.00\), \(p = .048\), \(df = 140\)). In addition, when the recipients were family members, the participants with high neuroticism preferred to use these messaging options more than the participants with low neuroticism (Option 1: \(t = 2.24\), \(p = .023\), \(df = 140\); Option 2: \(t = 2.43\), \(p = .016\), \(df = 140\); Fig. 4 right). These results suggest that the preference of the disclosers with high neuroticism for Options 1 and 2 may be uniquely increased when messaging a family recipient.

3.7 Discussions

EFA suggested that mediator robots may need to have three different categories of messaging options: requesting-support type, concealing type, and recording type. Additionally, detailed analysis identified that two messaging options (Options 9 and 11) of the requesting-support type were preferred by most of the participants. This may imply that when designing behaviors to request support for recipients, it may be effective to add nonverbal expressions to allow the robot to convey its (Option 9) seriousness and (Option 11) anxiety. For example, another study suggested that the internal weight movements of mediator robots can lead them to be seen as serious, and the reactions of message recipients are influenced by such robot movements [40]. By applying such a novel expression modality to mediator robots, the impact of message delivering could be enhanced.

Fig. 5
figure 5

Preference for Option 14 which was classified into recording type. Significant interactions were found in the topics concerning finance and isolation

On the other hand, detailed options were also detected, and the preferences for them, as expected, turned out to vary with the situation. Overall, the significant effect of recipient type always interacted with some of the discloser’s factors, such as gender or neuroticism. Thus, even when the same recipient type was assumed, the preference for a messaging option may vary depending on such discloser factors. In addition, the disclosers’ preferences for messaging options differed with dialog topic, suggesting that this was also an important factor. For example, in the topic concerning finance, the male disclosers preferred to use messaging options of the concealing type for friend recipients. This finding is consistent with the implication of the preliminary survey (found in Appendix B), in which the male disclosers tended to feel strong anxiety when they talked about their financial issues with their friends. Thus, the messaging options were possibly chosen by the elderly disclosers to avoid their negative feelings in specific situations of self-disclosure.

However, the obtained results were based on an online survey. To determine the effectiveness of the messaging options in this study, verification with an empirical HRI experiment, particularly focusing on the effects on human anxiety, is needed. Therefore, in the next section, we investigate whether interactions with a robot that allows elderly users to choose messaging options can reduce their anxiety in self-disclosure.

4 Study 2: Empirical Study

An HRI study was carried out to test the feasibility of the messaging options designed in the previous section. More specifically, elderly participants (\(N=36\)) were asked to assume a situation in which they self-disclose their loss experience to a family member living in a remote place by interacting with either of the following two mediator robots.

  • Mediator robot 1: When it delivers a message, it repeats the original message contents of the elderly discloser by using its voice (conventional mediator robot).

  • Mediator robot 2: When it delivers a message, it applies the preferred messaging option of the elderly discloser. The neutral messaging options developed in Study 1 were implemented in this mediator robot.

However, human factors such as fear of negative evaluation (FNE) [41] influence the anxiety of human disclosers. It is known that people with a stronger FNE are likely to be more nervous than those with a lower FNE, particularly when they do not know how to make good impressions on others [7]. Therefore, when analyzing the effects of robot behaviors, hierarchical linear regression analysis was applied to statistical control for human factors such as the following:

  • Anxiety felt in self-disclosure over the phone.

  • Fear of negative evaluation.

  • Personality traits of the disclosers/recipients based on the Big Five.

  • The genders of the discloser and recipients.

4.1 Robot Implementation

4.1.1 Robot Platform

The platform used in this study was the Type-B robot introduced in [1], which is a table-top dialog robot specifically developed to study elderly individuals’ self-disclosures (Fig. 6). The robot (185 mm (H), 155 mm (W), and 130 mm (D)) was designed with a simple appearance to minimize the influence of the image and the prior knowledge of commercial robots. It can exhibit physical expressions via arm and/or head movements while verbally communicating with the elderly user by controlling the servo motorsFootnote 7. For this study, a modification was provided to the actuators of the robot. More specifically, an actuator in each joint was replaced with a Kondo KRS-3302 ICS servo motor from SG92R hobby servo motors to achieve smoother and silent body movements. These servo motors are controlled by an Arduino microcontroller, which is connected to a laptop running Python scripts. The other hardware specifications are the same as those described in [1].

