Personality Traits Assessment using P.A.D. Emotional Space in

Human-robot Interaction

Zuhair Zafar, Ashita Ashok and Karsten Berns

Robotics Research Lab, Department of Computer Science,

Technische Universit

at Kaiserslautern, Germany

Keywords:

Personality Trait Assessment, PAD Emotional Space, Robot Adaptivity, Human-Robot Interaction.

Abstract:

Cognitive social robotics is the ﬁeld of research that is committed to building social robots that facilitate to

draw parallels with human beings. Humans assess the behavior and personality of their counterparts to adapt

their behavior and show empathy to ﬂourish human-human interaction. Similarly, assessment of human per-

sonality is highly critical in realizing natural and intelligent human-robot interaction. Numerous personality

traits assessment systems have been reported in the literature; however, most of them target the big ﬁve per-

sonality traits. From only visual information, this work proposes to use pleasure, arousal, and dominance

emotional space for the assessment of personality traits based on the work of Mehrabian. To validate the sys-

tem, three different scenarios have been developed to assess 12 different personality traits on a social humanoid

robot. Experimental results show that the system can assess human personality traits with 84% accuracy in

real-time and, hence, it can adapt its behavior according to the perceived personality of the interaction partner.

1 INTRODUCTION

With the technological advent and constant research

in the ﬁeld of robotics, it is now quite practical to ac-

knowledge the actuality of social robots being a part

of human’s daily life in the next decades. It neces-

sitates and inspires the motivation of creating robots

that can perceive the various learnings of life similar

to humans, especially in a real-world Human-Robot

Interaction (HRI). Concerning HRI, the basic expec-

tations from a social robot are to perceive words, emo-

tions, behaviors, and so on, in order to draw several

conclusions and informed decisions for realizing nat-

ural HRI. Henceforth, assessment of human person-

ality traits is essential to bring a sense of appeal and

acceptance towards the robot during the interaction.

Personality plays a vital role in Human-Human In-

teraction (HHI) as it guides the conversation towards

a level of satisfaction and comfort for humans. Ac-

cording to psychologists, human behavior is known

to be a combination of verbal cues together with non-

verbal cues. Nonverbal cues such as temperamen-

tal characteristics are known to be innate in human

beings, sometimes existing subtly or visibly. These

temperaments aggregate into traits, e.g., extrovert or

introvert, throughout human life through daily expe-

riences. The signiﬁcance of personality in HHI can

be better exempliﬁed by two renowned theories from

the ﬁeld of human psychology, i.e., the chameleon

effect (Chartrand and Bargh, 1999) and similarity/at-

traction theory (Henderson and Furnham, 1982). The

chameleon effect explains the non-conscious human

tendency to passively mimic the behavior of one’s in-

teraction partner in a social environment.

In contrast, the similarity/attraction theory empha-

sizes that humans are generally attracted to and prefer

the company of others who maintain morals and at-

titudes similar to their own. For example, it is quite

often observed that there exists a sense of shared per-

sonality among friends than among random pairs of

strangers. People tend to change their behavior ac-

cording to their interlocutor’s behavior. If he/she is

talkative and expressive, one also tends to be more

expressive. Therefore, assessment of human person-

ality is highly crucial for a robotic system in order to

interact and adapt naturally for intelligent HRI.

The primary goal of this work is the assessment

of human personality traits in the real-world using

the temperament framework presented by Mehra-

bian (Mehrabian, 1996). The author has exploited the

pleasure, arousal, and dominance emotional space to

describe and measure individual differences on dif-

ferent personality scales. This work formulates a

hypothesis that using P.A.D. emotional space, com-

puted through human nonverbal cues, can provide

a successful assessment of human personality traits

Zafar, Z., Ashok, A. and Berns, K.

Personality Traits Assessment using P.A.D. Emotional Space in Human-robot Interaction.

DOI: 10.5220/0010161801110118

In Proceedings of the 16th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2021) - Volume 2: HUCAPP, pages

111-118

ISBN: 978-989-758-488-6

111

in the context of HRI. To realize the hypothesis,

P.A.D. dimensions are computed using nonverbal

cues, namely, human posture, head gesture, hand ges-

ture, proximity, body activity, and facial expression.

