=Paper=
{{Paper
|id=Vol-2897/session3_paper2
|storemode=property
|title=Towards A Six-Level Framework of Emotional Intelligence for Customer Service Chatbots
|pdfUrl=https://ceur-ws.org/Vol-2897/AffconAAAI-21_paper5.pdf
|volume=Vol-2897
|authors=Qicheng Ding,V. Ivan Sanchez Carmona,Mingming Liu,Fangzhen Peng,Yu Zhang,Changjian Hu
|dblpUrl=https://dblp.org/rec/conf/aaai/DingCLPZH21
}}
==Towards A Six-Level Framework of Emotional Intelligence for Customer Service Chatbots==
https://ceur-ws.org/Vol-2897/AffconAAAI-21_paper5.pdf
Towards A Six-Level Framework of Emotional Intelligence for Customer
Service Chatbots
Qicheng Ding1, V. Ivan Sanchez Carmona2, Mingming Liu1,
Fangzhen Peng1, Yu Zhang1, Changjian Hu2
1
Technical Strategy and Innovation Platform, Lenovo Research, China
2
AI Lab, Lenovo Research, China
dingqc1, vcarmona, liumm9, pengfz1, zhangyu29, hucj1@lenovo.com
Abstract show frustration. Ideally, the chatbot would make a paren-
A chatbot with emotional intelligence can improve the inter- thesis to let the user know that his frustration has been no-
action experience and efficiency with human users. However, ticed, and that such an emotional expression has triggered a
there is no standard scheme or methods to evaluate the degree request to verify that he is understanding the information
of emotional intelligence of chatbots. Based on our industrial clearly. This ideal solution arises from a chatbot with emo-
practice, this paper aims to propose a preliminary framework
to evaluate to which extent a chatbot is emotionally intelli- tional intelligence.
gent. We divide it into two parts: emotional understanding Chatbots that have no model of emotional intelligence
and emotional strategy. Each part contains several factors may be at risk of jeopardizing the whole interaction since
that are important to emotional intelligence. Furthermore, we different users may have different preferences for the chat-
propose a six-level scheme as a guideline to evaluate the bot’s behavior. For example, some users may prefer a repet-
emotional intelligence. This paper provides valuable sugges-
tion in the entire lifecycle of a chatbot including design, im- itive chatbot that frequently rephrases its replies in different
plementation, evaluation, and maintenance. ways for the sake of clarity over a chatbot that never repeats
information; some users may feel comfortable interacting
with a succinct chatbot that formulates statements in as short
1 Introduction and concise sentences as possible; some users may prefer a
Emotional intelligence is a fundamental part of the overall chatbot that is constantly checking whether the user is still
intelligence of a human being. Emotions mediate the cogni- engaged in the interaction and is understanding every reply
tive processes related to individual intelligence. In other from the chatbot. Emotional indicators such as facial ges-
words, a person’s capacities such as memory, decision-mak- tures, voice tone, body gesture, or specific sentences are nat-
ing, and reasoning are affected by his or her emotional ural indicators of confusion, boredom, frustration, and other
states. Furthermore, emotions also affect social interactions. emotions that show the dissatisfaction from the user. Chat-
It is easier to achieve a consensus when group members are bots failing to perceive, attend, and act upon these emotional
in pleasant moods rather than bad ones. To achieve these signals may lead the user to a stressful situation, and to
pleasant moods, it may be necessary for group members to waste a lot of time to solve a problem, or even worse, to
identify and manage emotions, abilities that are the building abruptly quit the session and leave with problem unsolved.
blocks of emotional intelligence. In this paper, we propose a 6-level framework of emo-
Chatbots, which refer to dialog systems that aim to solve tional intelligence for chatbots. This framework aims to
task-oriented problems or provide social interactions, are al- characterize the abilities required by chatbots on different
ways inevitably embedded in human-computer interactions. levels of emotional intelligence. To do so, we first provide
Chatbots with emotional intelligence can correctly identify a model of emotional intelligence for chatbots, which we
users’ emotions and provide concise responses, which rep- build based on the human-centric model of emotional intel-
resent a stronger interaction ability and better user satisfac- ligence of (Salovey & Mayer, 1990; Salovey P. G., 2005;
tion. For example, when a user is unsatisfied with the chat- Mayer J. D., 1999; Mayer J. D., 2016). Similar to the work
bot, the user may not speak out, but his face will clearly of Salovey, we propose a two-branch model of emotional
Copyright © 2021 for this paper by its authors. Use permitted under Crea-
tive Commons License Attribution 4.0 International (CC BY 4.0).
