You are on page 1of 38

Group Decision and Negotiation (2019) 28:377–413

https://doi.org/10.1007/s10726-018-09611-4

Are You Angry (Happy, Sad) or Aren’t You? Emotion


Detection Difficulty in Email Negotiation

Christoph Laubert1 · Jennifer Parlamis2

Published online: 9 January 2019


© Springer Nature B.V. 2019

Abstract
This research investigates consistency of emotion detection in email negotiations.
Conveying and detecting emotions in negotiation is important because emotions can
function strategically. Therefore, this research systematically explores in four sepa-
rate studies how consistently individuals detect discrete emotions in text-based (email)
negotiations. Study 1 compared the ratings from two coders using a high quantity of
thought units (n  1317) and a negative bargaining zone negotiation scenario. In stud-
ies 2 and 3, three different negotiation scenarios were explored, first on a thought unit
level and then on a message unit level using a hierarchical emotion coding scheme.
In all three studies, coders’ perceptions were also compared with the text analysis
program LIWC. Study 4 compared coding from seven of the actual negotiators with
that of an independent coder and a computerized text program. All four studies found
low emotion recognition consistency across 14 different coders with only one nego-
tiation scenario in study 3 showing a moderate level of consistency. Comparisons of
computerized coding with human coders did not show improved agreement. High
amounts of contrary coding by independent coders were also found. Our research
makes an important contribution to the literature by challenging the common assump-
tion that emotions can be reliably detected in email negotiation. Factors that might
influence more consistent emotion recognition and conveyance as well as implications
for practice and future research are discussed.

Keywords Email negotiation · Emotion detection · Discrete emotion · Computerized


text analysis · LIWC

B Christoph Laubert
christoph.laubert@fu-berlin.de
Jennifer Parlamis
jparlamis@usfca.edu

1 School of Business and Economics, Marketing Department, Freie Universität Berlin, Arnimallee
11, 14195 Berlin, Germany
2 School of Management, University of San Francisco, 2130 Fulton Street, San Francisco, CA 94117,
USA

123
378 C. Laubert

1 Introduction

Employees face a constant stream of incoming emails during a working day. In 2015,
an individual employee in the U.S sent and received, on average, 122 business emails
per day and this number is projected to grow (Radicati Group 2016). Global business
email use is in the billions of messages sent and received per day (Radicati Group
2016). While some emails contain irrelevant information and are deleted quickly,
other emails convey important information, entail instructions from a supervisor, or
comprise whole business conversations over days, weeks, or months with colleagues,
suppliers, or customers.
Importantly, many of these conversations involve negotiations, a “process by which
two or more parties attempt to resolve their opposing interests” (Lewicki et al. 2011,
p. 6). Whether the negotiations are intra-organizational (e.g., two colleagues negotiate
about project responsibilities) or inter-organizational (e.g., buyer and seller negoti-
ate about delivery conditions), it is clear that many business negotiations take place
electronically (Griessmair et al. 2015). Given this pervasive use of email in business
communication and knowing that individuals commonly use email for negotiations,
being able to negotiate successfully via email is a critical component of an individual’s
performance and a company’s success.
Despite the ample opportunities for practice, email negotiations remain a challenge.
Some of the main hurdles include diminished cooperation, trust and information shar-
ing along with increased negative attributions and contentiousness (Ebner 2011). This
increase in contentiousness is characterized by a greater use of threats (Morris et al.
2002), greater risk of escalation (Friedman and Currall 2003) and a tendency to become
angry, insulting and hostile (Thompson and Nadler 2002). Negotiations are already
demanding, but when negotiations occur over email the emotionally loaded content
adds layers of complexity.
There are over two decades of literature showing that emotions are an important
determinant of outcomes in negotiations (e.g., Allred et al. 1997; Barry and Oliver
1996; Kopelman et al. 2006; Morris and Keltner 2000a, b). According to the view of
social function of emotions (see Keltner and Haidt 1999; Van Kleef 2009), emotions
provide information that “help individuals know others’ emotions, beliefs, and inten-
tions” (Keltner and Haidt 1999, p. 511) in order to coordinate behavior, make sense
of interactions and to solve social problems like establishing hierarchy, influencing
others, and creating norms. In line with the social function of emotions framework,
negotiation research has demonstrated that emotions can be used strategically to influ-
ence counterparts. For example, showing disappointment can result in more generous
offers from the counterpart, if it evokes guilt from the other party (Lelieveld et al.
2013). Communicating sadness can lead to more concessions, if the sender of the
emotion is perceived as lower in power (Sinaceur et al. 2015). However, the strate-
gic use of a discrete emotion requires that the receiver of the communication is able
to perceive and understand the sender’s intended emotion. In a meta-analysis exam-
ining the accuracy of emotional expressions across cultures, Elfenbein and Ambady
(2002) observed that people recognize emotions 58% of the time. While on the surface
this might seem high, their average result still means that four out of ten judgments
about emotions are not correct. More plainly put, conveying an emotion has a 40%