These P.A.D. values are then used in different person-

ality trait equations devised by Mehrabian (Mehra-

bian, 1996) for assessment of personality traits.

2 LITERATURE SURVEY

There are many complex models available in psy-

chology to recognize the personality traits of humans.

The most signiﬁcant theory that describes personality

traits in ﬁve dimensions is commonly known as the

big ﬁve (BF) personality theory or ﬁve-factor model.

Costa and McCrae (Costa Jr and McCrae, 1976) have

introduced the BF model, which is based on ﬁve di-

mensions, namely, extroversion, agreeableness, neu-

roticism, conscientiousness, and openness to new ex-

periences. Each dimension is a continuum, and a low

value in the dimension represents the opposite char-

acteristics of that dimension.

Some researchers have tried to map the cognitive

emotional models directly from humans to a robot, for

example (Rodi

c et al., 2016). However, conventional

approach is to use personality models from psychol-

ogy. The BF model has been used extensively in the

literature of personality traits assessment. In the case

of automatic personality detection from nonverbal be-

havioral cues, authors (Batrinca et al., 2011) have ap-

plied automatic detection of the BF personality traits

in scenarios of self-presentation and employment in-

terviews. Different features, 17 visual, 3 speech time,

and 9 acoustic cues have been used for classiﬁcation.

The visual nonverbal cues include eye-gaze, frowning

emotion, hand movements, head orientation, mouth

ﬁdgeting, and posture.

In their further work, authors (Batrinca et al.,

2012) have explored to detect the BF personality traits

in the human-computer interaction scenario using the

map task. The work aims to recognize the BF per-

sonality traits in a collaborative task setting. Features

used for classiﬁcation of personality traits are acous-

tic features, e.g., duration of speech, pitch, intensity,

and so on; visual features, e.g., motion vector mag-

nitude over skin computed by using discrete cosine

transform, and additional features such as the number

of speaking turns by the subject. Authors have used

support vector machines (SVMs) for the classiﬁcation

task. The results report an accuracy of greater than or

equal to 70 percent for emotional stability, extrover-

sion, and conscientiousness scales. Although authors

have argued having detected the BF traits accurately

except for open to new experiences trait, a question

that remains unanswered is how the lack of features,

for instance, the distance between the speakers, detec-

tion of facial expression, and many more have been

compensated.

Interesting research on the automatic analysis

of engagement in HRI is observed in the ﬁndings

by (Salam et al., 2016). This work aims to judge the

impact of personality traits of human participants on

the engagement with robots. Among the three phases

of analysis, the ﬁrst phase consisted of data collection

in the HRI triadic scenario, while the second phase in-

cluded extraction of individual and interpersonal fea-

tures based upon nonverbal cues from human par-

ticipants. Individual features include a histogram of

gradients (HOG), a histogram of optical ﬂow (HOF),

body activity, joint speed, motion features, etc. In-

terpersonal features include the visual focus of atten-

tion, the global quantity of movement, relative orien-

tation, and distance between the participants, and rel-

ative orientation with respect to the robot. The ﬁnal

phase included a prediction of the level of engage-

ment in two types of engagement, namely individ-

ual and group engagement, based upon the predicted

the BF personality traits and the features extracted in

phase 2. They have concluded that the prediction of

engagement using personality traits reports better re-

sults as compared to when personality trait informa-

tion is not used.

In the context of HRI, authors (Zafar et al., 2018a)

have presented a humanoid robot that can recognize

3 dimensions of the BF model in real-time. The

authors have discussed the importance of nonverbal

cues in automatic personality recognition. Nonverbal

features such as human postures, facial expressions,

body activity, head gestures, and proximity have been

used. SVMs are used for the classiﬁcation task. Au-

thors have claimed above 90% recognition rate for 3

dimensions, namely, extroversion, agreeableness, and

neuroticism.

The works mentioned above for personality as-

sessment are directly based on either visual features

or linguistic features or both of them. They also are

focused mainly on BF personality traits. However,

Mehrabian has presented a general framework for

describing and measuring individual temperaments

of personality that also covers more traits, such as

shyness, anxiety, and aggression (Mehrabian, 1996).