intelligence where each of the branches correspond to broad 2018). For example, Brackett et al. used MSCEIT, V2.0 and
abilities that we believe chatbots should possess, namely found out that emotional intelligence in males was associ-
Emotional Understanding and Emotional Strategy. The first ated with everyday behavior (Brackett, Mayer, & Warner,
branch corresponds to the ability of chatbots of perceiving 2004); Rosete & Ciarrochi argue that higher emotional in-
and recognizing emotional indicators from the users; the telligence prompted workplace performance (Rosete &
second branch corresponds to the ability of formulating and Ciarrochi, 2005).
carrying out a strategy to deal with the emotional states of In the field of Artificial Intelligence, efforts have been
users in order to successfully solve the target task the user made to both standardize (Sedoc, et al., 2019; Miller, et al.,
requires the chatbot to solve. 2018) and improve (Sedoc & Ungar, 2020) the evaluation of
Chatbots can then be classified based on our 6-level chatbots’ dialogue capabilities using both human judgments
framework. For example, while a null-intelligent chatbot and automatic metrics.
that falls at the Level 0 in our framework works with no ca- Also, research in the field of Human-Computer Interaction
pabilities to handle emotions, a highly-intelligent (Level 3) has pinpointed the importance of AI applications to match
chatbot can perceive, recognize and respond to users’ emo- social norms while avoiding social biases (Amershi, et al.,
tions by integrating information across multiple channels 2019). Moreover, it has been discussed the importance of
(such as text, voice, images from video camera) during the emotional intelligence in chatbots to better engage users into
entire conversation while carrying out strategies to handle a more fruitful conversation (Chaves & Gerosa, 2019).
the users’ emotions to solve the target task. At the top of our However, to our best knowledge, there is a lack of standard
framework we categorize a chatbot as fully-intelligent if it methods to evaluate the AI-based implementation (e.g., call
can fully understand, remember, predict and respond to us- center chatbot) from an emotional intelligence perspective.
ers’ emotions by integrating information from all visual, au- Therefore, this paper aims to provide a preliminary guide-
ditory and tactile channels while carrying out the appropri- line to support stakeholders to better understand and evalu-
ate strategy to both keep the user satisfied and successfully ate how emotionally intelligent their AI products are.
solve the target task.
Our 6-level framework can benefit in advancing research
in chatbots in different perspectives. For example, designers 3 Emotional Intelligence
and engineers can use our framework to check the abilities We stick with the definition of emotional intelligence from
of their chatbots; researchers can build roadmaps to guide the work of Mayer and Salovey (Salovey & Mayer, 1990;
future research directions for emotional intelligence in chat- Mayer J. D., 2016), due to its wide acceptance on both dis-
bots based on our definition of intelligence levels; industry ciplines Psychology and Affective Computing (Picard,
developers can use our emotional intelligence model to vis- 2000). According to Mayer and Salovey, emotional intelli-
ualize the technology required for chatbots to perceive and gence is characterized across four branches: 1) perception
recognize the emotional states of users. and expression of emotion, 2) using emotion to facilitate
thought, 3) understanding emotions, and 4) managing emo-
tions.
2 Related Work The first branch, perception and expression of emotion,
The lack of objective and automatic measures of emotional refers to the ability of recognizing one’s own emotions and
intelligence is one of the major limitations across diverse to accurately display them through different means such as
fields (Miners, Côté, & Lievens, 2017). In the field of Psy- physical actions. This ability extends to recognizing other’s
chology, a number of researchers have attempted to develop emotions also through different means such as language, fa-
self-reported scales of emotional intelligence; however, cial expressions, and behavior.
their emotional intelligence definitions are often incon- The second branch, using emotion to facilitate thought, re-
sistent with each other (Bar-On, 1997; Roger & Najarian, fers to the ability of using emotions to leverage cognitive
1989; Salovey & Mayer, 1990; Schutte, et al., 1998). processes, such as planning, reasoning, making decisions,
Mayer et al., (Mayer, Caruso, & Salovey, 1999) con- and so on. Moreover, this branch also includes the ability to
structed a Multi-factor Emotional Intelligence Scale (MEIS) adequately choose problems according to one’s own emo-
that is objective (has correct answers), reliable and less as- tional state in order to take advantage of such a state; for
sociated with personality. After that, they established a more example, happiness may stimulate a person’s creative think-
reliable scale named the Mayer-Salovey-Caruso Emotional ing, so choosing to solve problems that require creativity
Intelligence Test, Version 2.0 (MSCEIT, V2.0) (Mayer, when one is happy may be deemed as an intelligent decision.
