6670 IEEE ROBOTICS AND AUTOMATION LETTERS, VOL. 6, NO. 4, OCTOBER 2021 Using an Android Robot to Improve Social Connectedness by Sharing Recent Experiences of Group Members in Human-Robot Conversations Changzeng Fu , Chaoran Liu, Carlos Toshinori Ishi , Yuichiro Yoshikawa , Takamasa Iio, and Hiroshi Ishiguro Abstract—Social connectedness is vital for developing group cohesion and strengthening belongingness. However, with the accelerating pace of modern life, people have fewer opportunities to participate in group-building activities. Furthermore, owing to the teleworking and quarantine requirements necessitated by the Covid-19 pandemic, the social connectedness of group members may become weak. To address this issue, in this study, we used an android robot to conduct daily conversations, and as an intermediary to increase intra-group connectedness. Specifically, we constructed an android robot system for collecting and sharing recent member-related experiences. The system has a chatbot function based on BERT and a memory function with a neuralnetwork-based dialog action analysis model. We conducted a 3-day human-robot conversation experiment to verify the effectiveness of the proposed system. The results of a questionnaire-based evaluation and empirical analysis demonstrate that the proposed system can increase the familiarity and closeness of group members. This suggests that the proposed method is useful for enhancing social connectedness. Moreover, it can improve the closeness of the userrobot relation, as well as the performance of robots in conducting conversations with people. Index Terms—Long-term interaction, social HRI. Manuscript received February 24, 2021; accepted June 15, 2021. Date of publication July 7, 2021; date of current version July 20, 2021. This work was supported in part by JSPS, under Grants JP20H05576, 19H05691, 20H00101, and in part by JST, Moonshot R&D under Grant JPMJMS2011. This letter was recommended for publication by Associate Editor S. Rossi and Editor G. Venture upon evaluation of the reviewers’ comments. (Corresponding author: Changzeng Fu.) Changzeng Fu is with the Graduate School of Engineering Science, Osaka University, Osaka 560-8531, Japan, and also with the Interactive Robot Research Team, Guardian Robot Project, RIKEN Kyoto 619-0288, Japan (e-mail: changzeng.fu@irl.sys.es.osaka-u.ac.jp). Yuichiro Yoshikawa is with the Graduate School of Engineering Science, Osaka University, Osaka 560-8531, Japan (e-mail: yoshikawa@irl.sys.es.osakau.ac.jp). Chaoran Liu is with the Advanced Telecommunications Research Institute International Kyoto 619-0288, Japan (e-mail: chaoran.liu@atr.jp). Carlos Toshinori Ishi is with the Interactive Robot Research Team, Guardian Robot Project, RIKEN Kyoto 619-0288, Japan, and also with the Advanced Telecommunications Research Institute International Kyoto 619-0288, Japan (e-mail: carlos@atr.jp). Takamasa Iio is with the Faculty of Culture and Information Science, Doshisha University, Kyoto 610-0394, Japan (e-mail: tiio@mail.doshisha.ac.jp). Hiroshi Ishiguro is with the Graduate School of Engineering Science, Osaka University, Osaka 560-8531, Japan, and also with the Advanced Telecommunications Research Institute International Kyoto 619-0288, Japan (e-mail: ishiguro@sys.es.osaka-u.ac.jp). This letter has supplementary downloadable material available at https://doi. org/10.1109/LRA.2021.3094779, provided by the authors. Digital Object Identifier 10.1109/LRA.2021.3094779 I. INTRODUCTION OCIAL connectedness among group members is an important factor that affects group performance and determines the strength of group cohesion and the sense of belongingness [1], [2]. Good social connectedness is the basis for strong cohesion, which can bring many benefits to the group, such as goal commitment, efficient teamwork, smooth cooperation between members, and mutual understanding [3]–[5]. According to social psychology studies, collective events and perceived similarity are key factors for fostering connectedness [6]–[8]. However, with the accelerating pace of modern life, collective events for team building are barely available, as people have fewer opportunities to participate in group building activities. Furthermore, owing to the teleworking and quarantine requirements necessitated by the Covid-19 pandemic, the social connectedness of group members may become weak. Messaging applications may alleviate this and enhance communication among group members to some extent. However, such user-touser messaging occasionally add a certain social burden (e.g., replying may be awkward because of the low familiarity of the members during a chat) so that some people would be afraid of such direct message-based interaction [9], [10]. In this case, a robot serving as an intermediary to aid users in establishing a connection would be an effective approach [11]. Berger et al. [12] claimed that similarity-based attraction could produce robust interpersonal relationships. Moreover, according to McPherson et al. [13], the homophily principle reinforces people’s social relationships, which could be interpreted as interpersonal similarities that breed connections. This trend could be evidently