Emulating Empathetic Behavior in Online Service Encounters with Sentiment-Adaptive Responses: Insights from an Experiment with a Conversational Agent

Abstract

Conversational agents currently attract strong interest for technology-based service provision due to increased capabilities driven by advances in machine learning and natural language processing. The interaction via natural language in combination with a human-like design promises service that is always available, fast, and with a consistent quality and at the same time resembles a human service encounter. However, current conversational agents exhibit the same inherent limitation that every interactive technology has, which is a lack of social skills. In this study, we make a first step towards overcoming this limitation by presenting a design approach that combines automatic sentiment analysis with adaptive responses to emulate empathy in a service encounter. By means of an experiment with 112 participants, we evaluate the approach and find empirical support that a CA with sentiment-adaptive responses is perceived as more empathetic, human-like, and socially present and, in particular, yields a higher service encounter satisfaction.

Full text

This is the author’s version of a work that has been published in the following outlet:
Diederich, S.; Janßen-Müller, M.; Brendel, A.B.; Morana, S. (2019): Emulating Empathetic Behavior in
Online Service Encounters with Sentiment-Adaptive Responses: Insights from an Experiment with a
Conversational Agent, Proceedings of International Conference on Information Systems (ICIS), Munich,
Germany.
Chair of Information Management
Smart Mobility Research Group (SMRG)
Prof. Dr. Lutz M. Kolbe
Platz der Göttinger Sieben 5
37073 Göttingen – Germany
www.uni-goettingen.de/im
Dr. Alfred Benedikt Brendel
Humboldtallee 3
37073 Göttingen – Germany
www.uni-goettingen.de/smrg
Please note: The copyright is owned by the author and / or the publisher.
Commercial use is not allowed.
This work is licensed under a Creative Commons Attribution-NonCommercial-
NoDerivatives 4.0 International License.

Emulating Empathetic Behavior of Conversational Service Agents
Fortieth International Conference on Information Systems, Munich 2019 1
Emulating Empathetic Behavior in Online
Service Encounters with Sentiment-Adaptive
Responses: Insights from an Experiment
with a Conversational Agent
Completed Research Paper
Stephan Diederich
University of Göttingen
Humboldtallee 3, 37073
Göttingen, Germany
stephan.diederich@
stud.uni-goettingen.de
Max Janßen-Müller
University of Bremen
Bibliothekstraße 1, 28359
Bremen, Germany
maxjan@
uni-bremen.de
Alfred Benedikt Brendel
University of Göttingen
Humboldtallee 3, 37073
Göttingen, Germany
abrende1@uni-goettingen.de
Stefan Morana
Karlsruhe Institute of Technology
Fritz-Erler-Straße 23, 76131
Karlsruhe, Germany
stefan.morana@kit.edu
Abstract
Conversational agents currently attract strong interest for technology-based service
provision due to increased capabilities driven by advances in machine learning and
natural language processing. The interaction via natural language in combination with
a human-like design promises service that is always available, fast, and with a
consistent quality and at the same time resembles a human service encounter. However,
current conversational agents exhibit the same inherent limitation that every
interactive technology has, which is a lack of social skills. In this study, we make a first
step towards overcoming this limitation by presenting a design approach that combines
automatic sentiment analysis with adaptive responses to emulate empathy in a service
encounter. By means of an experiment with 112 participants, we evaluate the approach
and find empirical support that a CA with sentiment-adaptive responses is perceived as
more empathetic, human-like, and socially present and, in particular, yields a higher
service encounter satisfaction.
Keywords: Conversational agent, empathy, interactive service technology, social
response theory, anthropomorphic design
Introduction
Emerging technologies, in particular driven by machine learning and natural language processing,
continue to rapidly transform customer interactions (Marinova et al. 2017) and the service interface
evolves from being human-driven to becoming technology-dominant (Larivière et al. 2017). Innovating
and automating service provision through interactive, natural language technology currently attracts
strong interest in theory and practice alike (Wünderlich and Paluch 2017). So-called conversational agents
(CAs), defined as software with which users interact through natural language, are explored at the service

Emulating Empathetic Behavior of Conversational Service Agents
Fortieth International Conference on Information Systems, Munich 2019 2
interface, for example to provide customer service (Gnewuch et al. 2017; De Keyser et al. 2019), support
sales (Chattaraman et al. 2012) or offer financial advisory (Dolata et al. 2019). In the context of
technology-enabled service systems (Marinova et al. 2017), CAs as a technological component have the
potential to provide service that is always available, easy to use, and faster than human customer service
yet offers the feeling of a personal contact (Verhagen et al. 2014). Recent success stories by the railroad
company Amtrak, where the CA “Julie” answered over 5 million customer questions in its first year
(NextIT 2018), or the airline KLM, where “BlueBot” helps to respond to increasing volumes of service
requests and reduces workload of the hotline (Vogel-Meijer 2018), underline the potential of CAs for
service provision. From a theoretical perspective, CAs are a particular interesting phenomenon as the
human-like characteristics of such agents, such as the interaction via natural language, trigger social
responses by users (Nass et al. 1994; Reeves and Nass 1996). As the human-like design of CAs can
contribute to a positive perception and service experience (Gnewuch et al. 2018), different studies aim to
increase the human-likeness, for example with the disclosure of artificial feelings or with the
representation of the CA with a name or gender (Cowell and Stanney 2005).
Despite this potential, present CAs share the same inherent limitation of any other service technology: A
lack of human social skills, which in turn limits their potential for service provision (Larivière et al. 2017;
Yan et al. 2013). In particular a lack of empathy, comprising the ability to recognize a customer’s affective
state and respond in an appropriate manner in a service encounter, poses restrictions for technology-
based service design (Larivière et al. 2017). Empathy is often seen as a basis of social cooperation and
prosocial behavior (Leite et al. 2013) and can contribute to service encounter satisfaction (Yan et al.
2013). Yet, the limited social capabilities of interactive service technology have contributed to the
assumption that empathy is reserved exclusively for human service provision (Frey and Osborne 2017;
Yan et al. 2013). In our study, we challenge this assumption and explore the design and perception of
empathetic service agents with the following question (RQ): How does empathetic behavior of a
conversational agent in a service context affect customer perception compared to a non-empathetic
agent? We simulate empathy in a two-step approach by using automatic sentiment analysis to
approximate a user’s emotional state and providing sentiment-adaptive responses in real-time during a
service encounter. By means of an experiment with a CA in a customer service context and 112
participants, we assess the impact of sentiment-adaptive responses on user perception of the CA.
Our research makes three contributions: First, we make an initial step towards overcoming social limits of
current service technologies by proposing a basic design approach for sentiment-adaptive conversational
service agents. Second, our study sheds light on the role of empathy in service provision with interactive
technology, which was considered to be primarily relevant for human service provision in the past. Third,
we contribute to human-like CA design by presenting empathy as a social cue that leads to a more human-
like perception of conversational agents. We continue by providing an overview of related work on CAs in
general as well as in a service context, and describe the role of empathy in service encounters. Then, we
introduce our research design and present the results of our experiment. We afterwards discuss the
implications of these results for designing CAs in service encounters, describe limitations of our work and
propose opportunities for future research in this area.
Related Work and Theoretical Background
The idea to interact with software through natural language dates back decades to the 1960s, when the
first CA, called ELIZA, was developed by Joseph Weizenbaum (1966). While these early CAs executed
simple pattern matching to provide responses to users (McTear 2017), CAs now exhibit significantly
improved capabilities, in particular due to advances in natural language processing and machine learning
(Følstad and Brandtzæg 2017). With these increased capabilities, they currently attract significant interest
in research (McTear 2017) as well as practice (Oracle 2016) and a variety of different CAs emerged in the
last few years (Diederich, Brendel, and Kolbe 2019). Present-day CAs can be distinguished by three main
dimensions: the mode of interaction, the embodiment and their application context. With regard to the
interaction mode, users can interact with CAs via spoken language, often referred to as digital or voice
assistants, or written text, often described as chatbots, or both (Schroeder and Schroeder 2018).
Regarding the embodiment, or the representation of the CA (Seeger et al. 2018), CAs can have a physical
embodiment, such as service robots, a virtual interactive or static avatar (Wünderlich and Paluch 2017),
or be disembodied. Finally, CAs can be able to converse about a variety of topics or fulfill different tasks,