This framework is based on pleasure, arousal, and

dominance emotional space (Russell and Mehrabian,

1977). After extensive research, authors have deﬁned

these three domains as follows. Pleasure can be deter-

mined using cognitive judgments of evaluation, i.e.,

higher evaluations of stimuli associated with greater

HUCAPP 2021 - 5th International Conference on Human Computer Interaction Theory and Applications

112

pleasure induced by stimuli. Arousal corresponds to

judgments of high-low stimulus activity using mea-

sure of stimulus “information rate”. Dominance is

deﬁned as judgment of stimulus potency, with more

signiﬁcant the inﬂuence of stimuli corresponding to

lower values of dominance.

To the best of our knowledge, there exists no re-

search that has implemented the framework (Mehra-

bian, 1996) on human personality traits concerning

the adaptivity and behavior of social robots in HRI.

Our work is the ﬁrst to implement this psychology

framework for human personality traits assessment

using nonverbal cues in the context of HRI.

3 PERSONALITY TRAITS

ASSESSMENT

Although a limited number of technical systems have

been reported in the literature for real-time personal-

ity traits assessment, these systems at best can only

recognize the BF personality traits. They are unable

to distinguish between subtle personality traits, for

example, shyness and introversion or dominance and

aggression. According to (Watson and Clark, 1997),

extroversion can be subdivided into the more speciﬁc

facets of assertiveness, gregariousness, cheerfulness,

and energy. Similarly, neuroticism can be subdivided

into loneliness, anxiety, and sensitivity to rejection,

while shyness is the part of introversion trait. As

mentioned in the previous section, this work employs

the framework (Mehrabian, 1996) that uses pleasure,

arousal, and dominance (P.A.D.) emotional space for

personality traits assessment. In the following sub-

section, P.A.D. emotional space is deﬁned.

3.1 P.A.D. Emotional Space

In literature, human emotions are often deﬁned in

multiple dimensional spaces. However, the deﬁnition

of emotion varies for each researcher, who adopted

one or more dimensions to deﬁne it. For example, ac-

cording to (Wundt and Judd, 1897), the three dimen-

sions of emotions are namely, “pleasurable vs. un-

pleasurable”, “arousing vs. subduing” and “strain vs.

relaxation”. Many emotional spaces have been pre-

sented in psychology. Among them, the prominent

ones are the circumplex model by Russell (Russell,

1980) and the Positive Activation-Negative Activa-

tion (PANA) model (Watson and Tellegen, 1985).

There exists another renowned three-dimensional

emotional space, called Pleasure-Arousal-Dominance

emotional space (Russell and Mehrabian, 1977). The

P.A.D. model aims to describe and measure emotional

traits that correspond to human personality. The three

dimensions are deﬁned to be bipolar such that plea-

sure is described as a continuum that ranges from in-

tense pain or unhappiness on one end to intense hap-

piness or ecstasy on the other. Arousal has been re-

ported to range from sleepiness and drowsiness to a

high level of alertness and excitement. Dominance

varies from emotions of a complete absence of con-

trol or impact over events to feeling inﬂuential and in

control of the situation at the opposite extreme. In the

following sections, a methodology that uses nonver-

bal cues for the implementation of P.A.D. emotional

space is presented.

3.1.1 Pleasure

As previously mentioned, the value on the pleasure

scale describes how much the event is enjoyable for

a person. The study on facial expressions conducted

by (Boukricha et al., 2009) shows that pleasure is di-

rectly associated with facial expressions. If a person

is happy, the value on the pleasure scale is high. Sim-

ilarly, if a person is unhappy and exhibits facial ex-

pressions such as sadness, fear, anger, or disgust, then

the value on the pleasure scale is low.

Algorithm 1: Estimation of Pleasure Value.

pleasant = 0, unpleasant = 0

n ← number of frames in 10 seconds

for i ← 0 to n do

if Human.Face.Exist() then

expression

← Current.Expression

if P

expression

= happy OR surprise then

pleasant = pleasant +1

else

unpleasant = unpleasant +1

if pleasant ≥ unpleasant then

P ← pleasant

else

P ← unpleasant × (−1)

pleasure =

× P // average pleasure

return pleasure

To estimate the pleasure value from facial ex-

pressions, a system developed by authors (Al-Darraji

et al., 2017) has been used to recognize six basic fa-

cial expressions in real-time. The facial expressions,

extracted in every frame, are standardized and aver-

aged over a 10 second period. Happiness and surprise

expressions contribute towards the positive value of

pleasure scale, while sad, fear, disgust, and anger con-

tribute towards the negative value of pleasure scale.