seen on social media. People frequently assemble because of shared similarities, even if they are geographically separated and have never met before [14], [15]. Once the interpersonal similarities are discovered, the development of a connected strength relationship [16] as well as the group cohesiveness [17] are fostered. In Addition, Kaptein et al. [18] extended the similarity-attraction effect with computer-mediated communication. Their study suggests that computer-mediated interaction has the potential to improve social connectedness between people when there is an opportunity to identify similarities. Therefore, with the unavailability of collective events, we consider improving intra-group connectedness by discovering similarity among members (perceived similarity). To this end, we propose using an S 2377-3766 © 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See https://www.ieee.org/publications/rights/index.html for more information. Authorized licensed use limited to: Universita degli Studi di Bologna. Downloaded on August 10,2021 at 18:08:07 UTC from IEEE Xplore. Restrictions apply. FU et al.: USING AN ANDROID ROBOT TO IMPROVE SOCIAL CONNECTEDNESS BY SHARING RECENT EXPERIENCES android robot that conducts daily conversations, provides users with a sense of talking with people, and serves as an intermediary to increase social connectedness among group members. Accordingly, we should first develop interactive strategies for engaging people in human-robot conversations. Several studies have confirmed that presenting memory and learning abilities during an interaction can increase the willingness to talk with robots [19]–[22]. Moreover, the results reported by Fu et al. [23] suggest that remarking memorized user preferences in human-robot conversations can improve user engagement and has the potential to maintain long-term interaction. Following the same strategy, Uchida et al. [24] further verified that sharing preferences in group talk can tighten intra-group relationships. However, topics limited to preferences may not be sufficient to improve social connectedness, whereas discussions about offering more information on daily activities [25] as well as experience sharing [26] would be particularly helpful. Following the aforementioned directions, in this study, we developed a system using a neural-network-based model that enables a robot to automatically recognize and memorize recent member-related experiences during human-robot conversations, and share them in future interactions. Additionally, we conducted an experiment involving conversation continuation trials on three different days to verify the effectiveness of the proposed system. The contributions of this study are as follows: r By integrating neural-network-based models [27], [28], we develop a system whereby robots can automatically collect memories from the real world and share them in humanrobot conversations. r We propose an interactive strategy whereby an android robot (ERICA) can improve its performance and enhance intra-group connectedness. II. RELATED WORD The principal component of this study is robot memory. Herein, we review related work and summarize the main differences with the present study. Memory elicitation has been demonstrated to be effective in human-robot conversations in several respects, such as commonground establishment, intelligence [24], [30]–[32], personalization [33], [34], and long-term interaction enhancement [32], [34]. Kanda et al. [30], [31] designed a shop guide robot that remembers frequent consumers and gives different greetings based on the number of times it has met the customer. It was demonstrated that the proposed system could build rapport between users and the robot, promote the willingness of people to interact with the robot, and improve the evaluation of perceived intelligence. Leite et al. [32] proposed a method that allows robots to use information from prior conversations with the same child so that a sense of relationship can be fostered over time. It was verified that this method could improve perceived intelligence, increase friendliness, and facilitate engagement. Lee et al. [33] developed Snackbot, which can remember and share the snack choice and service usage of a user. It was demonstrated that this interaction strategy could reinforce the rapport, cooperation, and engagement of a person with the robot. Ahmad 6671 et al. [34] proposed a system for designing robot behavior based on past user performance on mathematics tests and emotion. Their experiment results demonstrated that a robot equipped with their proposed system could enhance the engagement and the learning performance of children. Although some studies have been concerned with the effectiveness of using robot memory in human-robot interaction, most of them have focused on enhancing the relationship between the robot and the current user by considering the memory of the robot, which can be understood as one-to-one rapport building. Uchida et al. [24] conducted a study on sharing individual preferences (perceived similarity) in a one-to-many group talk to enhance intra-group relationships; however, group talk is a collective event