Emulating Empathetic Behavior of Conversational Service Agents
Fortieth International Conference on Information Systems, Munich 2019 3
such as Cleverbot (2018) or be domain-specific (Gnewuch et al. 2017), such as for customer service
(Verhagen et al. 2014), financial advisory (Dolata et al. 2019) or team collaboration (Elson et al. 2018). In
the context of this study, we focus on a text-based CA with a virtual embodiment for customer service.
Conversational Agents in the Context of Service Systems
As the capabilities of CAs continue to improve, their application in service systems becomes increasingly
versatile. While customer service is one of the primary application areas at the moment where CAs can
fulfill requests like providing product information or handling complaints (Gnewuch et al. 2017), they can
be expected to support or even fully assume tasks currently performed by human service personnel
(Marinova et al. 2017; Verhagen et al. 2014), such as financial advisory (Dolata et al. 2019). Against this
background, CAs that use machine learning to continuously improve their performance over time meet
recent calls from service researchers, such as Beverungen et al. (2017) or Larivière et al. (2017), to explore
how different types of artificial intelligence can be embedded in service systems. As technological
components of service systems, which can be defined as “configurations of people, technologies, and other
resources that interact with other service systems to create mutual value” (Maglio et al. 2009, p. 395), CAs
have the potential to bridge the gap between service technology (which is always available yet lacks a
feeling of personal interaction), such as self-service online portals, and human service employees (who
offer a personal contact to a service provider yet have limited availability). In particular, CAs with a
human-like design can elicit feelings of social presence and personalization that are not present in typical
online service provision (Verhagen et al. 2014). Furthermore, conversational agents have the potential to
provide a steady level of service quality as well as efficiency and thus overcome inherent human
performance variability (Larivière et al. 2017). Overall, the evolving capabilities of CAs will require service
providers to evaluate which types of service can be better provided by human employees, (self-)service
technologies or conversational agents and to reflect on the optimal human-technology mix for service
provision to foster an appealing customer journey (Larivière et al. 2017).
Empathy in Service Encounters
In the context of service encounters, empathy can be understood as the caring, individualized attention
that a firm provides its customers (Parasuraman et al. 1985). In a human-to-human service encounter,
empathy can be expressed both in verbal communication, for example by indicating understanding of a
customer’s request and feelings, as well as in non-verbal communication, such as with nodding or through
frequent eye contact (Gabbott and Hogg 2004). In general, empathy is considered to influence the
perceived service quality alongside the dimensions responsiveness, assurance, reliability and tangibles
(Bolton and Drew 2002; Parasuraman et al. 1985, 1988). A prevalent assumption of technology in a
service encounter is that it dehumanizes the service interaction as it “cannot provide the empathy a
human agent can provide (Yan et al. 2013, p. 7). Understanding a conversation partners emotional state
and taking it into account during the encounter is considered to be one of the main factors that
differentiate service provision by humans and technology (Larivière et al. 2017). While service provision
through technology, such as in the form of self-service portals, has advantages regarding service
availability and fulfillment speed, it is usually associated with a lack of personal and emotional interaction
due to the social limits of technology (Frey and Osborne 2017). Hence, we suggest to make a first step
towards overcoming the social limits of interactive service technology by emulating empathetic behavior
of a CA through the use of sentiment analysis and sentiment-adaptive responses in a service encounter.
Social Response Theory and Anthropomorphic CA Design
Recent IS research on CAs draws on Social Response Theory to study user-CA interaction (Diederich,
Brendel, Lichtenberg, et al. 2019) and inform the design of such artifacts (Gnewuch et al. 2017, 2018).
Social Response Theory posits that humans mindlessly apply social rules and expectations to anything,
including computers, which demonstrates human-like traits or behavior (Nass and Moon 2000; Reeves
and Nass 1996). Nass and Moon (2000) explain how humans overuse social categories (e.g. gender) and
social behaviors (e.g. reciprocity), hypothesizing that “the more computers present characteristics that are
associated with humans, the more likely they are to elicit social behavior” (Nass and Moon 2000, p. 7).
Due to the fact that CAs usually exhibit a variety of social cues (Feine, Gnewuch, et al. 2019), including
turn-taking, the use of self-references or having a name, they trigger substantial social responses in users.

Emulating Empathetic Behavior of Conversational Service Agents
Fortieth International Conference on Information Systems, Munich 2019 4
These social responses pose opportunities for human-like CA design, yet at the same time foster high user
expectations as users would rather compare the CA to a human customer service employee than to a
computer system (Følstad and Brandtzæg 2017). Different studies have stated that a higher degree of
perceived humanness is often associated with a positive impact on other factors, such as trustworthiness
(Schroeder and Schroeder 2018), perceived competency (Araujo 2018), authenticity (Wünderlich and
Paluch 2017), or service encounter satisfaction (Gnewuch et al. 2018) despite potential feelings of
uncanniness and eeriness when interacting with a human-like technological artifact (MacDorman et al.
2009). Consequently, many scholars explore different approaches to design CAs as human-like as
possible. For the design of human-like CAs, Seeger et al. (2018) provide a conceptual framework which
distinguishes three dimensions: Human identity, nonverbal communication, and verbal communication.
In this framework, human identity includes cues which serve to identify a human being in computer-
mediated communication, such as a name or avatar (Cowell and Stanney 2005). Nonverbal
communication refers to any expression of emotional states not directly conveyed in the language itself,
such as blinking dots to simulate thinking and typing a response or the use of emoticons (Wang et al.
2008) to express emotions. Third, verbal communication includes the spoken or written sentences itself
and cues, such as greetings, small talk and anecdotes (Bickmore and Picard 2005). In the context of this
study, we investigate the effect of a verbal social cue (affective responses) to make CAs display empathy in
a human-CA service encounter.
Hypotheses
While various studies have been conducted on empathy in human-computer interaction (e.g. Leite et al.
(2013) or Niewiadomski and Pelachaud (2010)) and human service encounters (e.g. Yan et al. (2013)), to
the best of our knowledge little empirical research has been conducted on how the empathetic behavior of
an artificial virtual agent is perceived by a user in a service context and whether it affects service
encounter satisfaction. Different studies on CA design emphasize that CAs should have the capacity to
adequately display emotions (e.g. McQuiggan and Lester (2007)). However, we still lack an understanding
regarding the impact of synthetically displaying empathy in human-CA service encounters. As a first step
towards designing interactive service technology that is capable of understanding and adequately showing
affective communication behavior, we propose the combination of automatic sentiment analysis and
adaptive responses to emulate empathy. Sentiment analysis makes use of the fact that texts written by
humans convey valuable information about their emotional state (Liu 2012) and are capable of accurately
extracting positive or negative polarity in written text (Liu 2010). Based on the extracted polarity, a CA
can provide tailored responses and thus simulate empathy. Hence, we first hypothesize as follows:
H1: A CA that provides sentiment-adaptive responses in a service encounter yields a higher level of
perceived empathy than a CA that sends static responses independent of the sentiment.
In the context of human-like design, researchers explore the use of various cues that are intended to
contribute to the human-likeness of CAs. For example, Gnewuch et al. (2018) propose an approach to
dynamically delay response times of an agent to simulate thinking and typing of replies and find that such
delays lead to a higher perceived human-likeness than immediate responses in a service encounter. Thus,
the authors suggest that dynamic response delay represents a cue for human-like design as it can
contribute to simulating human communication behavior. Similarly, Araujo (2018) finds that giving a CA
a human name and using personalized greetings contribute to perceived humanness. In line with these
studies, we consider the display of empathy by understanding a person’s affective state and reacting
adequately as an essential part of human communication that significantly influences how we perceive
each other (Davis 2015). Thus, we suggest that empathetic communication contributes to perceived
humanness as follows:
H2: A CA that provides sentiment-adaptive responses in a service encounter yields a higher level of
perceived humanness than a CA that sends static responses independent of the sentiment.
Social cues in the form of avatars or emotions have been found to not only influence the perceived human-
likeness of an agent, but to also have an impact of the social presence one feels when interacting with
technology (Gefen and Straub 2004). Social presence was originally understood as “the degree of salience
of the other person in a mediated communication and the consequent salience of their interpersonal
interactions” (Short et al. 1976, p. 65) and has been shown to likewise exist without actual human contact