123
Are You Angry (Happy, Sad) or Aren’t You? Emotion… 379

chance of being interpreted incorrectly. This can be a problem in a challenging task


such as a negotiation in which meticulous formulation is essential to convey the right
preferences and priorities. Second, in their meta-analysis, the authors only include
studies in which emotions were displayed via facial photograph, photographs of the
body, video, or voice—all conveying non-textual communication cues. Therefore, the
chance of misinterpretation may further increase if the amount of communication cues
decreases, such as when we consider email instead of photographs or voice messages.
Thus, this popular meta-analysis leaves a blank spot in the literature about text-based
emotion detection. Similarly, Hawk et al. (2009) exclude text-based communication
in an extensive study about emotion recognition accuracy.
We find some indications in the literature that the accuracy of emotion recognition
is more difficult without non-verbal cues (Byron 2008; Kruger et al. 2005; Walther and
D’Addario 2001). In addition to the lack of nonverbal cues in email communication,
there is also less social presence and less synchronicity (Hatta and Ken-ichi 2008).
While this might suggest that emotion recognition in email negotiation is worse than
in face-to-face negotiation, there is a paucity of literature directly addressing this.
A surprising aspect in negotiation research is that the ability to recognize emotions
with high accuracy or the ability to detect emotions consistently has been seldom
questioned for email communication. While research in the analysis of emotions in
text-based communication (e.g., blogs, or newspaper articles) has gained traction in
different fields, such as computer science (Calvo and Mac Kim 2013; Ghazi et al.
2010), psychology (Mehl and Pennebaker 2003), information science (Li and Xu
2014), linguistics (Strapparava and Mihalcea 2008), or communication research
(Walther and Addario 2001), negotiation research has yet to investigate the decoding
of discrete emotions in email negotiations. Existing studies focused on the abstract
level of positive and negative emotions (Cheshin et al. 2011; Koeszegi et al. 2011;
Parlamis and Geiger 2015) or “friendly complaisant” and “unfriendly negatory” emo-
tions (Griessmair and Koeszegi 2009). Furthermore, many studies focus only on two
discrete emotions: happiness and anger (Olekalns and Druckman 2014). Thus, there is
a clear lack of testing around the decoding of a variety of discrete emotions in an email
negotiation context. It is critical for negotiators to know if their emotions (which are
usually more complex than just positive or negative) are being conveyed as intended
in an email. If emotions are not decoded reliably or as intended, negotiators will not be
able to use the emotions strategically; further, what is detected might be misinterpreted
over email and lead to misunderstanding and potentially worse outcomes.
Some argue that with less social presence in electronic communication, there are
fewer emotions conveyed with cues being filtered out (Walther and Parks 2002). On
the other hand, some studies show that emotions are present in electronic communica-
tion (e.g., Griessmair and Koeszegi 2009), and different cues (e.g., a time stamp of an
email (Rice 1990)) substitute for non-verbal information. However, based on concep-
tual insights from outside the negotiation realm, while emotions might be conveyed
generally or even with different cues, it is less clear if the receiver correctly perceives
emotions or even picks up on the alternative cues. Different cues and the increased
control a sender has when communicating an emotion via text (e.g., a sender can better
control his emotion while writing an email and thus tone down the intensity of the
emotion), all contribute to the lack of clarity around emotion detection in text-based