that allows people to communicate (e.g., verbally and non-verbally). Such behaviors could also be factors that affect these relationships. Compared with previous memory-related studies, the present study is aimed at developing intra-group connections by sharing recent member experiences stored as robot memories. These shared memories are supposed to be stimuli that enhance the perceived similarity of group members. It should be noted that no collective events are involved in this study; that is, we attempt to enhance one-to-many connectedness through one-to-one conversations between a user and the robot. III. METHOD A. System The core function of the proposed system is to enable robots to automatically detect people-related information (feelings and experiences) from human-robot conversations and store them as memories. To construct the people-related information detector, we first prepared a spoken Japanese dataset with two-way Japanese conversations (I-JAS [35]), from which we selected interview dialogs of native Japanese speakers as samples for model training. Subsequently, we annotated sentences (utterance level) with 13 labels by referring the method proposed by Stolke et al. [36] with some modifications. Table I lists the labels, definitions, and the corresponding samples. In this study, we regard “subjective information,” “objective information,” and “plan” as people-related information (experience) that can be shared in human-robot conversations. However, the annotated dataset was not sufficiently large to train a model with a large number of parameters, and our task requires a text classification model that does not have an excessively large number of parameters and is effective on a small-sample-size dataset. Accordingly, we adopted the SeMemNN model proposed by Fu et al. [27], which has been proved to have high training speed and fair performance when trained on a small-sample-size dataset. After SeMemNN was trained on the annotated dataset, the testing accuracy of detecting people-related information was 82.57%, which is reasonable for practical implementation. In addition to detecting people-related information, other categories, such as questions and agreement, are also used in the system. Moreover, for a dialog system, a chatbot (response generation) function is essential for providing replies based on user utterances. We trained the BERT [28] model on a Japanese Authorized licensed use limited to: Universita degli Studi di Bologna. Downloaded on August 10,2021 at 18:08:07 UTC from IEEE Xplore. Restrictions apply. 6672 IEEE ROBOTICS AND AUTOMATION LETTERS, VOL. 6, NO. 4, OCTOBER 2021 TABLE I TRAINING SAMPLES FOR THE PEOPLE-RELATED INFORMATION DETECTOR (JPS-DAPRINFO [29]) study), then it is considered that there is little people-related information shared in the current conversation. To promote an informative dialog, the system selects a question (e.g., do you exercise these days?) from the pre-defined question database so that the exchange of information may be facilitated. B. Verification Strategy Fig. 1. Flowchart of the proposed dialogue system. The bold blocks indicate the core functions of the system. The dashed lines indicate that a function refers to the connected database. conversation corpus [37] and integrated a question-answering function based on predefined stereotype information into an android robot. We now present the working flow of the proposed system. As shown in Fig. 1, the robot starts with a greeting and waits for the user to reply. Subsequently, by using the SeMeNN model, the system analyzes the utterance from the user and determines whether it is people-related information. In that case, the system stores this utterance in the memory database using the who and what components (e.g., Chason / went to Tokyo last week.). Then, the system calculates the similarity between the utterance and the memories stored in the database based on the semantic cosine distance and word overlapping rate by using the NLTK [38] and Word2Vec [39] toolkits, matching the most similar information. Subsequently, the system generates the robot response based on the matched memory by using a rule randomly selected from the template database (e.g., I talked to [who] a couple days ago, he/she told me [what]). Then, the system determines whether the conversation should be ended based on the round number (max round was set to 25 in this study). If the utterance of the user is not people-related, the system calls the chatbot function for responding. Specifically, if the system continually uses the chatbot function for N consecutive times (N equals 3 in this Fig. 2 shows the verification strategy. The main purpose of this study is to verify whether the proposed system can enhance social connectedness among group members. We invited subjects who knew each other to participate in conversation continuation trials conducted on three different days. In addition, the performance of the robot in conducting conversations should be evaluated. In order that the robot present learning as well as memory capabilities, the system collects people-related information in each conversation, and shares this information in future