Personality Traits Assessment using P.A.D. Emotional Space in Human-robot Interaction

113

Algorithm 1 shows the estimation of pleasure.

3.1.2 Arousal

The value on the arousal scale describes how much

the event is exciting and thrilling for a person.

The arousal can be assessed by the combination of

two nonverbal features, namely proximity and body

movements. According to (Nass and Lee, 2001), peo-

ple, when aroused, show frequent body movements.

Similarly, (Hirth et al., 2011) have established the re-

lationship between the proximity of a person from the

interlocutor and the arousal. Arousal of a person is

considered high if he/she moves towards or stands

close to the robot. If a person moves away or stands

farther from the robot, the arousal value goes down.

To calculate arousal value, a weighted sum of prox-

imity and body movements is used. In order to esti-

mate the proximity value, the concept of interpersonal

distances of humans during human-human interaction

has been used. According to (Hall, 1963), interper-

sonal distances of a person can be categorized into

four zones, namely intimate space, personal space, so-

cial space, and public space. If the robot is in the pub-

lic space of a person, the proximity value is negative

−1. If the robot is in the intimate space of a person,

the proximity value increases up to +1.

To detect human body movement during interac-

tion, skeleton joints positions of the upper body are

used, which are provided by the NiTE middleware li-

brary and analyze them over time. Activity is detected

if the change in values is exceeded from a threshold.

The proximity of a person with regards to a robot is

determined by using the depth information of tracked

humans. Equation 1 shows the weighted summation

of activity and proximity to estimate arousal.

Arousal =

∑

i=0

× P +W

× A)

= 0.8

= 0.2



4 P > 0.5m

= [0.6 − 0.8]

= [0.4 − 0.2]



0.2m < 4P ≤ 0.5m (1)

= 0.6

= 0.4



4 P ≤ 0.2m

In Equation 1, W

and W

are proximity and ac-

tivity weights, respectively. The weights are dynamic

and change according to the change in proximity, 4P,

of a person. If a person moves more than half a meter

during the interaction, the W

gets the higher value

to depict this sudden change on the arousal dimen-

sion. Moreover, the proximity value (P) is a continu-

ous value between −1 to +1 and depends on how far

the person is standing from the robot.

3.1.3 Dominance

In the HRI scenario, dominance can be estimated by

analyzing human behavior over time. As mentioned

by (Jensen, 2016), conﬁdent and dominant people

generally have a wide-open trunk during interactions,

which shows that they are approachable to others and

keeps them in a more open-minded attitude. Simi-

larly, threatening postures such as feet spread apart

with hands-on-hips posture and pointing postures are

correlated with aggression and dominance. Dominant

people are also physically active during interactions.

In contrast, submissive people tend to look down with

slumped body postures. Submissiveness is correlated

with self-touching postures, such as cross arms pos-

ture or thinking postures (Argyle, 1988). Submissive

people also avoid mutual eye gaze (Argyle, 1988).

They generally are passive during interactions.

The system developed in (Zafar et al., 2018b) has

been used to recognize different human postures us-

Algorithm 2: Estimation of Dominance Value.

dominant = 0, submissive = 0

n ← number of frames in 10 seconds

for i ← 0 to n do

if Human.Body.Exist() then

posture

← Current.Posture

head gesture

← Current.Head Gesture

body movements

←

Current.Body Movement

if (P

posture

= O.P OR P.P OR A.S) AND

head gesture

= L.A OR L.U) AND

body movements

= true) then

dominant = dominant + 1

else if (P

posture

= C.P OR T.P) AND

head gesture

= L.D OR L.A OR L.L

OR L.R) AND

body movements

= f alse) then

submissive = submissive + 1

else

do nothing

if dominant ≥ submissive then

D ← dominant

else

D ← submissive × (−1)

dominance =

× D // average dominance

return dominance

HUCAPP 2021 - 5th International Conference on Human Computer Interaction Theory and Applications