123
380 C. Laubert

communication (Byron 2008; see also Pesendorfer and Koeszegi 2006). Additionally,
abstract general emotions might be more easily decoded than discrete emotions and
electronic communication may have negative cues that override other emotional cues
(Walther and D’Addario 2001). Therefore, given the lack of clarity and status quo atti-
tude about emotion detection, we propose an exploratory research design with the aim
to systematically observe the phenomenon, identify key issues, and build a solid ground
for future empirical investigations. At this stage we do not test specific hypotheses;
rather, this exploratory research will investigate the following research question:
How consistent and reliable are individuals at detecting discrete emotions in
naturalistic, text-based (email) negotiation?
The structure of the paper is as follows. We first give an overview of the theoreti-
cal background regarding interpersonal emotion research in the negotiation context,
email negotiations, and computerized text analysis research about emotions. We then
describe four studies. In study 1, a high quantity of units (n  1317) were coded in
one negotiation scenario. Then we looked at three different negotiation scenarios, first
on a thought unit level (study 2) and then on a message unit level (study 3). In all three
studies, we compared coding from different human coders and a computerized text
analysis program. Last, we compared emotions which were indicated by the actual
negotiators with coding from a human and a computerized text analysis program (study
4). We then discuss the results, suggest implications, and conclude with limitations
and a research outlook.

2 Theoretical Background

2.1 Emotions in Email Negotiation

Despite the dominance of a rational approach in the early years of negotiation research,
it is now a given that emotions are inextricably linked with negotiation and significantly
impact the negotiation process and outcome (see Allred et al. 1997; Barry and Oliver
1996; Bazerman et al. 2000; Kopelman et al. 2006; Van Kleef 2009). More recently,
the social function approach has added valuable insights about the role of emotions
in negotiation for theory and practice (Morris and Keltner 2000a, b; Van Kleef et al.
2010). The approach postulates that emotions not only evoke cognitive and behav-
ioral changes within the emotion holder, but also affect the receiver of the emotions.
From this perspective, emotions are a consequence of interactions and a mechanism
by which relational problems in social exchange can be solved (Keltner and Kring
1998; Morris and Keltner 2000a, b). A variety of studies have tested how different
emotions influence a counterpart in a negotiation (for an overview see for example:
Druckman and Olekalns 2008; Olekalns and Druckman 2014). Many of these studies
have been conducted in a face-to-face (FtF) scenario (e.g., Overbeck et al. 2010) or
with emotionally primed text communication through (online) surveys (e.g., Harinck
and Van Kleef 2012).
Email communication has received remarkable attention in the scientific debate
in the last decade, too. This may be due to the high practical relevance because

123
Are You Angry (Happy, Sad) or Aren’t You? Emotion… 381

of the ubiquitous use of email communication in business and private life. As part
of computer-mediated communication (CMC), research about email communication
initiated more than 25 years ago (e.g., Trevino et al. 1987; Walther 1992). Since the
beginning, there have been discussions regarding if and how CMC is different than
FtF communication. Some argue that with less social presence in CMC, there are
fewer emotions conveyed, which is also described as cues filtered out (Walther and
Parks 2002). On the other hand, studies show that emotions are present in CMC (e.g.,
Griessmair and Koeszegi 2009), and different cues (e.g., a time stamp of an email or all
capital letters) substitute for non-verbal (e.g., facial expressions and posture) and para-
verbal (e.g., tone or pacing) information. This perspective is described as cues filtered
in (Walther and Parks 2002). Some even argue that CMC and FtF communication are
similar in the amount of conveyed emotions (Derks et al. 2008).
Research about email communication in negotiation also increased (see for an
overview: Griessmair et al. 2015). For example, positive and negative emotions
have been investigated as part of a negotiation phase analysis in a CMC context
(Koeszegi et al. 2011). In another extensive study, using over 800 electronic negotia-
tion transcripts, findings indicated that communicating more positive emotion resulted
in more negotiation success (Hine et al. 2009). Griessmair and Koeszegi (2009)
analyzed negotiation transcripts from electronic negotiations focusing on emotions
communicated in seemingly factual statements. They applied a multi-dimensional
scaling procedure on a macro level (entire negotiation), meso level (single mes-
sage) and micro level (single utterance) and created emotional maps that provided
information on an abstracted level (e.g., unfriendly negatory/friendly complaisant;
high/low assertiveness); they did not report information about discrete emotions.
Similarly, Filzmoser et al. (2016) applied multi-dimensional scaling to code for
emotions in a negotiation experiment, which was conducted with a CMC system.
In their study, emotions were analyzed to their degree of valence and activation.
They show a relationship between positive emotions and value creation as well
as negative emotion and value claiming in regards to strategic actions in negotia-
tion.
In sum, while there is some evidence that general emotions (i.e., positive or
negative) can be discerned and, in turn, influence negotiation outcomes, there is
little research that investigates how specific discrete emotions are perceived in a
CMC or email negotiation. It is especially relevant to understand how discrete emo-
tions can be conveyed in text-based communication, if we want to transfer the
findings from the social functional perspective of emotion from FtF to the email
context.