Salovey, & Caruso, 2002). MSCEIT is one of the most wide- The third branch, understanding emotions, refers to the
spread measure of emotional intelligence (Fiori, et al., 2014; knowledge that a person has about different types of emo-
Papadopoulos, Gkintoni, Halkiopoulos, & Antonopoulou, tions: how they may develop throughout time, how they may
interact with each other and what the possible causes and with emotions with proper
effects of emotions are. privacy consideration.
The last branch, managing emotions, is the ability to self- Table 1 Two-branch model of emotional intelligence for chatbots.
manage one’s own emotions, as well as to manage the emo- Each branch decomposes along different dimensions which char-
tions of other people, in order to achieve a desired goal. For acterize in a finer-grained detail the abilities required for chatbots.
example, when a dad is upset with his kid due to an improper
behavior, he knows that to have an effective conversation 4 Chatbot-Centric Model of Emotional In-
with his kid he should first manage his own emotions and
elicit a more calm and peaceful emotional state; moreover, telligence
if the kid is crying, the dad will try to comfort him in order Based on the four-branch model from Mayer and Salovey,
to have a fruitful conversation. we propose a two-branch model of emotional intelligence
Important to notice is that each of the branches in this for chatbots where we call these two branches as Emotional
framework are subject to social and cultural backgrounds; Understanding and Emotional Strategy. While the branch of
variations in emotional indicators across different cultures emotional understanding refers to the ability of a chatbot of
must be acknowledged and respected. For example, an overt understanding users’ emotions, the branch of emotional
smile may be interpreted with different levels of happiness strategy focuses on the capability of a chatbot in dealing
(such as happy vs. very happy) in different cultures. Thus, with the users’ emotions. These two branches are character-
for an individual to be emotionally intelligent, she must be ized across several dimensions, as is shown in Table 1.
aware of and understand the social and cultural context.
4.1 Emotional Understanding (EU)
Branch Dimen- Emotional understanding refers to chatbots’ ability in under-
Description
name sion standing customers’ emotions while solving their issues.
The range of emotions the Chatbots need to know customers’ emotional status before
Breadth acting properly. It includes the number/range of emotions
bot can identify.
The number of channels of chatbots can understand (breadth), the number of channels
Variety information the bot can uti- of information (variety) and the contextual information
Emotional chatbots can utilize (contextual-wise).
lize to understand emotions.
understand-
The contextual information
ing 4.1.1 Breadth
that the bot can utilize to un-
Contex-
derstand emotions, e.g., from The breadth of emotional understanding refers to the range
tual-wise
one conversation turn to the of emotions that a chatbot can identify. Generally, research-
entire conversation. ers treat emotions as either discrete (Ekman, 1994) or di-
The range of emotions that mensional (Mehrabian, 1980), and they both have pros and
the bot can deal with, e.g., cons (Gunes, 2016). We regard both classification methods
Breadth as a suitable choice to define the breadth of emotions to be
from dealing with one emo-
tion to seven emotions. recognized by chatbots since we believe both methods can
The different ways in which accurately capture users’ emotions. To characterize the
breadth of emotions that a chatbot should identify, we pro-
the bot can deal with one
pose a simple logic: the more types of emotions a chatbot
Variety type of emotion, e.g., from
can identify the smarter it is and thus the higher in our scale
pure a text-based strategy to it is positioned. Table 2 shows this logic. This scale can be
a multi-media strategy. effectively taken by stakeholders as a criterion to qualita-
Emotional The contextual information tively evaluate the level of emotional intelligence of a chat-
strategy that the bot can utilize to deal bot in terms of emotion identification.
Contex-
with emotions, e.g., from one
tual-wise
conversation turn to the en- Capac-
tire conversation. Level Description
ity
How well the bot’s strategy 0 No Cannot understand user’s emotion.
fit into the customer’s social Can understand limited (one or
Social and cultural background and 1 Basic
two) emotions.
norm avoid biased response, e.g., 2 Partial Can understand several emotions.
from being unaware of any Can understand major emotions,
customer privacy to deal 3 High
e.g., Ekman’s seven emotions
(Ekman, 1994) or PAD model to use the whole interaction with the user to extract and iden-
(Walter, 2011). tify emotional states from the user. In this way, a Level 1
Very chatbot may recognize that a user is angry at the current turn,
4 Can understand most emotions. but it will not be able to use this information to build up a
High
5 Full Can fully understand emotions. global model of the user’s emotional state; however, a Level
Table 2. Six-Level Scheme for Breadth of EU. 4 chatbot is not only able to build such a model, but it can
also recall information from previous interactions to build
4.1.2 Variety the emotional model in a finer-grained way.
Variety refers to the number of channels of information a
Capac-
chatbot can utilize to perceive and recognize users’ emo- Level Description
ity
tions. Similar to our logic in Section 4.1.1, the more chan-
nels a chatbot uses to identify customers’ emotions, the 0 No No continuity.
