Towards a Computational Approach to Emotion Elicitation

in Affective Agents

Joaqu

ın Taverner, Emilio Vivancos and Vicente Botti

Departamento de Sistemas Inform

aticos y Computaci

on, Universitat Polit

ecnica de Val

encia, Spain

Keywords:

Agent, Emotion Representation, Affective Computing.

Abstract:

Interest in affective computing is increasing in recent years. Different emotional approaches have been devel-

oped to incorporate emotions in multi-agent systems. However, most of these models do not offer an adequate

representation of emotions. An internal representation of emotions allows to deﬁne emotions according to

different affective variables. In addition, many of these approaches do not take into account factors such as

culture and language when deﬁning emotions. In this work we show the results obtained in an experiment

carried out to design an affective model for a multi-agent system taking into account factors such as language

and culture.

1 INTRODUCTION

Affective computing (Picard and Picard, 1997) is the

area of computing related to the recognition, process-

ing and simulation of different affective character-

istics including emotions, personality or mood (Al-

fonso et al., 2014; Taverner et al., 2018a). One of the

main goals in affective computing is to create com-

putational systems capable of simulate human emo-

tions. Currently, there is no consensus on the deﬁ-

nition of the term “emotion”. In general, an emotion

can be deﬁned as a rapid response to a given stimulus.

Therefore, emotions have a fundamental importance

in the modeling of intelligent affective agents. Con-

sidering that emotions inﬂuence the behavior of af-

fective agents, credibility of agents simulating human

behavior will depend heavily on the selected model of

emotion.

Emotions depend on language and culture (Rus-

sell et al., 1989). A direct translation of emotions

from one language into other can lead to errors that

produce strange and artiﬁcial agent behaviors. When

developing the affective processes in a multi-agent

system, these factors we must be taken into account.

However, despite the fact that in other domains, such

as personal assistants, cultural and language factors

are taken into account, to the best of our knowl-

edge, in affective computing there are still no propos-

als that really take cultural and language factors into

consideration. Furthermore, most proposals of affec-

tive agents use simple interpretations of psychological

theories that are not generally proposed to be incorpo-

rated into computational models. However, currently

there are very few models of emotions representation

explicitly developed to be used by computers such as

the Scherer’s model (Scherer, 2010). In this paper we

show the preliminary results of a method based on ex-

periments to create a model of emotions for a multi-

agent affective BDI architecture (Alfonso et al., 2017)

adapted to different languages. This emotional model

will improve the simulation of emotional human be-

havior in this multi-agent affective architecture.

This paper is structured as follows: Section 2

presents different psychological models developed to

explain the structure of emotions. In Section 3, some

related works on emotion modeling in affective agents

are presented. In Section 4, we discuss about emotion

modeling in affective agents. Section 5 shows the re-

sults of our method based on experiments to create a

model of emotions for a multi-agent system. Finally,

Section 6 presents the conclusions and future work.

2 PSYCHOLOGICAL MODELS

OF EMOTIONS

From a psychological perspective, there are different

theories to explain how emotions are elicited. The ba-

sic emotions theories (Ekman, 1992) hold that there is

a limited number of emotions and that each event that

can be detected by a human being produces an associ-

ated emotion. In addition, these theories suggest that

emotions have a universal meaning.

On the other hand, appraisal theories (Lazarus,

1991; Ortony et al., 1990) consider that an emotion

Taverner, J., Vivancos, E. and Botti, V.

Towards a Computational Approach to Emotion Elicitation in Affective Agents.

DOI: 10.5220/0007579302750280

In Proceedings of the 11th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2019), pages 275-280

ISBN: 978-989-758-350-6

275

is the result of the evaluation process that is activated

when an event is received. This process is executed

through a set of variables known as appraisal vari-

ables. The number and type of appraisal variables

usually vary depending on the psychological theory,

but appraisal theories have been widely used in affec-

tive computing to create emotional agents (Gebhard,

2005; Gratch and Marsella, 2004).