2.2 Emotion Recognition Difficulties in Email Communication

Against the backdrop of some empirical support that the email communication
medium conveys general emotional content and this content then impacts out-
comes (Griessmair and Koeszegi 2009; Walther and D’Addario 2001), it is not
clear if multiple receivers of a message perceive the same general emotion con-
tent as it is intended nor can it be confirmed that individuals can decode discrete

123
382 C. Laubert

emotions. Hence, we propose taking an individual difference approach to this phe-


nomenon and first replicate the finding that general positive and negative emotions
can be recognized, and then investigate if discrete emotions such as disappointment,
contempt, pride, or optimism, for example, are perceived similarly in an email nego-
tiation.
If it is possible to detect discrete emotions in email negotiation is not clear
yet. Derks et al. (2008) recognize that “[t]here is no further research, however,
on the potential misinterpretation of the expresser’s emotions in CMC, and thus
it is unclear whether this is a potential danger” (p. 777). They argue that the
reduced visibility in CMC has a negative effect on the ability to decode the sender’s
emotion. Furthermore, the lack of non-verbal signals can lead to either over- or
underestimated reactions towards the perceived emotion because of misinterpreta-
tion (Derks et al. 2008). In addition, it should be noted that the email negotiation
environment is ripe for exploiting the ambiguity in the emotions that are commu-
nicated over email. Intentional obfuscation or strategic manipulation of the email
medium to further the negotiator’s goal can also complicate or hinder emotion detec-
tion.
Another potential complication in emotion detection in email is egocentrism, which
has been shown to negatively impact how well we think we can communicate emotions
via email (Kruger et al. 2005). In their research, Kruger et al. (2005) hypothesized
that people routinely overestimate their ability to communicate over email. In a series
of studies, they show that it is not only difficult to communicate humor or sarcasm
over email, but also that the communicators are not aware of this shortcoming. Thus,
sender and receiver think they can convey or perceive emotions in emails, but actually
do that worse than they estimate. Riordan and Trichtinger’s (2017) findings showed
a similar pattern. They expected friendship and context to impact accuracy; how-
ever, findings indicated that friends did not show greater accuracy than strangers
even though participants were more confident that friends would be more accu-
rate.
In her conceptual paper, Byron (2008) reflected about variables that influ-
ence emotion misperception in emails. She conceptualizes a model of “Sender,
Receiver, Social Context, and Message Factor Effects on Receivers’ Emotion
Misperceptions in Emails”. In her model, she specifically proposed a relation-
ship among (a) sender factors (gender, relationship length, and relative status)
and (b) receiver factors (age and negative affectivity), and the dependent vari-
ables “neutrality effect” and “negativity effect”. “Neutrality effect” means that
actual positive emotions are perceived as more emotionally neutral and a “nega-
tive effect” describes that positive or neutral emotions are perceived more negative
than intended by the sender of the communication. Besides sender and receiver spe-
cific factors, Byron (2008) also argues that it is more difficult to correctly perceive
emotions in email, because a sender can better control his emotion while writ-
ing an email and thus temper the intensity of the emotion (see also Pesendorfer
and Koeszegi 2006). An empirical test of her proposition has not been conducted
yet.
Last, another source of difficulty in decoding discrete emotions can be that in
CMC negative cues have been found to override any other emotional cues (Walther

123
Are You Angry (Happy, Sad) or Aren’t You? Emotion… 383

and Addari 2001). In their study, Walther and D’Addario (2001) showed that when a
message contained a negative element the message was always rated more negative.
But if a positive component was added to a negative message, it did not change the
verbal connotation of the message significantly. In sum, based on the existing literature,
we have to challenge the assumption that negotiators can decode discrete emotions
conveyed via email.