114

Figure 1: Schematic ﬂow of personality traits assessment using P.A.D. emotional space from ﬁndings of (Mehrabian, 1996).

ing human skeleton joint angles, e.g., pointing pos-

ture, thinking posture, crossed arms posture, open

arms posture, aggressive posture and so on. To rec-

ognize head gestures, the system developed in (Saleh

and Berns, 2015) has been used. Head gestures such

as head nodding, head shaking, look left, look down,

look up, look right, look ahead, etc. are considered in

this study. Body movements of a person are detected

using the same method described in Arousal imple-

mentation. Algorithm 2 shows the estimation of the

dominance dimension.

In Algorithm 2, O.P = open posture, P.P = point-

ing posture, A.S = Aggressive stance posture,

C.P = crossed arms posture, T.P = thinking posture,

L.A = looking ahead gesture, L.U = looking up ges-

ture, L.D = look down gesture, L.L = look left gesture

and L.R = look right gesture.

3.2 Personality Trait Assessment

System

This work proposes to use the P.A.D. emotional space

for the assessment of human personality traits using

the Mehrabian’s framework (Mehrabian, 1996). Us-

ing the three dimensions, pleasure, arousal, and dom-

inance, he has formulated 59 individual measures that

correspond to human personality traits. It has been

demonstrated that traits are symmetrically related to

one another based upon the P.A.D. dimensions.

Although the formulated traits are of a wide range,

only 12 out of 59 traits are realized in this work.

These traits are chosen according to the experimen-

tal restrictions and based on the knowledge of non-

verbal cues associated with them. Personality traits

such as mysticism, loneliness, and anorexic require

either verbal or contextual information or both for an

accurate assessment. Furthermore, even humans ﬁnd

it challenging to assess these traits in human-human

interaction. Therefore, 12 realizable traits are consid-

ered. These traits’ equations, which are reported in

(Mehrabian, 1996), are shown in the Equation 2. Fig-

ure 1 shows the schematic ﬂow of the approach.

Intellect = 0.14P + 0.20A + 0.48D

Achievement = 0.13P + 0.60D

Extroversion = 0.21P + 0.17A + 0.50D

SocialDesirablility = 0.34P − 0.26A + 0.17D

ArousalSeeking = 0.14P + 0.26A + 0.55D

Aggression = −0.36P + 0.20A + 0.28D

TraitDominance = 0.72D

PhysicallyActive = 0.26P + 0.40D

Anxiety = 0.24A − 0.20D

Shyness = −0.29P + 0.13A − 0.56D

SensitivitytoRe jection = 0.14A − 0.71D

Nurturance = 0.41P + 0.12A + 0.17D

(2)

4 EXPERIMENTATION

The main hypothesis proposed is that the use of

P.A.D. emotional space, computed through human

nonverbal cues, can provide a successful assessment

of human personality traits in the context of HRI. Due

to the unavailability of ground truth, the validation of

the hypothesis is a challenging task. In order to vali-

date the authenticity of the proposed system, written

feedbacks have been compiled in the form of a ques-

Personality Traits Assessment using P.A.D. Emotional Space in Human-robot Interaction

115

tionnaire from psychology students. In the following

sub-sections, the experimentation and evaluation pro-

cedure is described in detail.

4.1 Experimental Setup

For experimentation purposes, 15 university students

(12 males, 3 females; age range 22-45 years) from

different ethnical backgrounds have participated. The

participants have been naive about the objective and

nature of the experiments. The participants provide

informed consent with regards to the guidelines of an

anonymous research group. Laboratory experiments

have been conducted in a closed ofﬁce room environ-

ment.

Participants have been instructed to stand in the

line of sight of ROBIN, the humanoid robot of TU

Kaiserslautern, as shown in Figure 2. Three cameras

are placed at different locations to record the interac-

tion. ROBIN perception GUI is also recorded for the

duration of the interaction. Artiﬁcial lights are used

during the experiments to make it consistent for all

the participants. An experimenter is also present in

the room to monitor the processes systematically tak-

ing place and only intervene if the system malfunc-

tions because of technical issues.