Finally, the constructivist emotion theories (Rus-

sell and Barrett, 1999; Thayer, 1997), in contrast to

the basic emotions theories, relate emotions to the lan-

guage and culture used by humans. These theories

explain the great differences using emotional terms in

different geographical and cultural areas. For exam-

ple, in the German language there is an emotion called

“Schadenfreude” whose meaning is “pleasure for the

suffering of others”. However, in languages such as

English or Spanish there is no word to deﬁne this

emotion. One of the most well-known constructionist

theorists is Russell (Russell and Mehrabian, 1977).

Russell proposes in his Circumplex Model of Af-

fect (Russell, 1980) that emotions can be represented

using two dimensions: pleasure and arousal. Accord-

ing to Russell’s experiments (Russell et al., 1989),

emotions represented by the same word in different

languages and cultures, can be associated with differ-

ent values of pleasure and arousal.

Russell’s model shares the Pleasure and Arousal

variables with the PAD model (Pleasure, Arousal, and

Dominance) (Mehrabian, 1996). This model is well

known in affective computing and is often used to rep-

resent the mood and emotions. The PAD model adds

the Dominance dimension to the representation of

emotions. However, generally this dimension is usu-

ally used as an appraisal variable while the Pleasure

and Arousal dimensions form what is known as the

core affect that structures the basic feelings (Marsella

et al., 2010; Russell, 2003).

3 RELATED WORKS

Over the last years, different proposals to use

emotions in multi-agent systems have been made

(Marsella et al., 2010; Esteban and Insua, 2017;

Gratch and Marsella, 2004; Alfonso et al., 2015).

These proposals adapt different theories of emotion

to create affective agents capable of selecting one

emotion when a certain event occurs. For example,

Marsella (Marsella and Gratch, 2009) proposes EMA

(a process model of appraisal dynamics) in which an

appraisal process based on the Lazarus’ appraisal the-

ory (Lazarus, 1991) is used. Marsella deﬁnes differ-

ent appraisal variables thresholds to elicit emotions.

Pleasure

Arousal

Tense

Nervous

Stressed

Upset

Unpleasant

Sad

Depressed

Bored

Fatigued

Tired

Calm

Relaxed

Serene

Contented

Happy

Elated

Exited

Alert

Fear

Anger

Disgust

Sadness

Surprise

Happiness

Figure 1: The circumplex model of affect. Source: (Russell

and Barrett, 1999).

In EMA, emotions are represented as labels and the

intensity is deﬁned as the product of different numeric

appraisal variables. The appraisal model proposed for

the EMA agent is also used in other affective agent

models such as GenIA

(Alfonso, 2017).

Other proposals for affective agents use dimen-

sional models of emotions. For example, Saldien

(Saldien et al., 2010) uses Russell’s Circumplex

Model of Affect to represent emotions in the robot

Probo. But in Saldien’s model, the Russell’s scheme

(Fig 1) is used as a reference for the situation of the

emotions without taking into consideration that this

scheme was not designed for that purpose.

Gebhard (Gebhard, 2005) uses the PAD space

in his emotional model. He deﬁnes an appraisal

model based on the OCC appraisal theory (Ortony

et al., 1990) called ALMA (A Layered Model Of Af-

fect). In this model, emotions obtained as a result of

the appraisal process are transformed into the PAD

space. Gebhard proposes to use a mapping that turns

each emotion into a point in the PAD space. Then,

these emotions are used to estimate the agent’s mood

(Mehrabian, 1996). In addition, personality is used to

deﬁne how the mood decay over time.

4 DISCUSSION

Despite the large numbers of proposals to model emo-

tions in multi-agent systems, there is still much work

to be done to obtain a realistic simulation of emotions.

Most emotional models were proposed to interpret

and analyze the way in which emotions are used by

human beings. But there are few models explicitly de-

veloped to be used in computational systems (Scherer,

ICAART 2019 - 11th International Conference on Agents and Artiﬁcial Intelligence

276

2010). In general, most affective computational mod-

els designed to create affective agents, adapt theories

from psychology. However, in most of these models,

emotions are represented as simple labels (as in most

proposals in emotion recognition). This simpliﬁca-

tion may not be appropriate when modeling complex

affective behaviors such as empathy, the effect of per-

sonality and mood on emotions, or the affect decay

rate process (which is the process that reduce the emo-

tion intensity over time). From a computational per-

spective, a dimensional representation of emotions, as

Russell (Russell and Barrett, 1999) proposes, seems

to be more appropriate to design this type of affective

behaviors.