2.2.1 Level of Emotion Analysis

As mentioned, there are studies that identify emotions in CMC/email negotiation.


An important limitation, however, is that these studies used an abstracted level of
emotion analysis. Emotions identified have been on the level of valence (positive or
negative direction) or rated in intensity (high/low).1 Undoubtedly, all these findings
are important for negotiation theory and practice, but they only touch the surface
of the role and importance of diverse, discrete emotions in negotiation. The cumu-
lated knowledge about how specific emotions work as a social function (Kopelman
et al. 2006) gives negotiators an instrument to shape the process more efficiently and
effectively to reach a mutual agreement. Thus, to be able to use emotion strategi-
cally in email negotiation, we need to find out if emotions on the discrete level can
be conveyed so that emotion detection between sender and receiver is consistent and
accurate.

2.2.2 Primed Versus Natural Communication

A further notable aspect of the existing research of emotion decoding in text-based


communication that might be problematic is due to the experimental design. Study
participants were either (a) told to write an angry/sad/happy message; (b) asked to
rate messages that were pre-tested for high emotional intensity; or (c) manipulated
to feel a specific emotion for an experiment. These procedures make sense in terms
of testing effects between variables in a laboratory setting; however, it is not sur-
prising that negotiators recognize an angry sentence, if the writer of the message
was explicitly instructed to write angry or if the message was pretested in regards
to being an intense angry emotion. But this kind of emotional primed communica-
tion lacks external validity. The emotional intensity may be totally different in natural
communication we face every day. Similar, Hine et al. (2009) argue that the magni-
tude of emotional intensity is different in simulated CMC negotiations and calls for
research in more naturalistic settings, as other researchers do as well (Hawk et al.
2009).
The use of primed communication is also criticized by Hancock et al. (2008). They
argue that in these cases negotiators are often asked to express a specific emotion
they do not necessarily feel. Hence, looking at discrete emotion decoding with natural
communication adds to the research. In the same vein, Druckman and Olekalns (2008)
argue for a better balance between internal and external validity and see “a loss in

1 Interestingly, research about decoding of discrete emotions with face or voice stimuli has been conducted
(e.g., see Hawk et al. 2009).

123
384 C. Laubert

relevance to real-world negotiations” (p. 5) by using computer controlled messages in


emotion research.

2.2.3 Importance of Emotion Detection Consistency

Investigating discrete emotions in naturalistic text-based negotiation is important,


because consistent emotion detection may have a positive influence on different vari-
ables. To start with, emotionally intelligent negotiators process more emotional cues
and hence can gain more information about the opponents’ issues (Barry et al. 2004,
p. 102). Similarly, Elfenbein et al. (2007) find that emotion recognition accuracy of
facial expressions predicts objective performance. In their study, sellers that performed
better in an emotion recognition test created more value and negotiated a bigger claim.
Their rationale is that negotiators who score higher in emotion recognition accuracy
can better use implicitly revealed information and subtle communication signals to
their benefit in a negotiation. Further, when female managers have a better perception
of non-verbal emotions, the subordinates report to be more satisfied and their respec-
tive supervisors rate them higher in performance (Byron 2007). Hawk et al. (2009)
also conclude that high decoding accuracy of emotions helps to establish important
functions of emotions such as increasing social distance or reduction of interactions
through negative emotions.
On the other hand, if emotion perception consistency and accuracy is low, emotions
may be perceived which are not actually there, or emotions are misinterpreted. This
can have a negative influence on workplace relationships and also on organizational
effectiveness, since communication that is emotionally misinterpreted will be more
difficult to make sense of (Byron and Baldridge 2005).
In the following, we will discuss four studies that address the basic question of
whether individuals are able to consistently detect discrete emotions in naturalistic,
text-based (email) negotiation. Table 1 gives an overview of the characteristics of the
following studies.
Table 1 Overview of the studies