conversations. It should be noted that at the beginning of the experiment, the robot had no memories, and thus only the chatbot function was used on the first day. On the second and third days, the robot shared the collected people-related information. With such a design, each day’s trails could regard historical trails as a baseline for comparing the members’ connectedness as well as the performance of the robot. Considering the possibility of the mere-exposure effect [40], we designed a control condition (the right part of Fig. 2) to verify whether or not the results were due to the proposed strategy. The configuration under the control condition was essentially identical to that under the experimental condition, but only the chatbot function was employed in all trials. Thus, there was no people-related information shared in conversations. IV. EXPERIMENT A. Subject We invited 36 subjects from six different groups to participate in the experiment. The profiles for each group and the assignment of subjects to the experimental and control groups are presented Authorized licensed use limited to: Universita degli Studi di Bologna. Downloaded on August 10,2021 at 18:08:07 UTC from IEEE Xplore. Restrictions apply. FU et al.: USING AN ANDROID ROBOT TO IMPROVE SOCIAL CONNECTEDNESS BY SHARING RECENT EXPERIENCES 6673 Fig. 2. Experiment design. The left part is the design of the experimental condition, in which only the chatbot function is used on the first day, and people-related information is shared on the next two days. The right part is the design for the control condition, in which only the chatbot function is employed for three days, and no people-related information is shared. TABLE II PARTICIPANTS FOR EACH CONDITION Uchida et al. [24] on a 7-level Likert scale to evaluate factors 3 and 4 (see the following text box). It should be noted that factors 1 and 2 were evaluated before the experiment and after each conversation, whereas the other factors were evaluated after each conversation. - Familiarity of the group members: r I am getting familiar with group members’ recent activities. r I would be able to know better the group members’ mental states/experiences. r I would be able to think of something fun that we could in Table II, where M and F denote “male” and “female,” respectively, whereas EXP. and CON. denote the experimental and control conditions, respectively. We carefully assigned groups to either the experimental or the control condition according occupation, gender, and age. Thus, 18 subjects (M = 9 and F = 9) with an average age of 31.11 (SD = 8.03) participated in the experimental condition, and the remaining 18 subjects (M = 9 F = 9) with an average age of 28.78 (SD = 6.21) participated in the control condition. We adopted a within-subject strategy for each condition with respect to the temporal dimension, and a between-subject strategy with respect to the conditional dimension. That is, the members of a group only engaged in the experimental or the control condition for three separate days. Moreover, the experiment was group-dependent, implying that the sharing of memories was limited to information relevant to the members of a group. B. Measurement and Hypothesis To comprehensively evaluate the effectiveness of the proposed system, we first designed a questionnaire-based evaluation for the factors, that is, 1) closeness with the group, 2) closeness with the robot (ERICA), 3) familiarity of the group members, 4) evaluation on the system/robot. Inspired by Uchida et al. [24], we adopted the inclusion of other in the self scale (IOS) [41] to evaluate factors 1 and 2. We used the questions designed by all do together with group members. - Evaluation on system/robot:. r The robot knows something about my group’s members. r The robot has the ability to have a conversation with people. r The robot has the ability to understand people’s utterances. r I would like to have this robot in the place where my group’s members and I are always around. r Other members of my group would like to interact with this robot. In addition to the questionnaire-based evaluation, we also conducted some empirical observations, that is, 1) the number of times the subjects feel good or bad about the robot’s reply, 2) the total words of the utterances of the subjects, 3) the number of times the subjects share people-related information (experience) in each conversation. To collect the feelings of the subjects, we developed a touch-panel interface (see Fig. 3) with two (good)’ and ‘ (bad)’) to let subjects provide buttons (‘ feedback regarding their feelings during the conversation. The subjects were asked to evaluate the content of the reply of the robot by pressing the “good” button when the participant felt the reply was good, or the “bad” if the reply was considered bad. The buttons could be pressed an unlimited number of times during the conversation. Regarding the total number of words of the utterance of the subject and the number of times that Authorized licensed use limited to: Universita degli Studi di Bologna. Downloaded on August 10,2021 at 18:08:07 UTC from IEEE Xplore. Restrictions apply. 