The Circumplex Model of Affect (Russell, 1980)

explains the occurrence of emotions using a two-

dimensional model based on the pleasure and arousal

values. Through this model, Russell proves that emo-

tions follow a circular pattern along the pleasure-

arousal dimensions. In his experiments he provides an

approximate model to relate each emotions to a pair of

pleasure and arousal values. This model was designed

to demonstrate the circular relationship of emotions

with the pleasure and arousal levels. The resulting

scheme (see Figure 1) uses a simpliﬁed location for

each emotion, because its main purpose is to proof

the circular relation between emotions and their plea-

sure and arousal levels. Apparently, a pair of values

(pleasure, arousal) corresponds to one emotion in the

circular representation of the Russell model and this

interpretation of the circular representation is directly

used in some affective agent models (Saldien et al.,

2010). But that use was not the original purpose of

the Rusell’s model and therefore, this representation

should not be translated directly into a computational

model. For example, in the scheme proposed by Rus-

sell, all emotions have exactly the same area. That

is completely unnatural, and consequently, an agent

using this simpliﬁed representation of emotions will

show an erratic behaviour, very different to the real

behaviour of a human being.

Even more important than this misuse of the origi-

nal Rusell’s model, the enormous differences between

emotions in different cultures, geographical areas,

and languages must also be taken into consideration.

When a literal translation of the English words used

to represent emotions is used in other language, a high

level of inaccuracies is introduced. When a human be-

ing uses a word to identify his/her emotion, this word

cannot be directly translated into other language. But,

how could we design an emotional agent interacting

with people if its emotional model is based on exper-

iments where people expressed their emotions using

words in a different language? Obviously, a lot of

inaccuracies will be introduced and the perceived be-

haviour of the emotional agent could look artiﬁcial

and erratic. Consequently, we propose a method to

adapt the Russell’s Circumplex Model of Affect to be

effectively used in multi-agent systems. This adap-

tation is ready to be used in a computational model

and it is based on an experiment where human beings

are expressing their emotions in the same language in

which the agent will be recognizing and expressing

emotions.

5 A CIRCUMPLEX MODEL TO

REPRESENT EMOTIONS IN A

CULTURAL CONTEXT

As we have argued before, one of the main challenges

that arise when designing affective agents is to de-

velop the process to select emotions. The behavior

of the agent will depend to a great extent on this pro-

cess. Therefore, a bad design can lead to unrealistic

agents showing an artiﬁcial and strange behavior. We

propose a computational model to represent emotions

in a dimensional space similar to the one proposed by

Russell’s (Russell, 1980) but our model is adapted to

a speciﬁc cultural environment. This representation

is appropriate to be used in computational models be-

cause emotions can be easily and effectively repre-

sented using two numerical variables: pleasure and

arousal. This bi-dimensional numerical representa-

tion uses variables instead of labels and therefore is

more accurate and appropriate to be use in a compu-

tational model. Emotional processes can easily mod-

ify the pleasure an arousal values to simulate, in a

more natural way, transitions between emotions. On

the other hand, evidence shows that intensity is re-

lated to the pleasure and arousal values (Reisenzein,

1994). High levels of these two variables correspond

to high intensity of emotions while, low values of

these two variables correspond to a low emotional in-

tensity. Very low levels of arousal and pleasure can be

identiﬁed as absence of emotion. This variation in the

pleasure and arousal values can be used to implement

the affective decay rate function of an affective agent.