Study 1 Study 2 Study 3 Study 4

Number of coded 1317 718 116 107


units
Level of analysis Thought unit Thought unit Whole message Whole message
Classification Flat Hierarchical Hierarchical Hierarchical
schemes
Negotiation 1 Scenario: 3 Scenarios: 3 Scenarios: 1 Scenario:
scenario Negative Negative, single Negative, single Single point
bargaining zone point, and point, and bargaining zone
positive positive
bargaining zone bargaining zone
Coders Two coders Two coders Two coders One coder/seven
negotiators as
coders
LIWC analysis Yes Yes Yes Yes

123
Are You Angry (Happy, Sad) or Aren’t You? Emotion… 385

3 Study 1

3.1 Method

In study 1, 46 graduate students from an American East Coast university participated


in the negotiation. They negotiated in 23 dyads as part of a class on “Negotiation and
Bargaining for Managers”. Negotiation role information and email addresses for their
negotiation counterparts were distributed in class. The students negotiated the Texoil
case study (Goldberg 1998). This negotiation is between a service station owner and a
Texoil representative. They are negotiating the sale of a service station near the port of
Los Angeles. This negotiation has a negative bargaining zone and requires creativity
for optimal solutions. The students were instructed that they were to use only email
to conduct the negotiation and had one week to come to a mutual agreement. If they
could not come to an agreement by the designated time, they were to consider the
result an impasse. Upon completion of the negotiation, students were asked to provide
the entire email exchange to the researchers.
For the coding of the emotions, we used a thought unit as the level of analysis. A
thought unit has been identified as a useful unit for content analysis in general (Srnka
and Koeszegi 2007), and has been applied to the negotiation context, too (Pesendorfer
et al. 2007). Pesendorfer et al. (2007) defined a thought unit as “conveying one thought
communicated by a negotiator” (p. 1322).
The authors of this article conducted thought unitizing for five dyads (i.e., more
than 20% of all data). Then the quality measure was compared (Guetzkow’s U  0.03).
Guetzkow’s U reliability scores of 0.10 or below are indicated as acceptable (Folger
et al. 1984). Because of the highly satisfying result, only one author coded the rest of
all dyads into thought units. This unitizing process resulted in 1317 units from 366
messages covering all of the 23 dyads.
After the thought unitizing, two trained research assistants, who were blind to the
research question, independently coded all 1317 thought units in terms of discrete
emotions. These two coders were used to explore the emotion detection consistency,
which is necessary because of a paucity of empirical studies about it. Hence, the study
setting provides first empirical insight rather than statistical correlations. The data was
prepared in an Excel sheet that combined all thought units. The Excel sheet structure
shows the coder when a message and/or a whole negotiation begins and ends. The
coders only had to tick a box to indicate which emotion is present. In this process,
they had to decide for each thought unit individually if it can be classified as one of
the eight suggested discrete emotions from the model of Morris and Keltner (2000a,
b): liking, interest, anger, contempt, embarrassment, empathy, pain or exasperation.
Hence, this is called a flat classification scheme (Ghazi et al. 2010). We further added
the categories “no emotion” and “not classifiable but an emotion is present”, in case
the coders could not decide on one of the eight emotions. The coders received detailed
instructions about the emotions and the coding process. The instruction sheet was
shown to three colleagues to check for comprehension and refined based on their
feedback (see Appendix 1). To assure high quality coding decisions, we asked the
coders to constantly check with the emotion descriptions, take regularly breaks, and