Emulating Empathetic Behavior of Conversational Service Agents
Fortieth International Conference on Information Systems, Munich 2019 5
(Gefen and Straub 2004). The concept has been studied in different contexts, such as avatars (Von Der
Pütten et al. 2010) or virtual agents (Qiu and Benbasat 2009), with researchers discovering different cues,
such as human photos or the use of emotional statements directed towards the user, equally contribute to
perceptions of social presence (Hassanein and Head 2007). In line with our view of empathy as a social
cue, we argue that empathetic responses convey a feeling of higher social presence than static statements:
H3: A CA that provides sentiment-adaptive responses in a service encounter yields a higher level of
social presence than a CA that sends static responses independent of the sentiment.
Finally, different studies on the role of empathy in human-CA interaction, often with physically or
virtually embodied agents, highlight positive effects on further factors. For example, Bickmore and Picard
(2005) and Liu and Picard (2004) describe positive effects of suitable expressions of emotion on
perceptions of persuasiveness and credibility. Similarly, Brave et al. (2005) found that an agent that is
perceived as empathetic is considered to be significantly more caring, likable, trustworthy and supportive.
Overall, these studies indicate that adequate displays of empathy by an artificial agent contribute to a
better perception of the agent by its user. In the context of human service encounters, empathy has been
shown to be associated with greater perceptions of service quality and higher customer satisfaction
(Caruana et al. 2000; Mohr and Bitner 1991; Price et al. 1995). Service employees that display empathy
are perceived as approachable (Parasuraman et al. 1985), caring (Johnston 1995), and additionally make
an effort to understand consumers’ needs (Wels-Lips et al. 1998). In addition, Feine, Morana, et al. (2019)
find a significant correlation between subjectively measured service encounter satisfaction and scores
from a sentiment analysis of customer dialogues, underlining the potential of sentiment detection in the
context of service provision with CAs. Thus, we hypothesize that an empathetic CA fosters a comparable
effect in a human-technology encounter with regard to the overall satisfaction with the provided service.
H4: Customers are more satisfied with a service encounter when the CA gives sentiment-adaptive,
empathetic responses than with a CA that gives static, sentiment-independent responses.
Figure 1 visualizes the hypotheses regarding the impact of sentiment-adaptive responses on perceived
empathy, perceived humanness, social presence, and service encounter satisfaction in our model.
Figure 1. Research Model
Method
We tested the four hypotheses in an online experiment that took place over a span of two months at the
beginning of 2019. For the experiment, we selected a between-subjects design to prevent carryover effects
(Boudreau et al. 2001) and chose a customer service context with a fictitious telecommunications
company to provide participants with an intuitively understandable service encounter setting. In the
following, we describe our data collection procedure and sample, the two experimental configurations
with a focus on the design of sentiment-adaptive responses, and the measures as well as underlying items
used in the survey after the interaction with the virtual service agent.

Emulating Empathetic Behavior of Conversational Service Agents
Fortieth International Conference on Information Systems, Munich 2019 6
Data Collection Procedure and Sample
At the beginning of the experiment, the participants received a briefing website in which the context
(service encounter), the structure of the experiment (interaction with CA as service agent and subsequent
survey) as well as the experimental task were explained. We provided every participant with the exact
same sets of information for the experiment (Dennis and Valacich 2001) and participants were assigned
randomly to the control or treatment configuration. Every participant was presented with a fictitious
mobile phone invoice and asked to contact the customer service chatbot of the company, find out why the
invoice amount was much higher than usual and afterwards file a complaint for an incorrect invoice.
After completing the task, the participants received a link to the online survey from the agent. In total, the
experiment took around 10 minutes per participant. With an a priori power analysis using G*Power (Faul
et al. 2007), we determined a sample size of at least 102 subjects (effect size = .50, alpha = .05,
power = .95). We recruited 112 participants via personal networks, comprising mainly information
systems students, and mailing lists, who were not compensated for their participation. We identified six
invalid responses with inverted control questions and excluded them from the analysis as participants
provided straight line answers, thus reducing the sample size from 112 to 106 participants. The
participants’ age ranged from 17 to 68 years (mean: 25.6 years) and our sample had a share of 32% female
persons. Most of the participants indicated that they never use digital assistants, such as Siri or Alexa (n =
57) and indicated that they never use chatbots (n = 70).
Control and Treatment Configurations
We prepared two instances (control instance with static responses, treatment instance with sentiment-
adaptive responses) of one CA using the design platform for natural language software Dialogflow by
Google (2019). We further implemented a custom-built web interface to provide convenient access to the
CAs and minimize distraction. Both CAs received the same training phrases, i.e. exemplary statements
which customers might make during the service encounter that indicate a user’s intent and trigger a reply.
The CAs were able to process different variations of sentences with the same meaning and could extract
parameters, such as invoice numbers, and use them throughout the dialogue for paraphrasing.
Furthermore, both CAs were designed to state that they did not understand a user’s request at one point
in the service encounter to explicitly provoke the user in addition to the expensive mobile phone invoice.
The CAs received different cues for human-like CA design according to the three dimensions (human
identity, verbal, non-verbal) as suggested by Seeger et al. (2018) to establish a baseline for perceived
humanness and social presence: With regard to the human identity, we equipped the CA with the name
"Sarah" (Cowell and Stanney 2000), a female gender (Nunamaker et al. 2011) and a comic avatar
representing a female customer service employee. Concerning verbal communication, the CA was
designed to use self-references, such as “I fully understand…”, (Sah and Peng 2015), turn-taking and a
personal introduction (“Hello, my name is Sarah and I am part of the customer service team. How can I
help you?”) as well as polite greeting in the form of a welcome message (Cafaro et al. 2016). Regarding the
non-verbal human-like CA design dimension, we implemented blinking dots in combination with
dynamic response delays depending on the length of the preceding message as suggested by Gnewuch et
al. (2018) to simulate thinking and typing of replies by the CAs.
Overall, both instances of the CA were identical except for the use of sentiment analysis and adaptive
replies (Figure 2).The treatment instance additionally detected the sentiment of a user statement in the
background using the automatic sentiment analysis provided by Aylien (2019) and adapted its response to
the detected sentiment at three points in the encounter: When the user describes the service request (1),
after the CA presents the elements of the invoice (2), and, finally, when the participant complaints about
the subscription for mobile games included in the invoice amount (3). If the automatic sentiment analysis
determined a positive or neutral sentiment of the participant’s message, the treatment instance used the
same responses as the control instance. In case of a negative sentiment, the agent added an empathetic
statement before the response at the respective interaction point. The empathetic responses aimed to
simulate empathetic behavior that a human service agent would most likely show by indicating
understanding of the customer’s affective state and politely apologizing for the inconvenient situation.