6674 Fig. 3. IEEE ROBOTICS AND AUTOMATION LETTERS, VOL. 6, NO. 4, OCTOBER 2021 Experimental setting. The proposed system was applied to ERICA. people-related information was obtained in each conversation, we set a function in the system to calculate the result. The hypotheses are the following: 1) Sharing similar recent feelings or experiences is supposed to enhance perceived similarity, thus enhancing the social connectedness and familiarity of group members. 2) Compared with a plain chatbot, the proposed system can improve the conversational ability of the robot and the robot-user relationship. C. Apparatus and Procedure Fig. 3 shows the experimental setting. The proposed system was applied to an android robot (ERICA), and a touch panel was placed in front of the subjects so that the replies of the robot could be evaluated (“good” or “bad”), as explained in Measurement and Hypothesis section. Herein, we describe the procedure. Before the experiment, we obtained approval from the ethics committee regarding the involvement of human subjects. The purpose and steps of the experiment were introduced to the subjects, who were asked to fill out a consent form. Before the first-day conversation, we asked the subjects to answer a pre-hoc questionnaire, related only to closeness within the group and closeness to the robot. Subsequently, the human-robot conversation was started, the subjects casually talked to the robot, and they provided evaluation by pushing the corresponding button. After finishing the conversation, the subjects filled out another questionnaire related to all the questionnaire-based measurements. Moreover, to ensure that the good/bad measurement was properly conducted, we reminded of the subjects of this operation before each conversation. After the experiment, we confirmed whether the subjects pressed the corresponding button at the appropriate time. The procedure for the conversations conducted on the second and third days was the same as that on the first day, except that there was no pre-hoc questionnaire. the increase by performing a person-wise subtraction from the scores collected by post-hoc questionnaires each day. Repeated two-way ANOVA of two independent variables (condition and temporal dimension) was conducted on the calculated growth of closeness, the scores collected by post-hoc questionnaire regarding other factors, as well as the empirical observations. Fig. 4 shows the results for each factor. Tables III and IV present the detailed statistical results. Regarding the growth of closeness with the group and the familiarity of group members, significant differences were found in the main and the interaction effect, as well as the multiple comparisons in the experiment condition. It can be concluded that sharing recent experiences better enhances intra-group connectedness and increases familiarity among group members with time than the system equipped with only the plain chatbot. A similar observation can be made regarding the growth of the closeness with the robot, indicating that, compared with the plain chatbot, the proposed system, by sharing recent experiences, considerably enhances human-robot interaction. Regarding the evaluation on the system/robot, the significant differences in the main and the interaction effect suggest that the proposed system performs better in human-robot conversation than the plain chatbot system. Regarding the empirical observation, there are also significant differences. The experimental condition allowed subjects to experience more good feelings about the robot’s reply than the control condition. However, in the multiple comparisons, no significant differences were found between day2 and day3 under the experimental condition. Regarding the number of bad feelings, a significant difference was found only in the temporal dimension. Regarding the number of words and the amount of sharing people-related information, significant differences were found in both the main effect and the interaction effect. These results imply that, as compared to the plain chatbot system, sharing recent experiences can better increase the willingness of subjects to talk with the robot and induce them to share their experiences. VI. DISCUSSION Herein, we use the experimental results to support our hypotheses, and we compare our findings with those of previous memory-related studies on human-robot interaction. In addition, we discuss the limitations of the proposed method and indicate future research directions. A. Verification of Hypotheses V. RESULTS First, we used one-way ANOVA to compare the scores collected from the pre-hoc questionnaire (regarding closeness with the group and closeness with the robot) between the experimental and the control group. The statistical results demonstrated that there was no significant difference in the closeness with the group (F (1, 34) = 0.256, p = 0.616) and the closeness with the robot (F (1, 34) = 0.739, p = 0.396) between the experimental and control groups. Then, we regarded the scores collected before the experiment as initial scores, and calculated Hypothesis 1: The increase in the closeness with the group and the familiarity of group members indicate that, without a collective event, memorizing