A robot counseling system — What kinds of topics do we prefer to disclose to robots?

Full text

A Robot Counseling System
–What kinds of topics do we prefer to disclose to robots?–
Takahisa Uchida1,2, Hideyuki Takahashi1,2, Midori Ban3, Jiro Shimaya1,2,
Yuichiro Yoshikawa1,2and Hiroshi Ishiguro1,2
Abstract— Our research goal was to develop a robot coun-
seling system. It is important for a counselor to promote self-
disclosure of clients to reduce their anxiety feelings. However,
when a counselor is human, clients sometimes hesitate to
disclose intrusive topics due to embarrassment and self-esteem
issues. We hypothesized that a robot counselor, on account
of its unique kind of agency, could remove mental barriers
between the counselor and the client, and promote in-depth self-
disclosure about negative topics. In this study, we prepared two
robots (an android and a desktop robot) as robot counselors.
First, we confirmed that subjects eagerly self-disclosed to
these prepared robots from the numbers of spoken words
about self-disclosure in preliminary experiment. And next, we
conducted the experiment to verify whether it is possible to
expose more of subjects’ weakness to robots than humans. The
experimental result suggested that robots can draw out subjects’
self-disclosure about negative topics than the human counselor.
I. INTRODUCTION
The goal of this research was to construct a counseling
system using robots. In counseling, counselors are required
to reduce clients’ anxiety and stress by interacting with them.
In order to promote anxiety and stress reduction in clients
through a dialogue, encouraging clients’ self-disclosure can
be an effective strategy. Self-disclosure is an act of notifying
others of one’s personal information [1]. Jourard insisted
that self-disclosure was a sign of personality health [1].
Since then various studies have found that there is a correla-
tion between indicators of mental health and self-disclosure
(e.g., [2]). Furthermore, it is reported that depression and
physical symptoms are reduced ([3], [4], [5]) by discussing
with others matters related to negative events. In other words,
it can be said that encouraging self-disclosure about negative
events greatly contributes in reducing anxiety and stress
among clients.
However, clients do not always aggressively practice self-
disclosure of negative topics. For example, it is pointed out
that disclosing personal events related to negative topics is
usually considered as lowering one’s relative position with
This work was supported by JST Ishiguro ERATO Project
1Takahisa Uchida, Hideyuki Takahashi, Jiro Shimaya,
Yuichiro Yoshikawa and Hiroshi Ishiguro are with Graduate
School of Engineering Science, Osaka University, Osaka, Japan
uchida.takahisa@irl.sys.es.osaka-u.ac.jp,
takahashi@irl.sys.es.osaka-u.ac.jp,
shimaya.jiro@irl.sys.es.osaka-u.ac.jp,
yoshikawa@irl.sys.es.osaka-u.ac.jp and
ishiguro@irl.sys.es.osaka-u.ac.jp
2JST ERATO
3Midori Ban is with Doshisya University Faculty of Psychology, Kyoto,
Japan ban@ams.eng.osaka-u.ac.jp
(a) ERICA (b) CommU
Fig. 1. Robots
respect to the partner through exposure of his/her weakness
by self-disclosing negative content [6]. Things concerning
such negative topics cannot necessarily be uttered positively
in consideration of their own facial aspects.
In this study, therefore, we consider to use a robot to solve
this dilemma. In the case of human counselor, especially in
the first meeting, clients are likely to avoid the disclosure of
negative content for reasons such as avoiding for lowering
their relative position. On the other hand, if the counselor is
a robot, the client’s impression that the robots are linked to
his/her own society, that is, the perception of the closeness
of relationship of the robot with the client, may be extremely
small. For these reasons, we hypothesized that by adopting a
robot as a counselor, clients may be able to expose their own
weaknesses with less hesitation than a human counselor; and
thus, the robot counselor could effectively encourage clients’
self-disclosure of negative topics.