more intelligent can be deemed. Thus, the main challenge Can identify emotion from the on-
1 Basic
remains in integrating the information across the different going conversation turn.
channels rather than merely augmenting the number of Can continually identify emotion
2 Partial
channels. Nevertheless, the more channels are used, the from multiple turns.
more the privacy concerns to be considered: a multi-media Can continually identify emotion
chatbot does not mean that it can open and use all capable 3 High from the entire ongoing conversa-
channels every time; when a customer expresses his pre- tion.
ferred channel(s), the chatbot should adjust its variety to fit Can continually identify and re-
the customer’s choice. Table 3 shows the degrees of com- Very member emotion from the entire
4
plexity in handling different channel types as the number of High ongoing and historical conversa-
channels increment. tion.
Can fully identify, remember emo-
Capac- tion as well as anticipate emotion
Level Description 5 Full
ity from the entire ongoing and histor-
0 No No usable channel. ical conversation.
Can use one channel to understand Table 4. Six-Level Scheme for Contextual-wise of EU.
1 Basic
emotion e.g., text, audio etc.
Can use multiple channels sepa- 4.2 Emotional Strategy (ES)
2 Partial rately but cannot integrate the in- Emotional strategy describes the chatbots’ ability to deal
formation together. with customer emotions. It includes the range of emotions
Can integrate multiple channels, (breadth), the expression channels (variety) and the contex-
3 High i.e., can combine multi-channel tual information it can use, and the ability to fit into the so-
emotional information together. cial and cultural background (social norm).
Can acquire and integrate multi-
4
Very channel emotional information; 4.2.1 Breadth
High also, can distribute different Breadth in emotional strategy refers to the range of emotions
weights for each channel. that a chatbot can respond to, e.g., from one emotion to
Can use all channels (visual, audi- seven emotions. Table 5 shows our characterization of this
5 Full tory, tactile etc.) flexibly and dimension across varying levels of complexity. For exam-
properly to understand emotion. ple, a chatbot capable of dealing with customers’ major
Table 3. Six-Level Scheme for Variety of EU. emotions can be deemed as very-highly intelligent.
4.1.3 Contextual-wise Capac-
Level Description
Contextual-wise refers to the contextual information that the ity
chatbot can utilize to understand emotions, e.g., from using 0 No Cannot deal with emotion.
a single turn from the conversation to using the entire con- 1 Basic Can deal with up to two emotions.
versation. The contextual information the chatbot can use 2 Partial Can deal with several emotions.
include, but is not limited to: conversation history, user pro- 3 High Can deal with major emotions.
file, the interaction with other customers, etc. Table 4 shows Very
the incremental capability of a chatbot in using contextual 4 Can deal with most emotions.
High
information for understanding users’ emotions. While a 5 Full Can fully deal with emotions.
chatbot with a basic ability can extract emotional indicators Table 5. Six-Level Scheme for Breadth of ES.
from a single turn at a time, a highly capable chatbot is able
4.2.2 Variety 1 Basic
Can deal with emotion from the
Variety refers to the different ways in which a chatbot can ongoing conversation turn.
deal with one type of emotion, e.g., from a pure text-based Can continually deal with emotion
2 Partial
strategy to a multi-media strategy. Here we define two types from multiple turns.
of strategy: emotion-focused and task-focused. An emotion- Can continually deal with emotion
focused strategy refers to a chatbot dealing only with the 3 High from the entire ongoing conversa-
user’s emotion without progressing on the target task (e.g., tion.
fixing a broken smartphone). A task-focused strategy, on the Can continually deal with emotion
Very
other hand, refers to a chatbot dealing with the user’s emo- 4 from the entire ongoing and histor-
High
tion and problem at the same time. We notice that while a ical conversation.
chit-chat chatbot only requires an emotional-focused strat- Can fully deal with emotion from
egy to deal with the user’s emotions, a customer service 5 Full the entire ongoing, historical and
chatbot should be able to solve the user’s problem while predicted conversation.
handling his or her emotions; in other words, for a task-ori- Table 7. Six-Level Scheme for Contextual-wise of ES.