Contrary to other proposals based in Russell’s

model, our emotion representation model redeﬁnes

the values of pleasure and arousal assigned to each

emotion. As we mentioned in the previous sections,

the circumplex scheme was proposed to support Rus-

sell’s theory in which each emotion could be repre-

sented in a circle according to its levels of pleasure

and arousal. This scheme can be easily criticized

when used in a real computational system, because

Towards a Computational Approach to Emotion Elicitation in Affective Agents

277

that was not the original use of this model. The ﬁrst

unjustiﬁed decision is the size of the areas assigned

to each emotion. In the original model, every emo-

tion receives exactly the same area of the circle, but

this decision looks very artiﬁcial and, as shown in

our experiment, it does not correspond to the men-

tal representation in real humans. The second impor-

tant problem of this model, when used to represent

human emotions in an agent, is the identiﬁcation of

emotions in the borders between the areas assigned to

each emotion. We know that these borders are arti-

ﬁcial because they do not exist in the mental repre-

sentation of emotions in humans, but this fact is fre-

quently ignored. Finally, Russell’s model is based on

the results obtained in experiments in which British

English-speakers identiﬁes emotions represented by

English words. However, when designing an agent

to interact with human expressing their emotions in

a different language, a new level of inaccuracy is in-

troduced. Our method uses an experiment to adapt

the emotional vocabulary and mental representation

of emotions to the language and culture of the envi-

ronment in which the agent will be located.

The ﬁrst step of our methodology to represent

emotions is the data acquisition. The data acqui-

sition allows us to learn the levels of pleasure and

arousal related to different emotions in the environ-

ment (language and cultural area) in which the agent

will be located. In our experiment we analyze Eu-

ropean Spanish-speakers because our agent will be

located in this environment. In this experiment 10

emotions are identiﬁed by the participants according

to their pleasure and arousal levels: fear, angry, dis-

gust, sadness, boredom, sleepiness, calm, happiness,

excitement, and surprise.

Methodology. A hundred volunteers from different

age, sex, and cultural level, were asked to assign a

level of pleasure and arousal to the 10 selected emo-

tions.

Participants assigned to each emotion a pleasure

value using an integer number between 1 and 7 (from

”very misery” to ”very pleased”) in a questionnaire.

Note that the value 4 corresponds to a neutral value

for this variable. On the other hand, participants as-

sign an arousal value between 1 and 7 (where 1 cor-

responds to ”very sleep” and 7 to ”very aroused”) to

the same set of 10 emotions.

Results. The results of the experiment are summa-

rized in Figure 2. These results show how the 10 emo-

In the experiment, we use the Spanish words corre-

sponding to these English words.

−2

−2 0 2

Pleasure

Arousal

(a) Fear

112

−2

−2 0 2

Pleasure

Arousal

(b) Angry

−2

−2 0 2

Pleasure

Arousal

−2

−2 0 2

Pleasure

Arousal

(d) Sad

157

−2

−2 0 2

Pleasure

Arousal

(e) Bored

9 1

−2

−2 0 2

Pleasure

Arousal

(f) Sleepy

−2

−2 0 2

Pleasure

Arousal

(g) Calm

−2

−2 0 2

Pleasure

Arousal

(h) Happy

−2

−2 0 2

Pleasure

Arousal

(i) Excited

−2

−2 0 2

Pleasure

Arousal

(j) Surprise

Figure 2: Number of people who selected each value of

pleasure and arousal for the ten emotions.

ICAART 2019 - 11th International Conference on Agents and Artiﬁcial Intelligence

278

tions are related with the pleasure and arousal dimen-

sions, as assumed in the circumplex model. For ex-

ample, the fear emotion is related to high levels of

arousal and low levels of pleasure, while the happy

emotion is more related to high levels of pleasure and

arousal.

Figure 3a shows the means obtained in the exper-

iment for the emotions expressed by Spanish words

(emotions that Russell did not represent in his ex-

periments with British Engish-speakers are shown

in gray). Figure 3b shows the means obtained by

Russell for emotions expressed by English words.

Some differences can be easily detected. For ex-

ample, for British English-speakers happiness is re-

lated to a lower level of activation than for Euro-

pean Spanish-speakers. This corroborates the results

obtained by Russell in his cross-cultural experiment

(Russell et al., 1989), in which he appreciated signif-

icant differences between emotions in different lan-

guages. On the other hand, these results agree with

the constructivist theories in which emotion depends

on factors such as language and culture.