and sharing similar recent experiences with group members can enhance social connectedness. Moreover, we collected feedback from participants after the experiment. Regarding the experimental condition, we received feedback from nine subjects, and all of them were positive regarding the effect of the proposed system. They felt that the group members became familiar and talked with each other a little bit more because of the people-related information shared by Authorized licensed use limited to: Universita degli Studi di Bologna. Downloaded on August 10,2021 at 18:08:07 UTC from IEEE Xplore. Restrictions apply. FU et al.: USING AN ANDROID ROBOT TO IMPROVE SOCIAL CONNECTEDNESS BY SHARING RECENT EXPERIENCES 6675 Fig. 4. Statistical results for each factor by using repeated two-way ANOVA (∗ p < .05, ∗∗ < .01). The horizontal curly brackets with ‘*’ indicate a significant difference in the temporal dimension (main effect), the vertical curly brackets with ‘*’ indicate a significant difference between conditions (main effect). The lines connecting sub-groups indicate significant difference in multiple comparisons (interaction effect). TABLE III STATISTICAL RESULTS USING TWO-WAY ANOVA TABLE IV STATISTICAL RESULTS OF MULTIPLE COMPARISONS Authorized licensed use limited to: Universita degli Studi di Bologna. Downloaded on August 10,2021 at 18:08:07 UTC from IEEE Xplore. Restrictions apply. 6676 the robot (e.g., ”I got to know other members’ recent feelings or activities through the conversation”, ”Sharing members’ information is a catalyst that triggers dialogue between members”, and ”The robot had positive effect on our group relationship”). Regarding the control condition, we received feedback from seven subjects. They felt the robot had no particular effect (e.g., ”So far, no effect”, ”I don’t think there is any particular effect right now”, and ”I didn’t feel any effect from the conversation”). This also demonstrates the effectiveness of the proposed system in improving social connectedness. Hypothesis 2: The second hypothesis can be verified by the increase in the closeness to the robot and the evaluation of the system/robot. Although the chatbot can increase robot-human closeness in the three-day consecutive interactions, the proposed system improved this score more significantly. The empirical observations also support the second hypothesis to some extent. On the days (day2 and day3 in the experimental condition) of sharing people-related information, users reported a larger number of good feelings. Additionally, the number of words and the amount of sharing people-related information indicate that sharing people-related information with group members can promote the volume of user utterances and the willingness to share more information. These results suggest that sharing people-related information could stimulate human-robot conversations. In addition to using people-related information to engage users in human-robot conversations, by memorizing shared people-related information, the system can find common ground with the user and provide appropriate information in different groups even for the same topic. This suggests that the proposed system has good generalizability for wide-range use. IEEE ROBOTICS AND AUTOMATION LETTERS, VOL. 6, NO. 4, OCTOBER 2021 proposed method is quite effective, some limitations must be acknowledged. Subject diversity: The subjects who participated in the experiment were adults, and more experiments are required to verify whether the proposed method is useful for children. To this end, more factors should be considered in the experimental setting for the children group (e.g., whether the evaluation method in the form of questionnaires can be valid for children, or whether the appearance of the robot scares children). Therefore, in future work, adjustments should be made to the experimental setting that would allow the participation of children subjects. Use of people-related information: In this study, we let the robot share people-related information under a maximum setting (all memory). More effective methods of using information obtained from the current user and other members should be developed to enhance intra-group connectedness. Moreover, the optimal order of the stages of rapport building should be studied (first establishing user-robot connections and then user-member connections, or first showing member-robot connections and then enhancing user-robot and user-member connections). These aspects should be explored in detail in a separate study. In addition, we should consider the effect of the ratio of the current-user memory to member-related memory on member connectedness and bot evaluation. Furthermore, in such an interaction strategy, private information should be carefully handled. More research should be conducted on this issue, and identifiers should be developed so that robots may distinguish how private the information is and choose what to share in human-robot interaction. Such privacy-preserving behavior may enhance the trustworthiness of robots. VII. CONCLUSION B. Findings Comparison We proposed a memory-sharing-based strategy to improve human-robot interaction. The