Emulating Empathetic Behavior of Conversational Service Agents
Fortieth International Conference on Information Systems, Munich 2019 7
Figure 2. Web Interface with Control (static, left) and Treatment (adaptive, right)
Configurations (material translated to English)
The architecture used in this experiment consisted of four components (Figure 3) that were connected via
webservices. Once the user entered a query in the web interface it was transmitted to the agent
implemented on Google Dialogflow. Dialogflow then detected the user’s intent from the statement based
on a set of predefined intents (e.g. file invoice complaint) and extracted parameters (e.g. invoice number)
that were for example used for paraphrasing by the agent. Next, Google Dialogflow forwarded the
detected intent name, parameters, and original user statement to a custom-built fulfillment component.
If Dialogflow was not able to successfully detect a user intent from a statement, it directly sent a context-
specific fallback response back to the web interface (e.g. “Unfortunately, I did not understand your
response. Please provide the invoice number so we can take a look at the amount for last month”). In case
of successful intent detection, the fulfillment component then transmitted a request for sentiment
analysis to the external text analytics provider Aylien (2019) and received information on the polarity
(negative, neutral, positive) of the original user statement. The fulfillment component then selected the
agent’s response for this intent and identified polarity (static response in case of positive or neutral
polarity, empathetic response in case of negative polarity) and forwarded it to Dialogflow. Finally,
Dialogflow sends the response to the Web Interface, which displayed it to the user after the dynamic
response delay and indicator (blinking dots) to simulate thinking and typing (Gnewuch et al. 2018).

Emulating Empathetic Behavior of Conversational Service Agents
Fortieth International Conference on Information Systems, Munich 2019 8
Figure 3. Agent Architecture
To ensure that the real-time sentiment analysis did not delay the CA’s responses, we tested whether the
sentiment analysis and adaptation of responses lead to higher response times than the dynamic response
delays. The test indicated that the designed delays did always exceed the time required for analyzing the
sentiment of the previous user statement. Table 1 provides an overview of the static as well as sentiment-
adaptive responses used in the interaction.
Service interaction
point
Static
response
Sentiment-adaptive response
(for negative sentiment)
Interaction point #1:
Description of
service request
“To view the invoice, please
provide your customer and
invoice numbers as well as your
date of birth for authentication”
“I understand your concern and would be
happy to take a look at your invoice.
To view the invoice, please provide your
customer and invoice number as well as
your date of birth for authentication.”
Interaction point #2:
Presentation of
invoice elements
“If you like, I can provide a
detailed list of costs for last
month.”
“Please excuse the confusion about the
invoice amount.
If you like, I can provide a detailed list of
costs for last month.”
Interaction point #3:
Complaint about
games subscription
“Our system indicates that you
subscribed to our mobile games
package on the 14th of December
2018.”
“Ok, I fully understand that this is
annoying.
Notwithstanding, our system indicates
that you subscribed to our mobile games
package on the 14th of December 2018.”
Table 1. Static and Sentiment-Adaptive Responses in the Service Encounter
Manipulation Check
After the experiment, we conducted a manipulation check to ensure that participants in the treatment
condition actually triggered and received empathetic replies. We reviewed conversation logs to identify
responses where a negative polarity was identified at the three interaction points in the service encounter.
At the first interaction point, where the participant described its request at the beginning of the service

Emulating Empathetic Behavior of Conversational Service Agents
Fortieth International Conference on Information Systems, Munich 2019 9
encounter, 41% of the responses were neutral and 59% empathetic. At interaction point two after the
elements of the incorrect invoice were presented, 14% of the agent’s responses were neutral and 86%
empathetic. Finally, at the third interaction point where the participant complains about the mobile
games subscription, the CA replied neutrally in 28% of the responses and empathetically in 72% of the
responses. Overall, around 72% of responses in the treatment instance were empathetic, i.e. the CA
detected a negative sentiment and tailored its response accordingly. Thus, the treatment CA did indeed
make use of sentiment-adaptive responses nearly three out of four times in the interaction.
Measures
Following the encounter with the virtual service agent, the participants completed a survey in which we
measured how the participants perceived the virtual agent and whether they were satisfied with the
service encounter. For the design of the questionnaire, we used established measurement instruments
from previous studies that were used in the context of technology-driven service encounters (Gnewuch et
al. 2018; Yan et al. 2013). We measured four constructs corresponding to our hypotheses: Perceived
empathy (Yan et al. 2013), perceived humanness (Holtgraves and Han 2007), social presence (Gefen and
Straub 1997) and service encounter satisfaction (Verhagen et al. 2014). All constructs were measured with
two to six items each. As suggested in the original studies, we used a 9-point semantic differential scale to
measure perceived humanness and a 7-point Likert scale to measure the constructs perceived empathy,
social presence, and service encounter satisfaction. Furthermore, we added two control questions to the
survey by inverting two items, one for perceived humanness and one for social presence to check whether
participants provided straight line responses in the questionnaire. In the final part of the survey, we
collected demographic information (gender and age), information on the frequency of use of digital
assistants (e.g. Siri or Alexa) and chatbots on websites and social media ranging from never over monthly
to weekly to daily. We further asked for free form feedback for the service encounter and the experiment.
Constructs and items
Loadings
Scale and source
Perceived empathy (
a
= .847, CR = .849, AVE = .738)
The agent gives customers individual attention.
The agent gives customers personal attention.
.860
.858
7-point Likert scale
(Yan et al. 2013)
Perceived humanness (
a
= .917, CR = .917, AVE = .689)
How human-like did you perceive the agent?
How skilled do you perceive the agent?
How thoughtful do you perceive the agent?
How polite do you perceive the agent? (dropped)
How responsive do you perceive the agent?
How engaging do you perceive the agent?
.811
.870
.825
(0.549)
.817
.826
9-point semantic
differential scale
(Holtgraves
and Han 2007)
Social presence (
a
= .861, CR = .858, AVE = .602)
I felt a sense of human contact with the agent.
I felt a sense of personalness with the agent.
I felt a sense of sociability with the agent (dropped)
I felt a sense of human warmth with the agent.
I felt sense of human sensitivity with the agent.
.777
.814
(0.490)
.729
.781
7-point Likert scale
(Gefen and
Straub 1997)
Service satisfaction (
a
= .848, CR = .851, AVE = .659)
How satisfied are you with the agent ‘s advice?
…the way the agent treated you?
…the overall interaction with the agent?
.781
.721
.920
7-point Likert scale
(Verhagen
et al. 2014)
Table 2. Constructs, Items, and Factor Loadings