These results conﬁrm the importance of our pro-

posal to create a speciﬁc emotion representation for

affective agents depending on the environment in

which the agent will be located. This method allows

to adapt the emotional model to the language and cul-

ture of the humans that will be interacting with the

affective agent. From these results we can easily ap-

proximate the regions where emotions are most likely

to occur in the emotional model. Therefore, these

results can be easily used to create a computational

model that represents emotions using their levels of

pleasure and arousal. Moreover, the results can also

be used to train a model to classify emotions accord-

ing to the levels of pleasure and arousal. This model

can be useful to map emotions onto the pleasure and

arousal dimensions using sentiment analysis or facial

expressions.

6 CONCLUSIONS AND FUTURE

WORK

Most of the models proposed to create affective agents

are approximations of psychological theories. In gen-

eral, these models use emotions as simple labels.

Having a dimensional representation of emotions can

help to understand and develop the emotional behav-

ior of affective agents. This allows to design emo-

tional processes such as the continuous process to

evolve one agent from one emotional state to other,

or the affect decay rate process.

In this paper we have shown the preliminary re-

Angry

Pleasure

Arousal

0.5 1.00.0-0.5-1.0

-1.0 0.5 0.0 0.5 1.0

Fear

Disgust

Sad

Bored

Sleepy

Calm

Surprise

Excited

Happy

(a) Results obtained for the 10 emotions trans-

lated into Spanish.

Pleasure

Arousal

Angry

Sad

Bored

Sleepy

Calm

Excited

Happy

(b) Results obtained by Russell.

Figure 3: Comparison between the results obtained by our

experiment (emotions expressed in European Spanish) and

the Russell’s experiment (emotions expressed in British En-

glish).

sults of a method to create a model of emotions for a

multi-agent system adapted to a speciﬁc language and

cultural area. The results demonstrate the importance

of performing this type of emotional analysis to adapt

affective psychological models to the environment in

which the emotional agent will be located. We have

shown how emotions can be interpreted in a very dif-

ferent way depending on the language used. From the

obtained data we can deduce the values of pleasure

and arousal associated to the different emotions. This

association can be used by agents to determine what

emotion must be shown according to the level of plea-

sure and arousal detected.

Using the experiment results, we can determine

the levels of pleasure and arousal corresponding to

each emotion. Therefore, given a pair of values for

pleasure and arousal, the agent can deduce the ex-

pressed emotion. On the other hand, when an emo-

tion is detected, the emotion can be represented in the

Towards a Computational Approach to Emotion Elicitation in Affective Agents

279

pleasure and arousal space. Using the same represen-

tation for agent and human emotions, we facilitate the

development of affective abilities such as emotional

contagion or empathy.

Currently we are implementing an emotional ap-

praisal model based on the results shown in this arti-

cle. This model will use a dimensional representation

of emotions based on the pleasure and arousal vari-

ables. In addition, our appraisal model will take into

consideration the inﬂuence of personality and mood

when selecting the emotion (Taverner et al., 2018b).

As part of our future work, this model will be incor-

porated into the affective agent architecture GenIA

(Taverner et al., 2016). The incorporation of this emo-

tion representation into GenIA

will facilitate emotion

recognition, empathy, or emotional contagion.

ACKNOWLEDGEMENTS

This work is partially supported by the FPI

grant ACIF/2017/085, Spanish Government project

TIN2017-89156-R, and GVA-CEICE project PROM-

ETEO/2018/002.

REFERENCES

Alfonso, B. (2017). Agents with Affective Traits for

Decision-Making in Complex Environments. PhD

thesis, Universitat Polit

ecnica de Valencia, Spain.

http://hdl.handle.net/10251/90497.

Alfonso, B., Vivancos, E., and Botti, V. (2014). An open

arquitecture for affective traits in a BDI agent. Proc.

of the 6th ECTA 2014. Part of the 6th IJCCI 2014,

pages 320–325.