In this study, we conducted two experiments. In the first
experiment (preliminary investigation), we quantified the
amount of self-discourse to human, robots, and a loudspeaker
respectively by counting clients’ spoken words to confirm the
validity of prepared two counseling robots. And in the second
experiment (main investigation), we verified the hypothesis
that it is possible to expose more of self’s weakness to robots
than humans. As robot consolers , we prepared two different
types of consoling robots, a female android (Fig. 1(a)) and a
cute small child-like robot (fig. 1(b)). And then, we verified
whether the difference of consoling robots affected preferred
topics of client’s self-disclosure. To investigate the relation
between the kind of self-disclosure and the kind of agent
we used questionnaires about variety kind of self-discloure
2017 26th IEEE International Symposium on
Robot and Human Interactive Communication (RO-MAN)
Lisbon, Portugal, Aug 28 - Sept 1, 2017.
978-1-5386-3517-9/17/$31.00 ©2017 IEEE 207

topics and about agents’ impressions. Finally, based on
these experiments, we discussed what kinds of aspects work
effectively for constructing a counseling system using robots.
II. PRE LIM INARY EXPERIMENT
This experiment’s aim was to verify whether the subject
actually self-discloses to some extent to the robots we used.
A. Condition
We prepared four conditions; human condition, android
condition, small robot condition, and sound only condition.
In these experiments, five people (three male, two female,
average age 19.2) participated in human condition, five
people (four male, one female, average age 19.2) participated
in the android condition, six people (three male, three female,
average age 22.2) participated in the small robot condition,
and four people (one male, three female, average age 18.3)
participated in the sound only condition. Each agent asked
some questions and then the subjects answered the questions.
In the android condition, we used ERICA [8] (Fig. 1(a)).
As ERICA speaks, it moves its lips, head, and torso in
synchrony with the prosodic features of the voice. The lip,
head, and torso movements of ERICA are automatically
generated from its voice (using the systems developed by [9]
and [10]). In the small robot condition, we used CommU
(Fig. 1(b)). This is a desktop robot (about 30 cm) with a
cute little body that resembles a child. It is also equipped
with speakers in the chest, and it opens and closes the mouth
while it utters words. We used a loudspeaker in sound only
conditions.
With respect to voice of agents except for the human,
we used speech synthesis softwares. In the android only
and sound only conditions, we used VOICE TEXT ERICA
of HOYA CORPORATION1to utter words. Moreover, in
the small robot condition, we used AITalk Chihiro2to utter
words.
B. Procedure
At first, the subjects faced one of the four agents across the
table, as shown in Fig. 2, and after an appropriate greeting,
the agent asked five questions about subjects’ self-disclosure.
In this experiment, we adopted the inter-subject design. From
the question items created by Niwa [11] for Japanese people,
two items from the question group (Level I) on hobby, that is
low self-disclosure difficulty level, and three from the group
on the subject’s negative character and ability (level IV),
that is high disclosure difficulty, were selected. The agents
informed the subject that he/she could refuse to answer the
questions anytime. The question items are shown below.
1http://voicetext.jp/
2http://www.ai-j.jp/
Fig. 2. 4 conditions
0
50
100
150
200
250
300
350
400
Human ERICA CommU Sound only
Number of characters
Fig. 3. Length of Self-disclosure
Level I: Hobby
•Favorite things
•How to spend your holiday
•Events that you are looking for
•Experiences which you enjoyed lately
•What you are crazy about recently
•Hobby
•What would you like to do as a hobby
Level IV: Negative personality and ability
•Experiences wherein you hurt somebody
•Personality that you dislike
•Experiences that you hated because of your
disability
•Sickness about your ability
•Experience wherein the goal could not be
achieved because of lack the capacity
•Inferiority with ability
•Disappointed experience because of your limited
ability
C. Result
Fig. 3 shows the amount of utterances (the number of
characters) that the subject made when answering to each
agent’s questions about self-disclosure. The error bars show
the standard error of the mean.