Emulating Empathetic Behavior of Conversational Service Agents
Fortieth International Conference on Information Systems, Munich 2019 10
Table 2 summarizes the four constructs, items and factor loadings including Cronbach’s a, composite
reliability (CR) and average variance extracted (AVE). Two items were dropped from the analysis due
factor loadings lower than .60 as proposed by Gefen and Straub (2005). All constructs showed sufficient
Cronbach’s a (larger than .80), CR (larger than .80) and AVE (larger than .50) with respect to the levels
proposed by Urbach and Ahlemann (2010) and were thus included in the analysis.
Results
We analyzed the survey data by means of descriptive statistics and t-tests to understand the relation
between static or sentiment-adaptive agent responses and perceived empathy (H1), perceived humanness
(H2), social presence (H3), and service encounter satisfaction (H4). The analyses were carried out using
the statistical computing software R and SPSS version 25. Table 3 summarizes our results for the four
hypotheses. We first checked for homogeneity of variance and the Levene tests indicated equal variances
for perceived humanness (F = 5.20, p = 0.055), social presence (F = 1.42, p = 0.236), and service
encounter satisfaction (F = 0.244, p = 0.622). For perceived empathy, the Levene test showed unequal
variances (F = 5.20, p = 0.025). Thus, we used Student’s t-tests to analyze for differences between the
conditions regarding perceived humanness, social presence and service encounter satisfaction
(homogeneous variances) and Welch’s t-test for perceived empathy (non-homogeneous variance). All t-
tests were performed one-sided to examine whether sentiment-adaptive responses (treatment group)
positively affected customer perception of the CA in comparison to static responses (control group).
Our data reveals that participants in the treatment group perceived the CA as more empathetic (M = 4.28,
SD = 1.27) than participants in the control group (M = 3.39, SD = 1.65), t(97.4) = - 3.14, p = .001. Thus,
the data supports our first hypothesis that a CA with dynamic, sentiment-adaptive response is perceived
as more empathetic than a CA that provides static, neutral responses. Concerning perceived humanness,
the survey data indicates a significant difference between the control (M = 4.29, SD = 1.97) and treatment
groups (M = 5.35, SD = 1.49), t(104) = -3.12, p = .001, providing support for our second hypothesis that a
CA with sentiment-adaptive responses is perceived as more human-like than a CA with static replies.
Condition
t-value
(df)
p-value
Result
Control
(n = 53)
Treatment
(n = 53)
Perceived
empathy
Mean
SD
SE
3.39
1.65
0.23
4.28
1.27
0.17
-3.14
(97.4)
.001
H1
supported
Perceived
humanness
Mean
SD
SE
4.29
1.97
0.27
5.35
1.49
0.20
-3.12
(104)
.002
H2
supported
Social
presence
Mean
SD
SE
2.92
1.34
0.18
3.74
1.26
0.17
-3.23
(104)
< .001
H3
supported
Service
satisfaction
Mean
SD
SE
3.97
1.37
0.19
4.70
1.29
0.18
-2.61
(104)
.003
H4
supported
SD = Standard deviation, SE = Standard error
Table 3. Descriptive Statistics and t-Test Results
Furthermore, participants in the treatment group reported a higher social presence (M = 3.74, SD = 1.26)
than participants in the control group (M = 2.92, SD = 1.34), t(104) =-3.23, p < .001, providing support
for the third hypothesis. Finally, participants that interacted with the sentiment-adaptive CA showed a
higher service encounter satisfaction (M = 4.70, SD = 1.23) than participants that were assigned to the CA

Emulating Empathetic Behavior of Conversational Service Agents
Fortieth International Conference on Information Systems, Munich 2019 11
with non-empathetic responses (M = 3.97, SD = 1.37), t(104) = -2.61, p = 0.o03). Thus, we find support
for the fourth hypothesis, highlighting that participants who interacted with the sentiment-adaptive CA
were more satisfied with the service encounter than participants that engaged in a conversation with the
static CA in the control configuration.
Next, we conducted additional robustness tests. We analyzed for differences between the control and
treatment groups with regard to demographics. We found no significant differences in age (t(69.1) = -1.92,
p = 0.059), gender (χ²(1) = 0.17, p = 0.679), frequency of use for digital assistants like Siri or Alexa
(χ²(3) = 0.06, p = 0.996), and frequency of use for chatbots on social media or company websites
(χ²(3) = 5.81, p = 0.121) between the two groups. Overall, the results of our tests indicate that the
differences in perceived empathy, perceived humanness, social presence, and service encounter
satisfaction are not explained by differences in demographics or CA experience between the two
experimental groups. Furthermore, we conducted a multiple linear regression to investigate whether the
frequency of use for digital assistants or chatbots on websites and social media had an impact on
perceived empathy, perceived humanness, social presence, and service satisfaction yet the results were
not significant. In addition, we performed post-hoc power analyses using G*Power (Faul et al. 2007),
which suggested that all tests have sufficient power for the given sample size (powerempathy = .916,
powerperceived humannes s = .899, powersocial presence = .925, powerservice satisfaction = .836). Figure 4 visualizes the
differences between the control and treatment group with regard to the four measured constructs with the
error bars indicating the 95% confidence interval.
Figure 4. Differences for the Constructs between the Experimental Groups
We further conducted a post-hoc exploratory moderator analysis, examining whether the relation
between static or sentiment-adaptive responses and empathy, perceived humanness, social presence, and
service satisfaction was moderated by age, gender or frequency of use of digital assistants or chatbots. The
participants’ age and frequency of use were not identified as moderators. However, the interaction term
between the group and gender explained a significant increase in variance in empathy (ΔR2 = .058, F(1,
102) = 6.868, p = .010), thus the participant’s gender moderated the relationship between the
experimental group and empathy. The potential moderations by gender with regard to the relation
between the experimental group and perceived humanness (ΔR2 = .027, F(1, 102) = 3.118, p = .080),
social presence (ΔR2 = .009, F(1, 102) = 0.992, p = .322), and service encounter satisfaction (ΔR2 = .034,
F(1, 102) = 3.837, p = .053) were not statistically significant. Figure 5 visualizes the relationship between
experimental group, gender, and empathy, humanness, social presence, and service satisfaction.

Emulating Empathetic Behavior of Conversational Service Agents
Fortieth International Conference on Information Systems, Munich 2019 12
Figure 5. Mean Values Differentiated by Gender
(dotted lines indicate female participants)
Discussion
Our results provide empirical support that sentiment-adaptive responses increase perceived empathy,
perceived humanness, social presence, and service encounter satisfaction in comparison to static
responses. In the following, we discuss implications of our results for service provision with CAs and their
human-like design, state limitations of our work, and outline opportunities for future research.
Implications for CAs in Service Encounters and their Design
The results of our experiment demonstrate how detecting a user’s affective state and providing empathetic
responses contributes to a better perception of the CA and overall satisfaction with the service encounter,
even when the actual request cannot be directly fulfilled by the system. Against this background, the
real-time identification of negative sentiments in an interaction and the provision of empathetic
responses can help to reduce customer frustration and overcome social limits of current interactive
service technology (Yan et al. 2013). The fundamental idea of our sentiment-adaptive design, comprising
the combination of sentiment analysis and adaptive responses, can be adapted by practitioners
implementing CAs for service provision and its components (e.g. provider for sentiment analysis or CA
platform used for intent detection) can be exchanged to meet different organizational requirements, such
as for architectural standards or concerning data protection. In addition, the automatic identification of
sentiment in the dialogue can be used not only to provide tone-aware responses, but could also enable a
better transition of the service encounter with the virtual agent to human service personnel by forwarding
service requests of potentially frustrated customers. As the design of CAs that are able to deal with a high
variety of user input remains a challenging endeavor (Følstad and Brandtzæg 2017), it can be expected
that CAs will have to forward service requests, particularly the ones that are more complex or did not
occur frequently in the past, to human service personnel. Thus, a real-time sentiment analysis of the
human-CA dialogue can not only be used to increase customer satisfaction with the service encounter, but
also opens up further options for designing service encounter transitions in a way that remains appealing
and convenient for the customer (Lemon and Verhoef 2016). Future CAs for service provision could for
example apply automatic compensation strategies in case a frustrated user is recognized, similar to
compensation strategies of human service employees nowadays.