123
386 C. Laubert

to point out the word or a group of words for each thought unit, which is most important
for their decision for a discrete emotion.
Furthermore, we also investigated the data using automated, software-based cod-
ing. Though empirical research in this area is still new, some promising work on text
analysis of emotion has been conducted (for an overview see: Calvo and Mac Kim
2013). A validated tool for computerized text analysis for emotion detection is Lin-
guistic Inquiry and Word Count (LIWC; Pennebaker et al. 2001; Kahn et al. 2007;
Tausczik and Pennebaker 2010; Tov et al. 2013). Based on dictionary categories, it
allows for text analysis of positive and negative emotions as well as a few discrete emo-
tions. It has been successfully applied in research in real world contexts. For example,
Robinson et al. (2013) showed that LIWC scores in specific categories (e.g., punc-
tuation, word simplicity) can predict university course performance. Further, LIWC
has been used to investigate naturalistic language use of students in regards to their
social life stability (Mehl and Pennebaker 2003). Research by Hancock et al. (2008)
found that people who were induced with negative affect through sad music used
more negative emotions in a subsequent instant messaging conversation. Comput-
erized text analysis has also been used in negotiation research. For example, Brett
et al. (2007) used LIWC to analyze positive and negative emotions in online dispute
resolution.
However, researchers have mainly relied on computerized text programs to analyze
different text units and compare the results of these units. That means that only different
LIWC scores were compared, but not LIWC scores and other data points (e.g., human
judges). We want to address this gap by using both intercoder judgments and LIWC
coding. This will allow us to draw inferences both about the ability of LIWC to
detect emotions in email text as well as its consistency with human coders. In other
words, we will compare the results of “subjective” coders to the technical coding of
a computer program. Each thought unit is analyzed individually. We can only do this
comparison as one part of the analysis for positive and negative emotional valance
due to the lack of discrete emotion categories in LIWC. For the LIWC analysis, we
aggregated liking and interest as positive emotion. Anger, contempt, embarrassment,
pain and exasperation constitute the category negative emotion. No emotion was used
as neutral. We did not consider “not classifiable” in the LIWC analysis as well as
empathy. Empathy cannot clearly be classified as positive or negative valence. Next,
we discuss the results.

3.2 Results

To assess the coding, we provide below an intercoder consistency matrix as sug-


gested by Srnka and Koeszegi (2007). This table cross-tabulates the coding of coder 1
(columns) and coder 2 (rows). For example, the Table 2 reports that eight times coder
1 marked a thought unit as the emotion “liking” and coder 2 marked these units as
“interest”.
The overall agreement rate between the coders is 21.9% and Cohen’s kappa is
0.11. Cohen’s kappa was calculated including chance-agreement correction following
the procedure by Gwet (2014). The Cohen’s kappa score is judged as poor (Fleiss

123
Table 2 Intercoder consistency-matrix of study 1

Coder 2 Coder 1

Liking Interest Anger Contempt Embarrassment Empathy Pain Exasperation Not No emotion Total Agreement
classifiable (%)

Liking 17 65 3 3 17 39 – – 6 2 152 11.2


Interest 8 164 12 5 23 12 – 1 5 4 234 70.1
Anger 1 25 15 3 18 – – 2 1 – 65 23.1
Contempt 2 48 13 6 49 2 – 2 – – 122 4.9
Are You Angry (Happy, Sad) or Aren’t You? Emotion…

Embarrassment 1 4 4 – 4 1 – – – – 14 28.6
Empathy 6 58 7 2 34 16 – 1 – 1 125 12.8
Pain 1 1 – – 1 1 – – – – 4 0.0
Exasperation – 9 12 3 11 2 – 1 – 2 40 2.5
Not 11 29 10 – 14 16 – – 2 – 82 2.4
classifiable
No emotion 37 191 18 5 84 39 2 4 36 63 479 13.2
Total 84 549 94 27 255 128 2 11 50 72 1317
Agreement 20.2 27.6 16.0 22.2 1.6 12.5 0.0 9.1 4.0 87.5
(%)
A bold figure indicates the number of agreed units between both coders