208

D. Discussion
In the figure 3, the human condition, the android condition,
the small robot condition, and the sound only condition are
presented in the order of the amount of speech uttered by
the subjects. Moreover, the speech quantity in the sound only
condition was particularly small. From these results, it can
be said that robots can draw out certain self-disclosure as
compared to humans.
III. EXPERIMENT2
The purpose of this experiment was to verify the hypoth-
esis that clients tend to expose more their own weak points
to the robot than human counselors, and robots can draw out
clients’ self-disclosure about negative topics. In Experiment
2, we evaluated only the human, the android, and the small
robot condition, eliminating the sound only condition, based
on the results of the preliminary experiment.
A. Condition
Fifteen subjects (seven male, eight female, average age
20.5) participated in this experiment. In the human condition,
the android condition, and the small robot condition, we used
the same agents and the same systems and speech synthesis
of each robot as in the preliminary experiment.
B. Procedure
At first, the subjects faced the human, the android, and
the small robot as shown in Fig. 4, and then were offered
greetings in order. We randomized the position of the agents
and the order of the greeting. The script is described below.
Human: Nice to meet you. My name is Midori. Thank
you very much.
Android: Nice to meet you. My name is ERICA.
Thank you very much.
Small Robot: Nice to meet you. My name is CommU.
Thank you very much.
Thereafter, the subject was asked to go to a separate
room and respond to the questionnaire. Questionnaires were
divided into two types. In the first type, the subjects had
to answer about their impressions of each agent, and in the
second type, the subjects had to select with which agent they
wanted to talk about each given 45 items of self-disclosure.
Fig. 4. Scene of experiment2
Self
Self
Self Other
Other
Other
1
4
7
Fig. 5. IOS
４
7
1Leader
Follower
Fig. 6. Leader-Follower Scale
We also measured the impression of the agents to inves-
tigate the influence of the felt closeness or relationship with
each agent on the attitude for self disclosure. Namely, we
used IOS scale (The Inclusion of Other in the Self scale) [12],
Leader-Follower scale and Agency scale [13]. IOS scale
is a measure of relationship closeness between self and
other (Fig.5). Leader-Follower scale is a measure of leader-
follower relationship between a subject and an agent. As
shown in Fig. 6, illustration is depicted at the center of each
illustration, and it expresses what kind of positional relation
exists between the subject and each agent. For example, in
Fig. 6, number four means equivalent positional relationship,
number one means the positional relationship wherein agent
is in front of the subject, and the number seven means
the positional relationship wherein the agent is behind the
subject. The Agency scale is an impression evaluation scale
(Fig. 7) that can map a target agent on two axes, intellectual
level and emotional level, separately. From the Agancy scale,
we can know what kind of agency the agent has.
We used 45 topics listed in the Enomoto Self-Disclosure
Questionnaire-45 (ESDQ-45) [14] as the self-disclosure top-
ics, which are representative self-disclosure contents for
Japanese people. For each topic, the subjects were asked to
select one preferrable agent from three agents (the human,
the android, and the small robot) as the interlocutor to be
disclosed it. Examples of the 45 topic items are shown below.
209

Fig. 7. Agency Scale
0
1
2
3
4
5
6
IOS Follower
IOS & Follower Scale
Human Android CommU
Fig. 8. Result of IOS Scale and Leader-Follower Scale
Studying interests, Current goal, Things about living
and fulfillment, Hobby, Efforts to improve appear-
ance appeal, Badly injured experiences, Troubles in
friendship, Emptiness and anxiety in life, Loneliness
and alienation, Complaints and dissatisfaction with
society, Opinions on literature and art, Professional
appropriateness, Sports sense, Things to ask a friend,
Past love affairs experience.
C. Result
First, the results of three impression evaluation question-
naires (IOS, Leader - Follower scale, Agency scale) are
shown in Fig. 8, Fig. 9.