4.2 Experimental Scenarios

In the direction of validating the proposed hypothe-

sis, participants are assessed for personality traits in

three different scenarios. These scenarios have been

developed with regard to a student’s area of exposure

that consequently leads to three relatable tasks for ex-

perimentation. Each participant has been instructed to

enact the following scenarios one at a time. The ex-

perimenter also introduces participants with ROBIN

at the beginning of the experiments to get familiarized

with ROBIN.

The ﬁrst scenario involves an interaction between

ROBIN, role-playing as the professor, and the partic-

ipant, role-playing as a researcher. The second sce-

Figure 2: Humanoid robot ROBIN.

nario involves an interaction between ROBIN, role-

playing as a master student, and the participant, role-

playing as a supervisor. The last scenario involves an

interaction between ROBIN, role-playing as an inter-

viewer, and the participant, role-playing as a candi-

date. For each scenario, the robot takes the lead by

asking questions and responding generically.

5 PERFORMANCE ANALYSIS

In order to evaluate the personality assessment sys-

tem, summative evaluations are used from 5 psychol-

ogy students. All the interactions are video recorded,

and these videos are used for further analysis. After

the experimentation, each evaluator is presented with

a list of 12 personality traits and their corresponding

description as follows:

1. Intellect: A person who engages in critical think-

ing, research, and reﬂection about society, pro-

poses solutions for its problems

2. Achievement: Something done with effort & skill.

3. Extroversion: Extroverts are behaviorally more

dominant in face-to-face interactions with others.

4. Social Desirability: To answer questions in a man-

ner that is viewed to be favorable by others.

5. Arousal Seeking: A person that looks for excite-

ment, change, new environments, take risks, etc.

6. Aggression: Readiness to attack or confront

7. Dominance: Showing power and inﬂuence

8. Physically Active: A person that is continuously

active, working, organizing activities, etc.

9. Anxiety: Feeling of worry, nervousness, or un-

easiness about something with uncertain outcome.

10. Shyness: Nervous or timid in company of others.

11. Sensitivity to Rejection: People who are affected

easily by the negative remarks of others.

12. Nurturance: Emotional/physical care given to

someone.

Observer evaluations have been used to assess hu-

man personality traits. Evaluators assess the personal-

ity traits of all the subjects using the recorded videos

and descriptions of personality traits to establish the

ground truth. They use the provided descriptions and

their prior knowledge of human behavior to form an

informed judgment about the presence (active) or ab-

sence (inactive) of each personality trait. In order to

combine the outcomes from each evaluator to gen-

erate ground truth, the maximum occurrence of the

outcome is used as the ﬁnal outcome. For example,

HUCAPP 2021 - 5th International Conference on Human Computer Interaction Theory and Applications

116

if two evaluators report the subject’s anxiety trait as

active and three evaluators report it as inactive, then

the ground truth for anxiety trait of the subject is de-

scribed as inactive. These assessments are compared

with the system results. However, the proposed sys-

tem reports the trait values between −1 to +1 range.

For evaluation and validation, the trait with positive

value is considered as active, and the trait with a neg-

ative value as inactive. Table 1 shows the recognition

rates of each personality trait.

Table 1: Personality Traits and Recognition Rates.

ID Trait Dimension Accuracy (%)

1 Intellect 83.11

2 Achievement 78.22

3 Extroversion 91.11

4 Social Desirability 80.44

5 Arousal Seeking 87.55

6 Aggression 84.88

7 Trait Dominance 89.77

8 Physically Active 92.88

9 Anxiety 90.66

10 Shyness 78.66

11 Sensitivity to Rejection 73.33

12 Nurturance 77.77

Average 84.03

5.1 Discussion

It can be seen from Table 1 that extroversion, arousal

seeking, trait dominance, physically active, and anxi-

ety traits have higher recognition rates. It is because

these traits are distinct and easily distinguishable vi-

sually. For example, the system detects the presence

or absence of an extroversion trait with 90% accu-

racy. Most of the subjects that are identiﬁed as high

on extroversion trait are expressive and physically ac-

tive with an open body stance during interactions,

which in turn yields higher scores on the P.A.D. scale.