Fig. 6
figure 6

The appearance of the robot used in Study 2. This robot was originally developed in a previous study [1]

4.1.2 Dialog System Implementation

A Python script is run on the laptop, which is connected to the robot to provide dialog interactions. For implementation, we used the Google speech-to-text API for speech recognition. For speech synthesis, a text-to-speech API (VoiceText Web API) provided by HOYA was used. By combining these dialog functions with physical expressions (such as nodding), the robot could provide audiovisual feedback in the human message creation process (in this study, a nodding expression was presented for each pause in the user’s speech).

The interaction with the robot is initiated by the key phrase “send a message”, which is input by the user. In response to the user’s command, the robot provides the following utterance: “We shall send a message to RECIPIENT. Please tell me the message you want to send. When you finished the input, please say input finalized.” Then, the user can input any message to the robot until he or she says “input finalized”, which is the key phrase for closing the message creation process. After the message input is finished, the robot will try to interact with the user in two ways described as follows.

Mediator robot 1 is designed to behave similarly to conventional mediator robots. The robot repeats the message from the discloser exactly as being input (i.e., Option 13, defined in Sect. 3.1). When the user finalizes the input of the message, the robot makes the following speech response: “I will close the message input. Then, I will tell RECIPIENT in this way. DISCLOSER said MESSAGE.” As described in Sect. 3.1, DISCLOSER and RECIPIENT are replaced by the name of the elderly discloser and the message recipient, respectively. In the experiment, the experimenter asked each participant the names to user for each party, and they were used in the robot speech. In addition, MESSAGE is replaced by the elderly’s original message to the robot.

For mediator robot 2, the messaging options developed in Study 1 were implemented. More specifically, five messaging options classified as neutral options in the analysis performed in section 3.6.2 were used. In addition, the conventional behavior of the mediator robot (Option 13) was also implemented. In summary, mediator robot 2 has a total of six messaging options (Options 1, 2, 4, 5, 13, and 14). Each of the elderly users can select the messaging option that is preferred for him or her, followed by the instructions from the robot, such as “How should I convey this message to RECIPIENT? Please choose the desired number.” When the elderly discloser selects a messaging option, a number representing each messaging option was visually presented on an iPad put aside the robot (Fig. 7).

4.2 Participants

The participants for this study were recruited through municipal employment service centers for older people in Tokyo, Japan. We obtained 36 participants over 65 years of age (19 males, 17 females; age: \(M = 72.9\) years, \(SD = 3.84\)). The study was a between-participants design, and the participants were divided into two groups based on each robot condition (mediator robot 1: \(N=16\); mediator robot 2: \(N=20\)).

The target profiles of the participants were as follows: (1) persons who had communicated by phone or face-to-face with their own children at least once in the past three years; (2) persons who were living alone or with only their spouses; and (3) persons who had troubles regarding one of the loss experience topics (health, financial, isolation, and reasons to live)Footnote 8. Each participant was paid 5,000 JPY (90 min) for their cooperation. The study protocol was approved by the Research Ethics Committee in the Faculty of Engineering, Information, and Systems at the University of Tsukuba (2015R109-5), and all participants provided informed consent. The fieldwork period was from October 18th to December 15th, 2021.

4.3 Measurement

4.3.1 Anxiety

To measure the anxiety state of the participants when they self-disclosed their loss experiences, twenty question items included in the State-Trait Anxiety Inventory (STAI) were usedFootnote 9: e.g., “I feel stress.” And “I feel nervous.” A 4-point Likert scale was used (in this study, a Japanese version of the STAI [43] was used).

Fig. 7
figure 7

Experiment setup. Elderly participants sat in a chair in front of the robot, and the experimenter sat in a chair behind the robot. An iPad next to the robot displayed the robot’s dialog script. Messaging options were also displayed when mediator robot 2 was used

4.3.2 Fear of Negative Evaluation

Fear of negative evaluation (FNE) from other people has been considered to be closely related to social anxiety [7]. To measure FNE, Watson and Friend created a psychological scale [41], which is frequently used in the social anxiety literature (e.g., [44]). In this study, a short Japanese version of the scale was used [45]. It contains 12 question items (e.g., I often worry that other people may notice my shortcomings, I am always concerned about the impression I am making on other people, etc.). A 5-point Likert scale ranging from “1: strongly disagree” to “5: strongly agree” was used in answering the question items.