results are consistent with the findings of existing related studies [23], [30]–[32], [34], that is, using memory enhances user engagement and promotes user-robot closeness. However, this is a preliminary study that discusses whether a robot, in the absence of collective events, can improve social connectedness by only taking advantage of the perceived similarity of group members through the elicitation of people-related information. The principle of the rapport-building mechanism of the proposed method is to enhance one-to-many (user-members) connections through only one-to-one human-robot conversations. By contrast, existing memory-related studies focus on one-toone (user-robot) rapport building through one-to-one humanrobot conversations. That is, this study presents the possibility of using robots as a medium to establish multi-person connections by sharing people-related information. We proposed a dialog system that enables an android robot to memorize and share people-related information in humanrobot conversations. Moreover, we conducted a 3-day humanrobot conversation experiment to verify the effectiveness of the proposed system. Compared to the plain chatbot system, the results demonstrated that the proposed system could increase the familiarity and closeness of group members, as well as the performance of the robot in conducting conversations with people. Moreover, sharing people-related information could induce people to engage in conversations with robots and improve intra-group connectedness. In future work, we will validate our approach in a wider age range and investigate the most beneficial memory-sharing strategies for enhancing social connectedness. Furthermore, we will explore the effect of sharing people-related information on the trustworthiness of robots. Finally, we will investigate methods for enabling robots to discern whether a piece of people-related information can be shared or not. REFERENCES C. Limitations and Future Work In this study, we developed a system and designed an interaction strategy based on memory sharing under a maximum setting (all memory) to improve intra-group connectedness. Even though it was experimentally demonstrated that the [1] P. W. Speer, C. B. Jackson, and N. A. Peterson, “The relationship between social cohesion and empowerment: Support and new implications for theory,” Health Educ. Behav., vol. 28, no. 6, pp. 716–732, 2001. [2] R. S. Huckman, B. R. Staats, and D. M. Upton, “Team familiarity, role experience, and performance: Evidence from indian software services,” Manage. Sci., vol. 55, no. 1, pp. 85–100, 2009. Authorized licensed use limited to: Universita degli Studi di Bologna. Downloaded on August 10,2021 at 18:08:07 UTC from IEEE Xplore. Restrictions apply. FU et al.: USING AN ANDROID ROBOT TO IMPROVE SOCIAL CONNECTEDNESS BY SHARING RECENT EXPERIENCES [3] B. Franz, R. Leicht, K. Molenaar, and J. Messner, “Impact of team integration and group cohesion on project delivery performance,” J. Construction Eng. Manage., vol. 143, no. 1, 2017, Art. no. 0 4016088. [4] J. R. Terborg, C. H. Castore, and J. A. DeNinno, “A longitudinal field investigation of the impact of group composition on group performance and cohesion,” J. Personality Social Psychol., vol. 34, no. 5, pp. 782–790, 1976, doi: 10.1037/0022-3514.34.5.782. [5] P. F. Dominey, V. Paléologue, A. K. Pandey, and J. Ventre-Dominey, “Improving quality of life with a narrative companion,” in Proc. 26th IEEE Int. Symp. Robot Hum. Interactive Commun., 2017, pp. 127–134. [6] M. Stieler and C. C. Germelmann, “The ties that bind us: Feelings of social connectedness in socio-emotional experiences,” J. Consum. Marketing, vol. 33, no. 6, pp. 397–407, 2016, doi: 10.1108/JCM-03-2016-1749. [7] W. H. McNeill, Keeping Together in Time. Cambridge, MA, USA: Harvard Univ. Press, 1995. [8] F. Sani, M. Bowe, and M. Herrera, “Perceived collective continuity and social well-being: Exploring the connections,” Eur. J. Social Psychol., vol. 38, no. 2, pp. 365–374, 2008. [9] P. Shiozawa and R. R. Uchida, “Social media during a pandemic: Bridge or burden?” Sao Paulo Med. J., no. AHEAD, 2020. [10] S. Steinert, “Corona and value change. the role of social media and emotional contagion,” Ethics Inf. Technol., pp. 1–10, 2020. [11] J. Shimaya, Y. Yoshikawa, K. Ogawa, and H. Ishiguro, “Robotic question support system to reduce hesitation for face-to-face questions in lectures,” J. Comput. Assist. Learn., vol. 37, no. 3, pp. 621–631, 2021. [12] C. R. Berger, “Task performance and attributional communication as determinants of interpersonal attraction,” Commun. Monographs, vol. 40, no. 4, pp. 280–286, 1973. [13] M. McPherson, L. Smith-Lovin, and J. M. Cook, “Birds of a feather: Homophily in social networks,” Annu. Rev. Sociol., vol. 27, no. 1, pp. 415–444, 2001. [14] I. Guy, M. Jacovi, A. Perer, I. Ronen, and E. Uziel, “Same places, same things, same people? Mining user similarity on social media,” in Proc. ACM Conf. Comput. Supported Cooperative Work, 2010, pp. 41–50. [15] M. A. Javarone and G. Armano, “Perception of similarity: a model for social network dynamics,” J. Phys. A: Math. Theor., vol. 46, no. 45, 2013, Art. no. 455102. [16] R. Reagans, “Close encounters: Analyzing how social similarity and propinquity contribute to strong network connections,” Org. Sci., vol. 22, no. 4, pp. 835–849, 2011. [17] M. A. Hogg, E. A. Hardie, and K. J. Reynolds, “Prototypical similarity, self-categorization, and depersonalized attraction: A perspective on group cohesiveness,” Eur. J. Social Psychol., vol. 25, no. 2, pp. 159–177, 1995. [18] M. Kaptein, D. Castaneda, N. Fernandez, and C. Nass, “Extending the similarity-attraction effect: The effects of when-similarity in computermediated communication,” J. Comput.