Therefore, the system estimates the extroversion trait

as active for these subjects. On the other side, some of

the subjects appear to be nervous and shy during inter-

actions, which results in lower values for the P.A.D.

scale. Therefore, the system estimates the extrover-

sion trait as inactive for these subjects. Similarly,

some subjects show anxiety during interactions by ex-

hibiting dejected and self-touching postures. Subjects

are also found to be passive and restless (leaning side

to side) during interactions. It has been found after

analysis that dominance scores for such subjects have

been negative. However, their arousal score is posi-

tive due to their movement attributed to restlessness.

Therefore, the system estimates the anxiety trait as

active for these subjects.

The reason for the wrong assessment of some per-

sonality traits lies in the inaccurate recognition of fa-

cial expressions. The facial expression recognition

(FER) system works accurately when a person is ex-

pressing the emotions clearly. During the interac-

tion, the facial expressions of a subject are sometimes

wrongly interpreted, which affects the pleasure value.

Nurturance and social desirability strongly correlate

with pleasure scale as can be seen from equation 2,

but due to the technical limitation of the FER sys-

tem, these traits achieve low accuracy as depicted in

table 1.

Due to the subjective nature of the task, evaluators

themselves ﬁnd it difﬁcult sometimes to have a mu-

tual consensus on these traits. Because of the avail-

ability of additional information such as verbal, sit-

uational, and contextual cues, evaluators labeled the

ground truths of personality traits for each subject ac-

cordingly. However, the personality traits system uses

only visual information to analyze human behavior

and, therefore, sometimes wrongly reports a subject

on a particular trait. For example, sensitivity to rejec-

tion trait is highly subjective and needs a context as

well to recognize it accurately. Some subjects that ex-

hibit submissive body postures, such as crossed arms

and thinking postures, and avoid eye contact during

interactions are detected as high on sensitivity to re-

jection trait, which may not be accurate in every case.

For example, introverts also show similar body pos-

tures and head gestures, however, they may not be

sensitive to rejection.

Since contextual and situational cues are not con-

sidered in this work, the system is not able to differ-

entiate between fake personality and genuine person-

ality. The system assesses personality based on non-

verbal cues, which can also be sometimes expressed

artiﬁcially. Although the subjects are instructed to en-

act genuinely, some may have faked their responses.

Therefore, the personality assessed by the system may

differ from the actual personality of the subject, which

shows the importance of contextual and situational

cues. However, the highly engaging nature of scenar-

ios, along with the robot’s human-like gestures and

expressions, challenge the subjects to respond with

minimal artiﬁciality in this work. Hence, the system

can assess different personality traits with 84% accu-

racy.

6 CONCLUSION

This paper proposes to use the temperament frame-

work to describe and measure individual differences

on different personality scales using P.A.D. emotional

Personality Traits Assessment using P.A.D. Emotional Space in Human-robot Interaction

117

space. The proposed study has been conducted in psy-

chology but has never been realized in HRI. In order

to validate the research hypothesis, P.A.D. emotional

space has been developed using the recognition / de-

tection of nonverbal cues, such as human postures,

facial expressions, proximity, activity, head gestures,

etc. Using the trait equations provided in the litera-

ture (Mehrabian, 1996), a score has been calculated

for each trait. For evaluation, three scenarios have

been developed, and 15 university students have been

invited. Due to the unavailability of the labeled data,

ﬁve psychology students have been consulted to eval-

uate the personality traits. From validation studies, it

is clear that the framework presented by Mehrabian is

also applicable in HRI by using nonverbal cues.

REFERENCES

Al-Darraji, S., Berns, K., and Rodi

c, A. (2017). Action unit

based facial expression recognition using deep learn-

ing. In Rodi

c, A. and Borangiu, T., editors, Advances

in Robot Design and Intelligent Control, pages 413–

420, Cham. Springer International Publishing.

Argyle, M. (1988). Bodily communication, methuen.[aaw]

armstrong, d.(1984) the patients view. Social Science

and Medicine, 18:73744.

Batrinca, L., Lepri, B., Mana, N., and Pianesi, F.

(2012). Multimodal recognition of personality traits in

human-computer collaborative tasks. In Proceedings

of the 14th ACM international conference on multi-

modal interaction, pages 39–46. ACM.