Emulating Empathetic Behavior of Conversational Service Agents
Fortieth International Conference on Information Systems, Munich 2019 13
Concerning the human-like design of such service agents, our experiment shows that users indeed
anthropomorphize CAs in a service context in line with Social Response Theory (Nass and Moon 2000;
Reeves and Nass 1996), highlighting the potential of this technology to (partially) overcome current
limitations of human service employees (availability, consistent quality, response time) and self-service
technologies (lack of personal contact). While empathy was considered in the past to be primarily relevant
for human service encounters and not for service provision via technology due to its inherent social limits
(Yan et al. 2013), our results indicate a necessity to rethink how we design and position technology in a
service context and emphasize the potential of human-like CA design (Seeger et al. 2018). In the context
of continuously increasing capabilities of CAs, we expect that thinking carefully about the role of CAs
between existing interactive service technology and human service provision will become even more
relevant in the future, urging managers to reflect on which services are better provided by humans and
which by different types of seemingly intelligent and interactive technology (Scherer et al. 2015).
Finally, our analysis of gender-related moderation effects highlights the importance of empathy in
human-CA interaction in particular for female customers. According to our findings, empathetic
communication seemed to be relatively more important for women than for men regarding perceptions of
empathy and humanness of the service agent as well as with regard to service satisfaction. Female
participants in the control group perceived the agent as less empathetic and human-like and were less
satisfied with the service encounter than male participants. However, female participants in the treatment
group rated the CA as more empathetic and human-like and indicated a higher service satisfaction than
male participants in the same experimental group (Figure 5). These results are in line with studies on
human service provision. For example, Chiu (2002) as well as Chiu and Wu (2002) investigated gender-
related differences between the affective and cognitive components of attitude towards service quality
(Zajonc and Markus 2002). In this distinction, cognitive components refer to “what we know” whereas the
affective components describe “what we feel” for a particular product or service (Zanna and Rempel
1988). Their findings suggest that the affective component is perceived as more important for service
quality than the cognitive component for female customers while there are no differences in the relative
importance these components of service quality for male customers (Chiu 2002). Thus, our results
suggest that existing findings concerning empathy in human service encounters hold true for service
provision with human-like CAs. Consequently, CA designers could explore varying interaction styles to
account for the gender-related differences in communication and perception of service quality.
Limitations and Opportunities for Future Research
Our work is not free of limitations. First, our exploratory design with the automatic sentiment analysis
and sentiment-adaptive verbal responses for statements with a negative polarity, offers opportunities for
future design-oriented CA studies in a service context. While our study places emphasis on detecting and
addressing negative emotions in the service encounter, future studies can explore whether positive
feelings in an interaction can be reinforced by adaptive CA behavior as well, thus contributing to a better
perceived service experience. Furthermore, the empathetic responses in our experiment were exclusively
verbal without the use of non-verbal communication. Exploring the use of sentiment-adaptive non-verbal
communication, such as facial expressions of a virtually embodied service agent or the use of emoticons
(Beale and Creed 2009) to express empathy, represents a promising research endeavor.
In addition, our experiment was conducted in an online setting where we benefitted from control yet
lacked realism (Dennis and Valacich 2001). We provided the participants with a set of rather specific
tasks, which in turn allowed for a better design of the communication behavior of the CA as the variation
of user input was limited. Thus, we propose future research to investigate the design of sentiment-
adaptive responses of CAs in the field where user input exhibits a potentially higher variability and
fallback responses or transitions between service encounters with CAs and human personnel need to be
carefully designed to ensure a high-quality user experience. Furthermore, our research design and sample
did not account for cultural differences among participants that could moderate the relation between
sentiment-adaptive responses and the four measured constructs. Future studies could explore a culturally
informed design as for example suggested by Pereira and Baranauskas (2015) to account for cultural
differences related to the importance and display of empathy. Finally, the CA that we designed with the
selected set of human-like cues was perceived positively by the participants. However, recent research on
CA design indicated potentially unintended outcomes of combining human-like design cues to maximize
humanness (Seeger et al. 2018), such as a high degree of uncertainty concerning the human or

Emulating Empathetic Behavior of Conversational Service Agents
Fortieth International Conference on Information Systems, Munich 2019 14
technological nature of the service agent or unrealistic expectations towards the agent’s capabilities
(Wünderlich and Paluch 2017). Similarly, Schroeder and Schroeder (2018) found that increasing human-
likeness does not linearly increase with trust, but that users may feel threatened at one point. In this
context, the theory of Uncanny Valley (Mori 1970) describes a sharp drop of affinity for a human-like
object, where the attention of a user shifts from the human-like qualities to the aspects that seem to be
inhuman (MacDorman et al. 2009). Thus, we suggest that future studies investigate the risks associated
with increasing anthropomorphism by combining sentiment-adaptive responses with other social cues,
such as a representation by means of an interactive avatar and name, to determine if and where a human-
like CA could elicit negative responses (e.g. feelings of eeriness or uncanniness) in service encounters.
Concluding Remarks
Interactive service technology in the form of CAs promises to provide service that is always available,
easy-to-use and resembles human-to-human service encounters. Yet, the social capabilities of current CAs
with regard to recognizing and adequately reacting to a customer’s affective state are limited. To make a
first step towards overcoming the social limits of conversational service agents, we proposed and
evaluated the use of sentiment-adaptive responses in a customer service context. Our findings provide
empirical evidence that a sentiment-aware CA leads to a higher perceived empathy, perceived humanness,
social presence, as well as service encounter satisfaction in comparison with CAs operating with static
responses. In addition, our results suggest that existing findings from human service provision concerning
gender-related differences between the importance of cognitive and affective components for perceived
service quality hold true for service provision with interactive technology, offering new design options for
increasing service satisfaction through varying communication behavior.
Our study contributes to designing CAs as an emerging, innovative technology for service provision in
three ways: First, we suggest a design approach for emulating empathy in conversational service agents
through the combination of automatic sentiment analysis and sentiment-adaptive verbal responses.
Second, our results place emphasis on the potential of affective communication in interactive service
technologies, which was primarily considered to be relevant for human service encounters in the past, as
it contributes perceived empathy, perceived humanness, social presence, and, in particular, service
encounter satisfaction. Finally, our results add to the growing knowledge base on human-like CA design
by displaying how empathy represents a social cue, which can effectively contribute to a more human-like
perception of such virtual agents in service encounters.
References
Araujo, T. 2018. “Living up to the Chatbot Hype: The Influence of Anthropomorphic Design Cues and
Communicative Agency Framing on Conversational Agent and Company Perceptions,” Computers in
Human Behavior (85), pp. 183–189.
Aylien. 2019. “Aylien.” (https://aylien.com, accessed February 5, 2019).
Beale, R., and Creed, C. 2009. “Affective Interaction: How Emotional Agents Affect Users,” International
Journal of Human Computer Studies (67:9), pp. 755–776.
Beverungen, D., Breidbach, C. F., and Poeppelbuss, J. 2017. “Smart Service Systems: An Interdisciplinary
Perspective,” Information Systems Journal, pp. 1–4.
Bickmore, T. W., and Picard, R. W. 2005. “Establishing and Maintaining Long-Term Human-Computer
Relationships,” ACM Transactions on Computer-Human Interaction (12:2), pp. 293–327.
Bolton, R. N., and Drew, J. H. 2002. “A Multistage Model of Customers’ Assessments of Service Quality
and Value,” Journal of Consumer Research (17:4), p. 375.
Boudreau, M.-C., Gefen, D., and Straub, D. 2001. “Validation in Information Systems Research: A State-
of-the-Art Assessment,” Management Information Systems Quarterly (25:1), pp. 1–16.
Brave, S., Nass, C., and Hutchinson, K. 2005. “Computers That Care: Investigating the Effects of
Orientation of Emotion Exhibited by an Embodied Computer Agent,” International Journal of
Human Computer Studies (62:2), pp. 161–178.
Cafaro, A., Vilhjalmsson, H. H., and Bickmore, T. 2016. “First Impressions in Human-Agent Virtual
Encounters,” ACM Transactions on Computer-Human Interaction (24:4), pp. 1–40.
Caruana, A., Ewing, M. T., and Ramaseshan, B. 2000. “Assessment of the Three-Column Format
SERVQUAL: An Experimental Approach,” Journal of Business Research (49), pp. 57–65.