4.3.3 Big Five Personality Traits

A discloser’s personality traits may influence his or her perceptions of anxiety when making a self-disclosure [46, 47]. Thus, each of the participants was asked about their self-reported personality traits based on the ten-item personality inventory (TIPI-J [48]). This scale contains two items to measure each personality trait in the Big Five personality model (extroversion, neuroticism, agreeableness, consciousness, and openness). A seven-point scale ranging from “1: strongly disagree” to “7: strongly agree” was used in answering these question items. Furthermore, since the personality traits of the recipient may also affect the anxiety perceived by the discloser, each participant was asked to report the personality traits of the recipient by using this scale.

4.4 Procedure

The study was conducted under face-to-face conditions between an experimenter and each participant in the environment shown in Fig. 7. First, the participants were given a questionnaire by the experimenter, and after filling out the survey on their age and gender, they answered the questions about their FNE and personality traits. Next, the participants were asked to recall the family member in a younger generation who lives in a remote place and communicates with them the most as the recipient of the self-disclosure. The family recipients selected by the participants included daughters (63.9%), sons (25.0%), and sons’ wives (11.1%). 39 percent of the family recipients were the same gender as the disclosers. They had an average age of 42.6 years (\(SD=6.21\)). To have a better awareness of the recipients during the study, all participants were asked to complete forms with the initials of the recipients. After that, the participants were asked to answer questions regarding the personality traits of the recipient.

After the recipients were defined, the participants were asked about the troubles to would like to disclose. Each participant chose one of the topics of loss experiences concerning health, finance, isolation, or reasons to live that they had experienced. In addition, they were also asked what impression they wanted to give to the recipients when they self-disclose. More specifically, they were asked the following questions: (1) How do you want to be perceived by the recipient? and (2) What kind of help do you expect from the recipient? These questions were designed to facilitate each participant’s motivation for self-presentation.

Next, a specific scenario in which they disclosed their loss experiences to the recipient directly over the phone was given to the participants. The anxiety state of the participants, which will be perceived in such a situation, was measured by using the STAI.

Then, each participant was asked to watch videos to learn how about interacting with the mediator robot (e.g., about how each message is sent via the mediator robot or how it delivers the message to the recipient). The experimenter also supplemented the instructions by using Microsoft PowerPoint slides to help the participants understand. The participants assigned to the condition in which they interact with mediator robot 1 were instructed that “this robot repeats to the family recipient exactly what you have input to it”, and none of the other messaging options were provided. On the other hand, all of the messaging options and details of how to switch these options were provided to the participants assigned to the condition using mediator robot 2. When introducing mediator robot 2, the experimenter said “this robot has 6 patterns of behavior for delivering your message. You can choose the one that you like the most. This robot will deliver the message to the recipient according to your choice.” In addition to giving such instructions, each participant had an opportunity to practice using the mediator robotFootnote 10

After measuring the participants’ anxiety immediately prior to making a self-disclosure to the mediator robot, the participants interacted with the robot. While the participants were talking to the robot, the experimenter waited outside the room. The entire interaction procedure with the robot was recorded by a video camera set up behind the participant.

4.5 Statistical Analysis

To compare the effects of the two mediator robots on the anxiety perceived in self-disclosure, a hierarchical linear regression analysis was applied. This method is a special form of a multiple linear regression analysis in which variables are added to the model in separate steps called “blocks.” This is often done to statistically control for certain variables [49]. In this study, a model was built using three blocks, as summarized in Fig. 8. In the first block, a variable representing the type of robot was added. This variable was entered by the forced entry method. In the second block, two control variables that were expected to have strong effects on the dependent variable were included: anxiety perceived in making a self-disclosure over the phone (measured by STAI) and FNE. These variables were also entered into the model by the forced entry method. In the third block, several explanatory variables were added via the stepwise method. The personality traits of both participants and recipients and the gender factor of both participants and recipients were included in this block.

Fig. 8
figure 8

Configuration of the hierarchical linear regression analysis applied in this study. Each variable was entered into the regression model along with the other variables in the corresponding block

4.6 Results

Table 6 summarizes the results. As we expected, certain human factors had significant effects on the perceived anxiety in making a self-disclosure to the mediator robot. In particular, perceived anxiety in self-disclosure over the phone and FNE had significant positive coefficients in every regression model. The larger these values were, the greater the anxiety felt in making a self-disclosure to the mediator robot. In addition, significant regression coefficients on the personality traits of the discloser were calculated for model 3, model 4, and model 5. Specifically, negative coefficients were observed for agreeableness and openness, and a positive coefficient was observed for neuroticism. By statistically controlling for the influence of these human factors, we observed significant regression coefficients for robot type (in model 4 and model 5). This result suggests that the participants who used mediator robot 2 reported less anxiety than those who used mediator robot 1 (model 4: \(t=-2.36\), \(p=0.025\); model 5: \(t=-2.36\), \(p=0.026\)). On the other hand, none of the other variables (e.g., personality traits of the recipients and gender factors of the disclosers/recipients) had significant coefficients.