ented chatbot to be considered emotionally intelligent it
should be capable of carrying out a task-focused strategy. 4.2.4 Social Norm
Table 6 shows our characterization of this dimension across Social norm refers to how well a chatbot’s strategy fit into
levels of intelligence. Thus, according to this characteriza- the customer’s social and cultural background, and how well
tion, a customer service chatbot should be at least at Level 3 it avoids biased responses: from not considering any socio-
in order to effectively embody an emotional intelligence demographic, ethnicity and racial factor to fully and world-
model of the user by leveraging the user’s emotion in order wide modeling all sensitive factors by which a user may be
to solve the task at hand. identified. Based on the premise that a chatbot should both
respect the uniqueness of each type of background and avoid
Capac- any offensive behavior, we propose our characterization of
Level Description
ity the social norm dimension across increasing levels of social
0 No No strategy. knowledge as shown in Table 8.
Can use one channel to deal with
1 Basic emotion by of one or two emotion- Capac-
focused strategies. Level Description
ity
Can use multiple channels sepa- No consideration of social and cul-
2 Partial rately to deal with emotion by sev- 0 No
tural background.
eral emotion-focused strategies. Can consider limited social and
Can use multiple channels together 1 Basic cultural background, e.g. avoiding
to deal with emotion by varied mocking/insulting customer.
3 High emotion-focused strategies and be Can consider one or two major so-
able to generate task-focused strat- 2 Partial cial and cultural backgrounds (e.g.,
egies occasionally. Western/Eastern cultures).
Can use multiple channels flexibly Can consider common social and
to deal with emotion with varied 3 High
Very cultural background world widely.
4 emotion-focused and task-focused Very Can consider most social and cul-
High 4
strategies leveraging current con- High tural backgrounds world widely.
dition and history. Can fully consider all social and
Can use all channels properly to 5 Full
cultural backgrounds.
5 Full deal with emotion leveraging cur- Table 8. Six-Level Scheme for Social Norm of ES.
rent condition and history.
Table 6. Six-Level Scheme for Variety of ES.
5 Six-Level Framework of Emotional Intelli-
4.2.3 Contextual-wise
gence for Chatbots
Table 9 shows a summary of our 6-level framework of emo-
It refers to the contextual information that the chatbot can tional intelligence. This scale ranges from Level 0, where a
utilize to deal with emotions, e.g., from using one conversa- chatbot possesses no ability related to emotional intelli-
tion turn to using the entire conversation (see Table 7). gence, to Level 5, where chatbots are said to be fully emo-
tionally intelligent. The overall logic underlying our 6-level
Capac- characterization is that the smarter a chatbot is, the more
Level Description
ity kinds of emotions it can understand, the more information it
0 No No continuity. can utilize (multiple turns, multiple channels etc.), and the
more proper strategies it can generate (from just emotion- frustration- or anger-related emotions, nor capable of exe-
focused strategies to task-focused strategies). cuting any strategy to attend the emotional state of the user
by, for example, identifying what triggered the user’s emo-
tion and what actions may alleviate this emotional state (if
Emo-
required).
tional Chatbots at Level 1, on the other hand, are considered as
Level Description
Capac- basic-intelligent chatbots since they possess basic abilities
ity to produce the corresponding behavior that takes into ac-
Cannot understand nor deal with count one or two user’s emotions such as anger or happiness.
0 No
customer emotion. This behavior is carried out through a single channel type in
Can understand and respond to a both directions to perceive the emotional signals from the
limited number of emotion types ex- user and to respond to them. Moreover, chatbots Level 1 can
1 Basic pressed through one channel type in only model users’ emotions in a single-turn basis; this im-
a one-turn basis, and deal with it by plies that they are not able to aggregate emotional signals
a limited emotion-focused strategy. throughout the whole conversation leading to temporal or
Can understand and respond to sev- local models of emotion which are independent of each
eral emotion types expressed other across turns. Nevertheless, these chatbots are able to
through multiple channel types sep- constrain their responses based on a simple model of social
2 Partial norms according to the region they are deployed at; these
arately across multiple turns, and
deal with it by several emotion-fo- responses are constrained to avoid the most common or sim-
cused strategies. ple social- or cultural-related biases such as gender biases.