Emulating Empathetic Behavior of Conversational Service Agents
Fortieth International Conference on Information Systems, Munich 2019 15
Chattaraman, V., Kwon, W. S., and Gilbert, J. E. 2012. “Virtual Agents in Retail Web Sites: Benefits of
Simulated Social Interaction for Older Users,” Computers in Human Behavior (28), pp. 2055–2066.
Chiu, H. C. 2002. “A Study on the Cognitive and Affective Components of Service Quality,” Total Quality
Management (13:2), pp. 265–274.
Chiu, H. C., and Wu, H. C. 2002. “Exploring the Cognitive and Affective Roles of Service Quality Attidude
across Gender,” Service Industries Journal (22:3), pp. 63–76.
“Cleverbot.” 2018. (https://www.cleverbot.com, accessed June 6, 2018).
Cowell, A. J., and Stanney, K. M. 2005. “Manipulation of Non-Verbal Interaction Style and Demographic
Embodiment to Increase Anthropomorphic Computer Character Credibility,” International Journal
of Human Computer Studies (62:2), pp. 281–306.
Davis, M. H. 2015. “Empathy and Prosocial Behavior,” in The Oxford Handbook of Prosocial Behavior,
Oxford, England: Oxford University Press.
Dennis, A. R., and Valacich, J. S. 2001. “Conducting Experimental Research in Information Systems,”
Communications of the Association for Information Systems (7:5), pp. 1–41.
Diederich, S., Brendel, A. B., and Kolbe, L. M. 2019. “On Conversational Agents in Information Systems
Research: Analyzing the Past to Guide Future Work,” in Proceedings of the International
Conference on Wirtschaftsinformatik, pp. 1550–1564.
Diederich, S., Brendel, A. B., Lichtenberg, S., and Kolbe, L. M. 2019. “Design for Fast Request Fulfillment
or Natural Interaction? Insights from an Online Experiment with a Conversational Agent,” in
Proceedings of the European Conference on Information Systems (ECIS), Stockholm, Sweden, pp.
1–17.
Dolata, M., Kilic, M., and Schwabe, G. 2019. “When a Computer Speaks Institutional Talk: Exploring
Challenges and Potentials of Virtual Assistants in Face-to-Face Advisory Services,” in Proceedings of
the Hawaii International Conference on System Sciences (HICSS), Grand Wailea, Maui, pp. 105–
114.
Elson, J. S., Derrick, D. C., and Ligon, G. S. 2018. “Examining Trust and Reliance in Collaborations
between Humans and Automated Agents,” in Proceedings of the Hawaii International Conference
on System Sciences (HICSS), Waikoloa Village, Hawaii, USA, pp. 430–439.
Faul, F., Erdfelder, E., Lang, A.-G., and Buchner, A. 2007. “G*Power 3: A Flexible Statistical Power
Analysis Program for the Social, Behavioral, and Biomedical Sciences,” Behavior Research Methods
(39:2), pp. 175–191.
Feine, J., Gnewuch, U., Morana, S., and Maedche, A. 2019. “A Taxonomy of Social Cues for
Conversational Agents,” International Journal of Human-Computer Studies.
(https://doi.org/10.1016/j.ijhcs.2019.07.009).
Feine, J., Morana, S., and Gnewuch, U. 2019. “Measung Service Encounter Satisfaction with Customer
Service Chatbots Using Sentiment Analysis,” in Proceedings of the International Conference on
Wirtschaftsinformatik, pp. 0–11.
Følstad, A., and Brandtzæg, P. B. 2017. “Chatbots and the New World of HCI,” Interactions (24:4), pp.
38–42.
Frey, C. B., and Osborne, M. A. 2017. “The Future of Employment: How Susceptible Are Jobs to
Computerisation?,” Technological Forecasting and Social Change (114), pp. 254–280.
Gabbott, M., and Hogg, G. 2004. “The Role of Non-Verbal Communication in Service Encounters: A
Conceptual Framework,” Journal of Marketing Management (17:1–2), pp. 5–26.
Gefen, D., and Straub, D. 1997. “Gender Differences in the Perception and Use of E-Mail: An Extension to
the Technology Acceptance Model,” Management Information Systems Quarterly (21:4), pp. 389–
400.
Gefen, D., and Straub, D. 2004. “Consumer Trust in B2C E-Commerce and the Importance of Social
Presence: Experiments in e-Products and e-Services,” Omega (32:6), pp. 407–424.
Gefen, D., and Straub, D. 2005. “A Practical Guide to Factorial Validity Using PLS-Graph: Tutorial and
Annotated Example,” Communications of the Association for Information Systems, pp. 91–109.
Gnewuch, U., Morana, S., Adam, M. T. P., and Maedche, A. 2018. “Faster Is Not Always Better:
Understanding the Effect of Dynamic Response Delays in Human-Chatbot Interaction,” in
Proceedings of the European Conference on Information Systems (ECIS), Portsmouth, United
Kingdom, pp. 1–17.
Gnewuch, U., Morana, S., and Maedche, A. 2017. “Towards Designing Cooperative and Social
Conversational Agents for Customer Service,” in Proceedings of the International Conference on
Information Systems (ICIS), Seoul, Korea, pp. 1–13.