Table 6 Significant factors that influenced the anxiety of participants when they self-disclosed to mediator robots.

5 Discussion and Limitations

A mediator robot that delivers messages to family members in a way preferred by the elderly disclosers was shown to have the potential to suppress their anxiety about self-disclosure on loss experiences. Moreover, this study successfully demonstrated the effectiveness of the messaging options developed in Study 1. Self-disclosure in one-to-one human-robot interactions has been studied in HRI (e.g., [50,51,52]); however, few studies have been conducted on social robots that act as third-party mediators for human messaging. In the case of the social mediator robot discussed in this study, there is always a person behind the robot, and the user communicates with that person while being influenced by the (mediator) robot’s personality and interaction capabilities. In fact, it has been revealed that the design of the mediator robot’s personality influences the self-disclosures of elderly people [1]. The present study provided further experimental evidence suggesting that the behavior of robots when delivering a message may also have an important impact on the self-disclosures of the elderly. Therefore, this research contributes to the HRI field by presenting insight into the unique configuration of HRIs, which is different from previous works that studied simple one-to-one HRIs.

On the other hand, the effects of the robot behavior were subtle and could be observed only after controlling for the effects of several human factors. For example, the participants who reported strong anxiety when self-disclosing directly to a recipient over the phone also perceived strong anxiety when using a mediator robot. In addition, the discloser’s temperamental characteristics regarding FNE and neuroticism traits were also factors that contributed to the anxiety of the elderly participants. In contrast, the discloser’s agreeableness and openness traits were factors that suppress anxiety about self-disclosing to a robot. Meanwhile, agreeableness and openness traits have been shown to have weak relationships with social anxiety in a psychological study [46], thus the significant negative regression coefficients observed in our models may represent individual differences in affinity for robots. Agreeableness reflects the extent to which someone is cooperative and friendly [53]. Although few studies have reported the effects of this human trait on HRI, Bernotat and Eyssel found that high agreeableness predicts a positive evaluation of interactions with an intelligent robotic apartment, while the other traits did not [54]. Similarly, our study results may also suggest that elderly participants with higher agreeableness are more receptive to the robot and did not feel high anxiety. The openness trait represents the degree to which an individual is imaginative, curious, and broadminded [55]. This may suggest that elderly participants with higher openness are more willing to interact with robots without feeling anxiety.

However, we should note some limitations of this study. This paper does not present knowledge on the mechanism of anxiety reduction in humans when using a social mediator robot. In the future, it is necessary to examine the mechanism of anxiety reduction while using a mediator robot while using methods such as path analysis. At the same time, it is necessary to improve the accuracy in measuring human anxiety by incorporating the physiological indices commonly used in anxiety measurements, such as GSR and salivary cortisol. Additionally, more complicated human factors, such as the interaction among multiple factors, were not considered. There may be other human factors that influence human anxiety, and these structures might be complex. Additionally, our experiment aimed to verify the impact of the various robot behaviors, and all measurements were conducted in a relatively short time window. However, long-term trust development may be a key element for the use of mediator robots. Further exploration is needed to pursue this aspect of mediator robots. Another limitation is that the participants in this study were relatively healthy and active people who regularly communicated with their family members (at least once in three years). Although it is quite difficult to recruit only a certain number of truly isolated elderly people, empirical studies of them are also important. Last, the results obtained from this study may not be supported when using robots with different specifications (e.g., appearance and/or vocal features). Future work is needed to test our interaction model in other robot platforms, such as those with more human-like appearances.

6 Conclusion

To design effective social mediator robots that encourage elderly people to self-disclose to prevent them from being socially isolated, we investigated the behaviors of mediator robots, particularly the way messages are conveyed to recipients. Messaging options were designed through large-scale explorative online surveys (\(N=528\)) and were then implemented in a robot prototype. In an HRI study (\(N=36\)), two interaction models of social mediator robots were compared. The results showed that the elderly disclosers felt less anxiety when they could apply their preferred messaging option than when they could not. We conclude that the ability of mediator robots to deliver messages can affect the user’s interpersonal feelings (such as anxiety) and that the messaging options designed through this study were effective.