-Mediated Commun., vol. 19, no. 3, pp. 342–357, 2014. [19] M. M. De Graaf, S. B. Allouch, and T. Klamer, “Sharing a life with harvey: Exploring the acceptance of and relationship-building with a social robot,” Comput. Hum. Behav., vol. 43, pp. 1–14, 2015. [20] H. Mahzoon et al., “Effect of self-representation of interaction history by the robot on perceptions of mind and positive relationship: A case study on a home-use robot,” Adv. Robot., vol. 33, no. 21, pp. 1112–1128, 2019. [21] D. F. Glas, K. Wada, M. Shiomi, T. Kanda, H. Ishiguro, and N. Hagita, “Personal greetings: Personalizing robot utterances based on novelty of observed behavior,” Int. J. Social Robot., vol. 9, no. 2, pp. 181–198, 2017. [22] S. Mitsuno, Y. Yoshikawa, and H. Ishiguro, “Robot-on-robot gossiping to improve sense of human-robot conversation,” in Proc. 29th IEEE Int. Conf. Robot Hum. Interactive Commun., pp. 653–658. 6677 [23] C. Fu, Y. Yoshikawa, T. Iio, and H. Ishiguro, “Sharing experiences to help a robot present its mind and sociability,” Int. J. Social Robot., vol. 13, pp. 341–352, 2021, doi: 10.1007/s12369-020-00643-y. [24] T. Uchida, H. Ishiguro, and P. F. Dominey, “Improving quality of life with a narrative robot companion: Ii-creating group cohesion via shared narrative experience,” in Proc. 29th IEEE Int. Conf. Robot Hum. Interactive Commun., 2020, pp. 906–913. [25] K. Jeong et al., “Fribo: A social networking robot for increasing social connectedness through sharing daily home activities from living noise data,” in Proc. ACM/IEEE Int. Conf. Hum.-Robot Interact., 2018, pp. 114– 122. [26] T. Chen, J. Drennan, and L. Andrews, “Experience sharing,” J. Marketing Manage., vol. 28, no. 13-14, pp. 1535–1552, 2012. [27] C. Fu, C. Liu, C. Ishi, Y. Yoshikawa, and H. Ishiguro, “Sememnn: A semantic matrix-based memory neural network for text classification,” in Proc. IEEE 14th Int. Conf. Semantic Comput., 2020, pp. 123–127. [28] J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “Bert: Pre-training of deep bidirectional transformers for language understanding,” 2018, arXiv:1810.04805. [29] C. Fu, C. Liu, C. T. Ishi, and H. Ishiguro, “JPS-daprinfo: A dataset for japanese dialog act analysis and people-related information detection,” Mar. 2021. [Online]. Available: https://doi.org/10.5281/zenodo.4590253 [30] T. Kanda, R. Sato, N. Saiwaki, and H. Ishiguro, “Friendly social robot that understands human’s friendly relationships,” in Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst. (IEEE Cat. No 04CH37566), vol. 3, 2004, pp. 2215– 2222. [31] T. Kanda, M. Shiomi, Z. Miyashita, H. Ishiguro, and N. Hagita, “A communication robot in a shopping mall,” IEEE Trans. Robot., vol. 26, no. 5, pp. 897–913, Oct. 2010. [32] I. Leite, A. Pereira, and J. F. Lehman, “Persistent memory in repeated childrobot conversations,” in Proc. Conf. Interact. Des. Child., 2017, pp. 238– 247. [33] M. K. Lee, J. Forlizzi, S. Kiesler, P. Rybski, J. Antanitis, and S. Savetsila, “Personalization in hri: A longitudinal field experiment,” in Proc. 7th ACM/IEEE Int. Conf. Hum.-Robot Interact., 2012, pp. 319–326. [34] M. I. Ahmad and O. Mubin, “Emotion and memory model to promote mathematics learning-an exploratory long-term study,” in Proc. 6th Int. Conf. Hum.-Agent Interact., 2018, pp. 214–221. [35] K. Sakoda and Y. Hosoi, “International corpus of japanese as a second language (i-jas),” Nat. Inst. Japanese Lang. Linguistics, vol. 6, no. 3, pp. 93–110, 2016-03. [36] A. Stolcke et al., “Dialogue act modeling for automatic tagging and recognition of conversational speech,” Comput. Linguistics, vol. 26, no. 3, pp. 339–373, 2000. [37] I. Fujimura, S. Chiba, and M. Ohso, “Lexical and grammatical features of spoken and written japanese in contrast: Exploring a lexical profiling approach to comparing spoken and written corpora,” in Proc. VIIth GSCP Int. Conf. Speech Corpora, 2012, pp. 393–398. [38] S. Bird, “NLTK: the natural language toolkit,” in Proc. COLING/ACL 2006 Interactive Presentation Sessions, Jul. 2006, pp. 69–72. [39] Y. Goldberg and O. Levy, “word2vec explained: Deriving mikolov et al.’s negative-sampling word-embedding method,” Advances Neural Inf. Process. Syst., vol. 27, 2014, pp. 2177–2185. [40] R. F. Bornstein and P. R. D’agostino, “Stimulus recognition and the mere exposure effect,” J. Pers. Social Psychol., vol. 63, no. 4, pp. 545–552, 1992. [41] K. M. Woosnam, “The inclusion of other in the self (ios) scale,” Ann. Tourism Res., vol. 37, no. 3, pp. 857–860, 2010. Authorized licensed use limited to: Universita degli Studi di Bologna. Downloaded on August 10,2021 at 18:08:07 UTC from IEEE Xplore. Restrictions apply.