In addition, 10 subjects (two male, eight female, average
age 20.4), who were different from those of Experiment
2, evaluated the positive degree of each topic (Enomoto’s
45 topics) [14]. We categorized top 15 topics as positive
topics, middle 15 as neutral topics, and bottom 15 as negative
topics. Fig. 10 shows the selected ratio of each agent for the
disclosure of the negative topics, the neutral topics, and the
Fig. 9. Result of Agency Scale
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Negative Neutral Positive
Selected Proportion of Each Agent
Human Android CommU
Fig. 10. Category of Positive/Neutral/Negative
positive topics.
D. Discussion
Regarding the impression evaluation for each agent, as
shown in Fig. 8, the distance of the relationship with each
agent felt by the subjects with respect to the IOS scale was
the nearest with the human, and the distance between the
android and the small robot was almost of the same degree.
In addition, as shown in the Leader-Follower scale, the
subject and the android had the same level of relationship, the
human had a slightly higher degree of Leader, the CommU
had a higher degree of Follower. Also, looking at Fig. 9, we
observed that the Android had a low emotion score but high
intelligence score, while human and small robots condition
had high emotional but low intelligence scores.
Next, when we analyzed how many agents were selected
as counselors (Fig.10) after classifying self-disclosure items
into three categories of positive, negative, and neutral, the
positive topics had a large percentage of human selections,
and the proportion of robots increased as the topics became
neutral-negative topics. Therefore, in order to investigate
210

0
0.1
0.2
0.3
0.4
0.5
0.6
Negative Neutral Positive
Mean
Prefered Category in Robots
**p<.01
**p<.01
Fig. 11. Prefered Category in Robots
which topics (positive, neutral, negative topics of self-
disclosure) in the robot condition (the android condition and
the small robot condition) were preferred, we carried out
analysis of variance (ANOVA). Significant difference was
confirmed at the significance level p= 0.05, thus sub effect
tests was conducted by using the Ryan method. The result
is shown in Fig. 11. It can be confirmed that there was a
tendency to self-disclose negative and neutral contents to the
robot.
Furthermore, regarding self-disclosure of negative topics,
the results of the categories of emotional aspects classified
in advance by ESDQ-45 [14] (Fig. 12) revealed that the
small robot had the highest score for ”experiences badly
hurt the heart” and ”points that seem to be emotionally
immature,” while the android had the highest score for
”jealous experience.” Regarding the self-disclosure items on
negative topics, the items on the emotional aspect showed
that the robot could draw self-disclosure more effectively
than the human. A possible reason could be the distance
between the robot and the subject, as indicated by the robots’
IOS score and the degree of the Follower score. Therefore,
an impression that they are downgraded than him/her may
be able to draw out self-disclosure on his emotional, that is,
weakness aspect.
E. Future Work
Finally, we will now discuss the limitations of this re-
search. In the experiments, we compared the human, the
android, and the small robot assuming the scene of the
first meeting. In other words, the evaluation related to self-
disclosure was only based on the first impression. In order
to build a robot counseling system, it is important to take
into consideration not only the first impression but also what
kinds of dialogue the robots narrated. In the future, we would
like to conduct an advance research on what kind of dialogue
the robot should utter as a counselor.
In addition, the result of this experiment is insufficient to
conclude what kinds of differences in characteristics exited
between the android and the small robot. The result of the
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Selected Proportion of Each Agent
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
Emotionally injured
experience
Emotional immaturity Jealous experience
Emotional Aspect
Human Android CommU
Fig. 12. Emotional Aspect
Agency scale (Fig. 9) revealed that the intelligent index was
higher for the android than for the small robot, and the
emotion index was higher for the small robot than for the
android, but the relationship between self-disclosure items
was not clear. In the future, we would also like to conduct
a research on the differences in robot characteristics.
Furthermore, in this experiment, we adopted a female
experimenter as a human agent. Therefore, the result of the
human condition in this experiment may be peculiar to this
experiment only. Going forward, it is necessary to explore
other conditions even in human condition (for example,
include men as counselors).
Finally, in the experiment 2, the subjects were required to
choose one agent to disclose something about each topic.