Can understand and respond to ma- Even though this type of chatbot can build a very basic
jor emotion types by integrating in- model of emotions, it cannot leverage this model to solve
formation from multiple channel the target task the user expects the chatbot to solve.
types during the entire conversation, Moving one level above in our scale, we find the category
3 High of partially-intelligent chatbots—Level 2—which can be
and deal with it by varied emotion-
focused strategies while being able seen as a more advanced version of basic-intelligent chat-
to generate task-focused strategies bots. The number of emotion types which Level 2 chatbots
under some conditions. can identify and respond to increases by one or two orders
Can understand, remember and re- of magnitude, i.e. they can deal with up to 3 or 4 (even 5)
spond to most emotion types by in- types of emotion. Dealing with users’ emotions can be done
tegrating information from multiple through multiple channel types independently of each other;
Very for example, while the visual channel is dedicated exclu-
4 channel types during the entire on-
High sively to deal with happiness-related emotions, the text
going and historical conversation,
and deal with it flexibly leveraging channel is dedicated to distressful-related emotions; the
current condition and history. pairing of channel type with emotion type is selected based
Can fully understand, remember, on the feasibility to capture emotional patterns (it may be
predict and respond to all emotion easier to capture certain types of emotion by text patterns
types by integrating information than by visual cues). The information captured through the
5 Full channels cannot be integrated or aggregated into a single
from all channels, and deal with it
properly leveraging current condi- emotional model, however, by the chatbots which are thus
tion and history. required to possess different emotional strategies to deal
Table 9. Summary of the 6-Level Scheme. The descriptions with with each type of emotion separately. These strategies are
keywords “understand”,” predict” and “remember” are related to intended to modulate the users’ emotions to keep them in a
the Emotion Understanding branch; those with keywords “deal stable or neutral point that can lead to a maximum customer
with” and “strategy” correspond to the Emotion Strategy branch. satisfaction. However, the emotional strategies are carried
out independently of the target task to be solved which may
We refrain from proposing specific channel or emotion lead chatbots to be seen as partially intelligent by apologiz-
types for each level of intelligence, as well as specific tasks ing whenever a facial expression denoting anger from the
and metrics to evaluate the accuracy of chatbots’ abilities, user is recognized or by displaying grateful messages such
since these aspects may be culturally specific to the target as “thank you for patiently waiting” whenever keywords
regions where chatbots are to be deployed and may be better signaling distress from the user are read; these emotional
defined by the chatbots’ developers and designers. signals may not be related to usability or performance fac-
Level 0 includes chatbots which are neither equipped with tors such as a slow reply from the chatbot or a wrong inter-
any ability to perceive, recognize or identify any emotional pretation of the user’s query, they may just be due to the
signal from a user such as text keywords showing user’s mood, but chatbots at this level have no capacity to
integrate the user’s emotional state with the state of the task. may further alleviate the user’s emotional state while
In addition, a key feature of Level 2 chatbots is the ability to properly solving the target task, such as displaying a list of
cope with social norms; their behavior faithfully aligns to queries posed by previous users to let the current user man-
the norms of one or two cultures avoiding thus any social, ually select the closest one.
cultural or gender bias from the geographical region they are Finally, another useful feature of highly-intelligent chat-
deployed at. bots is the scope of the model of social norms; these chatbots
At the next level in our hierarchy we observe an inflec- can conform to most of the common social and cultural rules
tion point in the emotional intelligence of chatbots. Level 3 across the world.
chatbots not only increase in the range of emotion types, We consider this level in our framework to be the baseline
channel types and the scope of the context information from level of emotional intelligence; chatbots with the character-
which to extract emotional signals (the entire conversation istics described above better align to human-centric frame-
with the user) with respect to chatbots at Level 1 or 2, but works of emotional intelligence in psychology than chatbots
more importantly, they are able to integrate information and at lower levels by not only perceiving and identifying users’
inferences into a global emotional model of the user that can emotions but by applying the corresponding strategies based
be further integrated with the task’s model. Emotional sig- on such emotions to solve the target problem.
nals from diverse channels can be aggregated to resolve the The next category of chatbots in our framework are very-
emotional state of the user; for example, visual cues can be high intelligent chatbots which elaborate on top of the skills
used to infer an emotion type which may differ from the of highly-intelligent chatbots. The number of emotion types
emotion type inferred from text or vocal signals, but the to be handled increases as well as the number of social and
chatbot is able to resolve any conflict and to decide what is cultural rules to take into account. Furthermore, Level 4
the most likely emotional state. Moreover, chatbots can chatbots cannot only model the user’s emotions based on the
align this inference to the state of the task in order to resolve entire ongoing conversation, they can use previous conver-
further possible conflicts such as the causes of the user’s sations to resolve the user’s emotional state; for example,
emotion; for example, resolving that the main user’s emo- taking the case posed above, if across the entire ongoing
tion is distress requires the chatbot to resolve afterwards if conversation the chatbot is able to capture emotional de-
usability factors, such as taking too long or proposing too pendencies that when aggregated lead the chatbot to resolve
complex strategies to solve the user’s problem, are the de- for distress as the emotional state of the user, the chatbot still
terminants of this emotional state. needs to resolve for the emotional determinants such as
Resolving the conflicts described above is done while user’s baseline mood or usability factors; thus, the chatbot
considering the entire conversation with the user. Thus, an- can consult emotional models from previous conversations
other degree of freedom to be resolved by the chatbot is the to compare to what extent usability factors, for instance, lead
continuity of determinants of emotion across the conversa- to the observed emotional dependencies.