Alfonso, B., Vivancos, E., and Botti, V. (2017). Toward

formal modeling of affective agents in a BDI archi-

tecture. ACM Transactions on Internet Technology

(TOIT), 17(1):5.

Alfonso, B., Vivancos, E., Botti, V., and Hern

andez, P.

(2015). Building emotional agents for strategic de-

cision making. In Proceedings of the International

Conference on Agents and Artiﬁcial Intelligence, vol-

ume 2, pages 390–397. SCITEPRESS.

Ekman, P. (1992). An argument for basic emotions. Cogni-

tion & emotion, 6(3-4):169–200.

Esteban, P. G. and Insua, D. R. (2017). An affective model

for a non-expensive utility-based decision agent. IEEE

Transactions on Affective Computing.

Gebhard, P. (2005). Alma: a layered model of affect.

In Proc. of the fourth international joint conference

on Autonomous agents and multiagent systems, pages

29–36.

Gratch, J. and Marsella, S. (2004). A domain-independent

framework for modeling emotion. Cognitive Systems

Research, 5(4):269–306.

Lazarus, R. S. (1991). Emotion and adaptation. Oxford

University Press.

Marsella, S., Gratch, J., Petta, P., et al. (2010). Computa-

tional models of emotion. A Blueprint for Affective

Computing-A sourcebook and manual, 11(1):21–46.

Marsella, S. C. and Gratch, J. (2009). EMA: A process

model of appraisal dynamics. Cognitive Systems Re-

search, 10(1):70–90.

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.

Ortony, A., Clore, G. L., and Collins, A. (1990). The cog-

nitive structure of emotions. Cambridge University

Press.

Picard, R. W. and Picard, R. (1997). Affective computing,

volume 252. MIT press Cambridge.

Reisenzein, R. (1994). Pleasure-arousal theory and the in-

tensity of emotions. Journal of Personality and Social

Psychology, 67(3):525.

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

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

Russell, J. A. (2003). Core affect and the psychologi-

cal construction of emotion. Psychological review,

110(1):145.

Russell, J. A. and Barrett, L. F. (1999). Core affect, pro-

totypical emotional episodes, and other things called

emotion: dissecting the elephant. Journal of person-

ality and social psychology, 76(5):805.

Russell, J. A., Lewicka, M., and Niit, T. (1989). A cross-

cultural study of a circumplex model of affect. Journal

of personality and social psychology, 57(5):848.

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

three-factor theory of emotions. Journal of research

in Personality, 11(3):273–294.

Saldien, J., Goris, K., Vanderborght, B., Vanderfaeillie, J.,

and Lefeber, D. (2010). Expressing emotions with the

social robot probo. International Journal of Social

Robotics, 2(4):377–389.

Scherer, K. R. (2010). The component process model: Ar-

chitecture for a comprehensive computational model

of emergent emotion. Blueprint for affective comput-

ing: A sourcebook, pages 47–70.

Taverner, J., Alfonso, B., Vivancos, E., and Botti, V.

(2018a). Modeling Personality in the Affective Agent

Architecture GenIA3. In Proceedings of the 10th In-

ternational Conference on Agents and Artiﬁcial In-

telligence. ICAART’18, volume 1, pages 236–243.

SCITEPRESS.

Taverner, J., Alfonso, B., Vivancos, E., and Botti, V. J.

(2016). Integrating expectations into Jason for ap-

praisal in emotion modeling. In Proc. of the 8th Inter-

national Joint Conference on Computational Intelli-

gence, IJCCI 2016, Volume 1: ECTA, pages 231–238.

Taverner, J., Alfonso, B., Vivancos, E., and Botti, V. J.

(2018b). Using different agent personalities to ob-

tain different agent behaviours. In Proceedings of the

Workshop on Affective Computing and Context Aware-

ness in Ambient Intelligence (AfCAI 2018), Valencia,

Spain, April 19-20, 2018.

Thayer, R. E. (1997). The origin of everyday moods: Man-

aging energy, tension, and stress. Oxford University

Press, USA.

ICAART 2019 - 11th International Conference on Agents and Artiﬁcial Intelligence

280