Thus, it cannot be declared whether the subject actively
wanted to talk or not. Therefore, in future experiments, it
is necessary to evaluate not only which agent the subjects
want to talk to but also the willingness of the subject to
discuss the topics.
IV. CONCLUSION
In this paper, with the aim of introducing a robot for
counseling, we conducted experiments with the human, the
androids, and the small robot as a counselor and examined
the extent of study participants’ self-disclosure about subjects
and topics. As a result, when using the robots as a counselor,
the possibility of extracting self-disclosure amount similar
to human counselors was shown. In addition, as a result of
investigating by self-disclosure items, a higher tendency to
disclose negative topics to the robots than the human was
confirmed, especially with respect to items concerning emo-
tions. In the future, we would like to conduct a survey that
takes into consideration the subjects’ willingness to discuss
and the importance of self-disclosed topics, and clarify what
kinds of differences are likely to appear depending on the
type of robots.
211

ACKNOWLEDGMENT
This research was supported by the Japan Science and
Technology Agency, ERATO ISHIGURO Symbiotic Human-
Robot Interaction Project.
REF ERE NCES
[1] S. M. Jourard, “Self-disclosure: An experimental analysis of the
transparent self.” 1971.
[2] P. C. Cozby, “Self-disclosure: a literature review.” Psychological
bulletin, vol. 79, no. 2, p. 73, 1973.
[3] S. Cohen and T. A. Wills, “Stress, social support, and the buffering
hypothesis.” Psychological bulletin, vol. 98, no. 2, p. 310, 1985.
[4] J. W. Pennebaker and S. K. Beall, “Confronting a traumatic event: to-
ward an understanding of inhibition and disease.” Journal of abnormal
psychology, vol. 95, no. 3, p. 274, 1986.
[5] R. L. Silver and C. B. Wortman, “Coping with undesirable life events,”
Human helplessness: Theory and applications, vol. 279, p. 375, 1980.
[6] E. Hatfield, “The dangers of intimacy,” Communication, intimacy, and
close relationships, pp. 207–220, 1984.
[7] L. J. Wood, K. Dautenhahn, A. Rainer, B. Robins, H. Lehmann, and
D. S. Syrdal, “Robot-mediated interviews-how effective is a humanoid
robot as a tool for interviewing young children?” PloS one, vol. 8,
no. 3, p. e59448, 2013.
[8] D. F. Glas, T. Minato, C. T. Ishi, T. Kawahara, and H. Ishiguro, “Erica:
The erato intelligent conversational android,” in Robot and Human
Interactive Communication (RO-MAN), 2016 25th IEEE International
Symposium on. IEEE, 2016, pp. 22–29.
[9] C. T. Ishi, C. Liu, H. Ishiguro, and N. Hagita, “Evaluation of formant-
based lip motion generation in tele-operated humanoid robots,” in
Proc. of the IEEE/RSJ International Conference on Intelligent Robots
and Systems, 2012, pp. 2377–2382.
[10] K. Sakai, T. Minato, C. T. Ishi, and H. Ishiguro, “Speech driven
trunk motion generating system based on physical constraint,” in Robot
and Human Interactive Communication (RO-MAN), 2016 25th IEEE
International Symposium on. IEEE, 2016, pp. 232–239.
[11] S. Niwa and S. Maruno, “Development of a Scale to Assess the Depth
of Self-disclosure,” Personality Research, vol. 18, no. 3, pp. 196–209,
2010. (in Japanese)
[12] A. Aron, E. N. Aron, and D. Smollan, “Inclusion of other in the
self scale and the structure of interpersonal closeness.” Journal of
personality and social psychology, vol. 63, no. 4, p. 596, 1992.
[13] H. Takahashi, M. Ban, and M. Asada, “Semantic differential scale
method can reveal multi-dimensional aspects of mind perception,”
Frontiers in Psychology, vol. 7, 2016.
[14] Hiromoto Enomoto, “Psychological study of self-disclosure, Kitaoji
Shoboh,” 1997. (in Japanese)
212