tion; for example, the chatbot may infer distress as the user’s At the top of our framework, we propose the category of
emotional state from the initial states of the interaction up to fully-intelligent chatbots which represent a second point of
the current turns due to the perceived emotional signals; thus inflexion. Chatbots at Level 5 can handle all types of emo-
the chatbot may resolve that distress is both the user’s emo- tion across all world regions while respecting all social and
tional state and the baseline mood of the user since it was cultural norms of each region. Furthermore, they can handle
identified from the initial turns and it was maintained emotions by means of any channel type. This type of chatbot
throughout the conversation; in this way, the chatbot dis- can resolve a user’s emotional state by extracting emotional
cards distress being an emotional state caused by factors dependencies from the current conversation and from previ-
from the interaction; however, while the user’s baseline ous conversations archived. Also, it can resolve for the fac-
mood may be distress, he may indeed be induced to a tors that influenced or induced the user’s emotional state
strengthen distress state by factors related to the chatbot. while trying to both alleviate such emotional state (if re-
Chatbots at this level can accurately resolve this type of quired, such as in the case of anger) and optimally solving
problems and thus recognizing that the user’s baseline mood the user’s problem. Moreover, Level 5 chatbots are able to
is distress while some interaction factors are further distress- predict the possible emotional state resulting from applying
ing him. specific emotional- and task-focused strategies; for exam-
Furthermore, this type of chatbot is able to generate and ple, if the chatbot resolves for anger as the current emotional
carry out task-focused strategies that leverage the emotional state of the user where time was the factor that influenced
state of the user while considering the determinants of this the user to achieve this state, the chatbot then will try to
state to solve the target problem; for example, if the chatbot choose for both emotional- and task-focused strategies that
is not able to understand the user’s query after several at- will alleviate the anger while reducing the time to solve the
tempts, and the user’s emotion turns to anger, then the chat- problem; therefore, choosing or generating strategies that
bot should first modulate the user’s emotion by, for instance, optimize these two variables (reducing both anger and time)
displaying an apology message, and then stop asking the is resolved, indeed, as an optimization problem where the
user to reformulate her query and change to a strategy that chatbot can predict the structures and features of the
emotional- and task-focused strategies that maximize the intelligence test (MSCEIT) good for? An evaluation using item
probability of success. In this way, we conceptualize Level response theory. PLoS One, 9(6), p. e98827.
5 chatbots as ideal dialogue systems. Thus, it may be tempt- Gunes, H. a. (2016). Is automatic facial expression recognition of
emotions coming to a dead end? The rise of the new kids on the
ing to compare the emotional capacities of this type of chat-
block. Image Vis. Comput., 55(1), 6-8.
bots to that from trained humans in customer service; but Mayer, J. D. (1999). Emotional intel-ligence meets traditional
such a comparison implies a great complexity and we leave standards for an intelligence. Intelligence, pp. 267-298.
this task as future work. Mayer, J. D. (2002). Mayer-Salovey-Caruso emotional
intelligence test (MSCEIT) item booklet. MHS Assessments.
Mayer, J. D. (2016). The ability model of emotional intelligence:
6 Discussion Principles and updates. Emotion review, 8(4), 290--300.
Mayer, J. D. (2016). The ability model of emotional intelligence:
This paper aims to provide valuable suggestions regarding
Principles and updates. Emotion Review, 290-300.
emotional intelligence in the entire lifecycle of a chatbot in- Mehrabian, A. (1980). Basic dimensions for a general
cluding design, implementation, evaluation, and mainte- psychological theory: Implications for personality, social,
nance. At the design stage, this paper could support stake- environmental, and developmental studies. Oelgeschlager, Gunn
holders to draw a feasible blueprint the industry based on \& Hain Cambridge, MA.
factors derived from our framework. During the implemen- Miller, A., Feng, W., Fisch, A., Lu, J., Batra, D., Bordes, A., . . .