Batrinca, L. M., Mana, N., Lepri, B., Pianesi, F., and Sebe,

N. (2011). Please, tell me about yourself: automatic

personality assessment using short self-presentations.

In Proceedings of the 13th international conference on

multimodal interfaces, pages 255–262. ACM.

Boukricha, H., Wachsmuth, I., Hofst

atter, A., and Gram-

mer, K. (2009). Pleasure-arousal-dominance driven

facial expression simulation. In 2009 3rd Interna-

tional Conference on Affective Computing and Intel-

ligent Interaction and Workshops, pages 1–7. IEEE.

Chartrand, T. L. and Bargh, J. A. (1999). The chameleon

effect: the perception–behavior link and social inter-

action. Journal of personality and social psychology,

76(6):893.

Costa Jr, P. T. and McCrae, R. R. (1976). Age differences

in personality structure: A cluster analytic approach.

Journal of gerontology, 31(5):564–570.

Hall, E. T. (1963). A system for the notation of proxemic

behavior 1. American anthropologist, 65(5):1003–

1026.

Henderson, M. and Furnham, A. (1982). Similarity and

attraction: The relationship between personality, be-

liefs, skills, needs and friendship choice. Journal of

Adolescence, 5(2):111–123.

Hirth, J., Schmitz, N., and Berns, K. (2011). Towards social

robots: Designing an emotion-based architecture. In-

ternational Journal of Social Robotics, 3(3):273–290.

Jensen, M. (2016). Personality traits and nonverbal com-

munication patterns. Int’l J. Soc. Sci. Stud., 4:57.

Mehrabian, A. (1996). Pleasure-arousal-dominance: A gen-

eral framework for describing and measuring individ-

ual differences in temperament. Current Psychology,

14(4):261–292.

Nass, C. and Lee, K. M. (2001). Does computer-

synthesized speech manifest personality? experi-

mental tests of recognition, similarity-attraction, and

consistency-attraction. Journal of experimental psy-

chology: applied, 7(3):171.

Rodi

c, A., Urukalo, D., Vujovi

c, M., Spasojevi

c, S., Tomi

M., Berns, K., Al-Darraji, S., and Zafar, Z. (2016).

Embodiment of human personality with ei-robots by

mapping behaviour traits from live-model. In Interna-

tional Conference on Robotics in Alpe-Adria Danube

Region, pages 438–448. Springer.

Russell, J. A. (1980). A circumplex model of affect. Journal

of personality and social psychology, 39(6):1161.

Russell, J. A. and Mehrabian, A. (1977). Evidence for a

three-factor theory of emotions. Journal of research

in Personality, 11(3):273–294.

Salam, H., Celiktutan, O., Hupont, I., Gunes, H., and

Chetouani, M. (2016). Fully automatic analysis of en-

gagement and its relationship to personality in human-

robot interactions. IEEE Access, 5:705–721.

Saleh, S. and Berns, K. (2015). Nonverbal communica-

tion with a humanoid robot via head gestures. In Pro-

ceedings of the 8th ACM International Conference on

PErvasive Technologies Related to Assistive Environ-

ments, page 15. ACM.

Watson, D. and Clark, L. A. (1997). Measurement and mis-

measurement of mood: Recurrent and emergent is-

sues. Journal of personality assessment, 68(2):267–

296.

Watson, D. and Tellegen, A. (1985). Toward a consensual

structure of mood. Psychological bulletin, 98(2):219.

Wundt, W. M. and Judd, C. H. (1897). Outlines of psychol-

ogy, volume 1. Scholarly Press.

Zafar, Z., Paplu, S. H., and Berns, K. (2018a). Au-

tomatic assessment of human personality traits: A

step towards intelligent human-robot interaction. In

2018 IEEE-RAS 18th International Conference on

Humanoid Robots (Humanoids), pages 1–9. IEEE.

Zafar, Z., Venugopal, R., and Berns, K. (2018b). Real-

time recognition of human postures for human-robot

interaction. In 11th International Conference on

Advances in Computer-Human Interactions (ACHI),

2018, pages 29–35.

HUCAPP 2021 - 5th International Conference on Human Computer Interaction Theory and Applications

118