Emulating Empathetic Behavior of Conversational Service Agents
Fortieth International Conference on Information Systems, Munich 2019 16
Google. 2019. “Google Dialogflow.” (https://dialogflow.com, accessed March 25, 2019).
Hassanein, K., and Head, M. 2007. “Manipulating Perceived Social Presence through the Web Interface
and Its Impact on Attitude towards Online Shopping,” International Journal of Human Computer
Studies (65:8), pp. 689–708.
Holtgraves, T., and Han, T. L. 2007. “A Procedure for Studying Online Conversational Processing Using a
Chat Bot,” Behavior Research Methods (39:1), pp. 156–163.
Johnston, R. 1995. “The Determinants of Service Quality: Satisfiers and Dissatisfiers,” International
Journal of Service Industry Management (6:5), pp. 43–51.
De Keyser, A., Köcher, S., Alkire, L., Verbeeck, C., and Kandampully, J. 2019. “Frontline Service
Technology Infusion: Conceptual Archetypes and Future Research Directions,” Journal of Service
Management (30:1), pp. 156–183.
Larivière, B., Bowen, D., Andreassen, T. W., Kunz, W., Sirianni, N. J., Voss, C., Wünderlich, N. V., and De
Keyser, A. 2017. “‘Service Encounter 2.0’: An Investigation into the Roles of Technology, Employees
and Customers,” Journal of Business Research (79), pp. 238–246.
Leite, I., Pereira, A., Mascarenhas, S., Martinho, C., Prada, R., and Paiva, A. 2013. “The Influence of
Empathy in Human-Robot Relations,” International Journal of Human Computer Studies (71:3),
pp. 250–260.
Lemon, K. N., and Verhoef, P. C. 2016. “Understanding Customer Experience Throughout the Customer
Journey,” Journal of Marketing (80:6), pp. 69–96.
Liu, B. 2010. “Sentiment Analysis and Subjectivity in: Handbook of Natural Language Processing, Second
Edition,” in Handbook of Natural Language Processing, Second Edition, Taylor & Francis, p. 568.
Liu, B. 2012. Sentiment Analysis and Opinion Mining, Morgan & Claypool Publishers.
Liu, K. K., and Picard, R. W. 2004. “Embedded Empathy in Continuous, Interactive Health Assessment,”
in Proceedings of the Workshop on Computer-Human Interaction Challenges in Health
Assessment, pp. 1–4.
MacDorman, K. F., Green, R. D., Ho, C. C., and Koch, C. T. 2009. “Too Real for Comfort? Uncanny
Responses to Computer Generated Faces,” Computers in Human Behavior (25:3), pp. 695–710.
Maglio, P. P., Vargo, S. L., Caswell, N., and Spohrer, J. 2009. “The Service System Is the Basic Abstraction
of Service Science,” Information Systems and E-Business Management (7), pp. 395–406.
(https://doi.org/10.1007/s10257-008-0105-1).
Marinova, D., de Ruyter, K., Huang, M. H., Meuter, M. L., and Challagalla, G. 2017. “Getting Smart:
Learning From Technology-Empowered Frontline Interactions,” Journal of Service Research (20:1),
pp. 29–42.
McQuiggan, S. W., and Lester, J. C. 2007. “Modeling and Evaluating Empathy in Embodied Companion
Agents,” International Journal of Human Computer Studies (65:4), pp. 348–360.
McTear, M. 2017. “The Rise of the Conversational Interface: A New Kid on the Block?,” Lecture Notes in
Computer Science.
Mohr, L. a., and Bitner, M. J. 1991. “Mutual Understanding Between Customers and Employees In Service
Encounters.,” Advances in Consumer Research (18), pp. 611–617.
Mori, M. 1970. “The Uncanny Valley,” Energy.
Nass, C., and Moon, Y. 2000. “Machines and Mindlessness: Social Responses to Computers,” Journal of
Social Issues (56:1), pp. 81–103.
Nass, C., Steuer, J., and Tauber, E. R. 1994. “Computers Are Social Actors,” in Proceedings of the ACM
CHI Conference on Human Factors in Computing Systems, Boston, USA, p. 204.
NextIT. 2018. “Helping a Railroad Service Conduct Business.” (http://nextit.com/case-studies/amtrak,
accessed November 1, 2018).
Niewiadomski, R., and Pelachaud, C. 2010. “Affect Expression in ECAs: Application to Politeness
Displays,” International Journal of Human Computer Studies (68:11), pp. 851–871.
Nunamaker, J. F., Derrick, D. C., Elkins, A. C., Burgoon, J. K., and Patton, M. W. 2011. “Embodied
Conversational Agent-Based Kiosk for Automated Interviewing,” Journal of Management
Information Systems (28:1), pp. 17–48.
Oracle. 2016. Can Virtual Experiences Replace Reality? The Future Role for Humans in Delivering
Customer Experience, p. 19.
Parasuraman, A., Zeithaml, V. A., and Berry, L. L. 1985. “A Conceptual Model of Service Quality and Its
Implications for Future Research,” Journal of Marketing (49:4), pp. 41–50.
Parasuraman, A., Zeithaml, V. A., and Berry, L. L. 1988. “SERVQUAL: A Multiple-Item Scale for
Measuring Consumer Perceptions of Service Quality,” Journal of Retailing (64:1), pp. 12–40.

Emulating Empathetic Behavior of Conversational Service Agents
Fortieth International Conference on Information Systems, Munich 2019 17
Pereira, R., and Baranauskas, M. C. C. 2015. “A Value-Oriented and Culturally Informed Approach to the
Design of Interactive Systems,” International Journal of Human Computer Studies (80), Elsevier,
pp. 66–82.
Price, L., Arnould, E., and Deibler, S. 1995. “Consumers’ Emotional Responses to Service Encounters: The
Influence of the Service Provider,” International Journal of Service Industry Management (6:3),
pp. 34–63.
Von Der Pütten, A. M., Krämer, N. C., Gratch, J., and Kang, S. H. 2010. “‘It Doesn’t Matter What You Are!’
Explaining Social Effects of Agents and Avatars,” Computers in Human Behavior (26:6), pp. 1641–
1650.
Qiu, L., and Benbasat, I. 2009. “Evaluating Anthropomorphic Product Recommendation Agents: A Social
Relationship Perspective to Designing Information Systems,” Journal of Management Information
Systems (25:4), pp. 145–182.
Reeves, B., and Nass, C. 1996. The Media Equation: How People Treat Computers, Television and New
Media Like Real People and Places, The Center for the Study of Language and Information
Publications.
Sah, Y. J., and Peng, W. 2015. “Effects of Visual and Linguistic Anthropomorphic Cues on Social
Perception, Self-Awareness, and Information Disclosure in a Health Website,” Computers in Human
Behavior (45), pp. 392–401.
Scherer, A., Wünderlich, N. V., and von Wangenheim, F. 2015. “The Value of Self-Service: Long-Term
Effects of Technology-Based Self-Service Usage on Customer Retention,” Management Information
Systems Quarterly (39:1), pp. 177–200.
Schroeder, J., and Schroeder, M. 2018. “Trusting in Machines: How Mode of Interaction Affects
Willingness to Share Personal Information with Machines,” in Proceedings of the Hawaii
International Conference on System Sciences (HICSS) (Vol. 9), Waikoloa Village, Hawaii, USA, pp.
472–480.
Seeger, A.-M., Pfeiffer, J., and Heinzl, A. 2018. “Designing Anthropomorphic Conversational Agents:
Development and Empirical Evaluation of a Design Framework,” in Proceedings of the
International Conference on Information Systems (ICIS), San Francisco, USA, pp. 1–17.
Short, J., Williams, E., and Christie, B. 1976. The Social Psychology of Telecommunications, London,
United Kingdom.
Urbach, N., and Ahlemann, F. 2010. “Structural Equation Modeling in Information Systems Research
Using Partial Least Squares,” Journal of Information Technology Theory and Application (JITTA)
(11:2), pp. 5–40.
Verhagen, T., van Nes, J., Feldberg, F., and van Dolen, W. 2014. “Virtual Customer Service Agents: Using
Social Presence and Personalization to Shape Online Service Encounters,” Journal of Computer-
Mediated Communication (19:3), pp. 529–545.
Vogel-Meijer, K. 2018. “KLM: Making Airline Customer Service Soar with Messenger.”
(https://www.facebook.com/business/success/klm-messenger, accessed November 1, 2018).
Wang, N., Johnson, W. L., Mayer, R. E., Rizzo, P., Shaw, E., and Collins, H. 2008. “The Politeness Effect:
Pedagogical Agents and Learning Outcomes,” International Journal of Human Computer Studies
(66:2), pp. 98–112.
Weizenbaum, J. 1966. “ELIZA—a Computer Program for the Study of Natural Language Communication
between Man and Machine,” Communications of the ACM (9:1), pp. 36–45.
Wels-Lips, I., van der Ven, M., and Pieters, R. 1998. “Critical Services Dimensions: An Empirical
Investigation across Six Industries,” International Journal of Service Industry Management (9:3),
pp. 286–309.
Wünderlich, N. V., and Paluch, S. 2017. “A Nice and Friendly Chat With a Bot: User Perceptions of AI-
Based Service Agents,” in Proceedings of the International Conference on Information Systems
(ICIS), Seoul, Korea, pp. 1–11.
Yan, A., Solomon, S., Mirchandani, D., Lacity, M., and Porra, J. 2013. “The Role Of Service Agent, Service
Quality , And User Satisfaction In Self -Service,” in Proceedings of the International Conference on
Information Systems (ICIS), Milan, Italy, pp. 1–21.
Zajonc, R. B., and Markus, H. 2002. “Affective and Cognitive Factors in Preferences,” Journal of
Consumer Research (9:2), p. 123.
Zanna, M. P., and Rempel, J. K. 1988. “Attitudes: A New Look at an Old Concept,” in The Social
Psychology of Knowledge, New York, USA: Cambridge University Press, pp. 315–334.