tation and evaluation stage, our framework could be used as Weston, J. (2018). ParlAI: A Dialog Research Software Platform.
a look-up table to assess the current intelligence level of a arXiv:1705.06476. arXiv.
target chatbot to guide the overall development schedule. At Miners, C. T., Côté, S., & Lievens, F. (2017). Assessing the
Validity of Emotional Intelligence Measures. Emotion Review,
maintenance stage, after the launch of a target chatbot, this
1(9).
paper could continually support to offer insights about next- Papadopoulos, D., Gkintoni, E., Halkiopoulos, C., &
generation design. Antonopoulou, H. (2018). A Computational Approach of
Apart from the industrial areas, this paper could also in- Consumer Decision Making Process Emotional Intelligence and
spire academic researchers to develop a quantitative toolkit Their Effects on E-Marketing. 6th International Conference on
to evaluate the overall intelligence performance of chatbots, Contemporary Marketing Issues (ICCMI)(2018: Athens, Greece)
as well as to tackle the difficulties that stop us from achiev- 6th International Conference on Contemporary Marketing Issues
ing smarter chatbots. (ICCMI): June 27-29, 2018, Athens, Greece/Co-Organized by
Even though this scale is not in its final version, as it will Alexander Technological Educational Institu, (p. 166).
Picard, W. R. (2000). Affective Computing. MIT Press.
nurture from future research from the AI and Psychology
Roger, D., & Najarian, B. (1989). The construction and validation
fields, it is self-contained, and it comes from our vision after of a new scale for measuring emotion control. Personality and
working at the trenches deploying chatbots. Individual Differences, 10(8), pp. 845-853.
In the future, we will include more factors that contribute Rosete, D., & Ciarrochi, J. (2005). Emotional intelligence and its
to the emotional intelligence of chatbots; for example, we relationship to workplace performance outcomes of leadership
will present a quantitative method to evaluate each factor’s effectiveness. . Leadership & Organization Development Journal.
performance. In addition, a case study will be performed to Salovey, P. G. (2005). The Science of Emotional Intelligence.
illustrate how to use our framework to get insights about in- Current Directions in Psychological Science, pp. 281-285.
telligent chatbot design and development. Salovey, P., & Mayer, J. D. (1990). Emotional intelligence.
Imagination, Cognition and Personality, 9(3), pp. 185--211.
Schutte, N. S., Malouff, J. M., Hall, L. E., Haggerty, D. J., Cooper,
J. T., Golden, C. J., & Dornheim, L. (1998). Development and
References validation of a measure of emotional intelligence. Personality and
Amershi, S., Weld, D., Vorvoreanu, M., Fourney, A., Nushi, B., Individual Differences, 25(2), pp. 167-177.
Collisson, P., . . . Horvitz, E. (2019). Guidelines for Human-AI Sedoc, J., & Ungar, L. (2020). Item Response Theory for Efficient
Interaction. CHI '19: Proceedings of the 2019 CHI Conference on Human Evaluation of Chatbots. Proceedings of the First Workshop
Human Factors in Computing Systems. Association for Computing on Evaluation and Comparison of NLP Systems (pp. 21-33).
Machinery. Association for Computational Linguistics.
Bar-On, R. (1997). The emotional intelligence inventory (EQ-I): Sedoc, J., Ippolito, D., Kirubarajan, A., Thirani, J., Un-gar, L., &
Technical manual. Toronto, Canada: Multi-Health Systems. Callison-Burch, C. (2019). ChatEval: A Tool for Chatbot
Brackett, M. A., Mayer, J. D., & Warner, R. M. (2004). Emotional Evaluation. Proceedings of the 2019 Conference of the North
intelligence and its relation to everyday behaviour. Personality and American Chapter of the Association for Computational
Individual Differences,, 36(6), pp. 1387-1402. Linguistics (Demonstrations) (pp. 60-65). Association for
Chaves, A. P., & Gerosa, M. A. (2019). How should my chatbot Computational Linguistics.
interact? A survey on human-chatbot interaction design. Walter, S. a. (2011). Multimodal emotion classification in
arXiv:1904.02743. arXiv. naturalistic user behavior. International Conference on Human-
Ekman, P. E. (1994). The nature of emotion: Fundamental Computer Interaction (pp. 603--611). Springer.
questions. Oxford University Press.
Fiori, M., Antonietti, J.-P., Mikolajczak, M., Luminet, O.,
Hansenne, M., & Rossier, J. (2014). What is the ability emotional