123
387
388 C. Laubert

1981), low (Green 1997), discard (Krippendorff 1980), or slight strength of agree-
ment (Landis and Koch 1977).2 Besides this low overall agreement rate, there are
further interesting results: First, the low agreement rate is consistent across the ten
categories. The agreement rate ranges from 0.0% for “pain” to a maximum of 27.6%
for “interest” for coder 1 and 0.0% for “pain” to a maximum of 28.6% for “em-
barrassment” for coder 2. The only two exceptions are “no emotion” (87.5%) for
coder 1 and “interest” (70.2%) for coder 2. This is due to a high focus of coder 1
on “interest” and a high focus of coder 2 on “no emotion”. Thus, if the other coder
judged a unit as “interest” or “no emotion” respectively, the likelihood was high to
get an agreement. Further, it has to be noted that coder 1 judged many units as pos-
itive (interest, liking, empathy; n  761 units; 56.92%) and coder 2 judged many
units as neutral (no emotion; n  479 units; 35.83%). Second, the intercoder matrix
displays that 205 units are coded differently regarding emotional valence from both
coders, which equals 15.6% of all units. That means coder 1 judged a unit as postive
and coder 2 judged it as negative, or vica versa. For example, this is the case where
coder 1 judged a unit 25 times as “interest” and coder 2 judged the same unit as
“anger”. Third, in 63 units (4.8%) both coders agreed that there is no emotion present.
That means in more than 95% of all units, at least one coder perceived an emotion
as being present. It is important to stress that the coding instructions clearly indi-
cated that “no emotion” is always a reasonable choice. Last, we provided the coders
with the category “not classifiable” an option if more than one emotion is present in
a unit. However, in this category a specifically low agreement rate (2.4%/4.0%) is
present.
The LIWC analysis shows similar results: Cohen‘s kappa between LIWC and coder
1 is 0.07. The same Cohen‘s kappa (0.07) is reached when comparing LIWC scores
to coder 2.

3.3 Discussion

Overall, the results showed very low agreement rates across coders. This offers first
empirical support for the notion that individuals are not able to consistently detect
emotion in email negotiation transcripts. Interestingly, we can determine a neutrality
effect for coder 2 and a positivity effect for coder 1. The neutrality effect has been
identified in the literature before (Byron 2008). The neutrality effect describes that
people tend to perceive communication as neutral even though the sender put positive
emotions in the message. This effect can be due to the lack of cues and feedback,
difficulty conveying emotional intensity, and the fact that people tend to see an email
as a serious task (Byron 2008). Hence, our study offers some empirical support for the
presence of this effect. On the other hand, in contrast to the proposed negativity effect
(Byron 2008), we find a positivity effect. That means coder 1 rates more than 50% of
the thought units as containing liking, interest, or empathy. While Byron (2008) argues
2 We used Cohen’s kappa as done by other researchers in a negotiation context (Brett et al. 2007; Koeszegi
et al. 2011; Pesendorfer and Koeszegi 2006). We also calculated Krippendorff’s alpha and found similar
results. The evaluation of Cohen’s kappa is not fixed defined. For the rest of the paper, we will use Landis
& Koch (1977) benchmark scale, because it is widely applied (Gwet 2014) and is the most detailed one
with its six value ranges.

123
Are You Angry (Happy, Sad) or Aren’t You? Emotion… 389

that a negativity effect may result from an ambiguous tone, the salience of negative
cues (see also Walther and D’Addario 2001), or less information available about the
recipient to create an effective email, we need to think about reasons that people
perceive negative or neutral messages as more positively. A reason may be an individual
difference variable such as a positive attitude of coder 2; someone who thinks that
people have in general good intentions. Furthermore, the context of negotiation can
lead to a positivity effect by some individuals. In negotiation, people voluntarily get
together to reach a mutual agreement. Thus, coder 2 may assume a willingness and
more positive attitude between the two negotiators to work successfully together.
However, this does not explain why the positivity effect is only present for one coder
in this context. Importantly, this result should not be over interpreted at this stage. We
will look to see if this positivity result is present in the studies that follow.
Puzzling is the fact that more than 15% of the units were assigned to completely
opposite emotions by the coders. It is surprising that two individuals think about the
same thought unit so differently. For example, one coder judges a unit to represent
optimism, while the other coder perceives it as sadness. Examples of contradictory
units can be found in Appendix 2. These are not slight differences in perceiving emo-
tion nuances in a text-based document, but provide evidence of opposing perception
of emotion valence. This finding supports the rationale about the danger of misinter-
preting emotions in email (Kruger et al. 2005) and it shows the potential of triggering
unwanted behavior because emotions are not communicated clearly enough via email.
This is especially relevant when we consider that both coders together nearly always
(> 95%) perceived an emotion in these email conversations. It means when you send an
email to a group of people, surely some individuals will perceive emotions, even if the
sender did not intend to convey an emotion. Furthermore, different email receivers of
this group email will most likely experience different emotions. Thus, it is negligent to
communicate in email negotiations without paying special attention to how emotions
are incorporated in or omitted from the text.
Though this first study provides a fertile base for questioning the ability to convey
discrete emotions in email negotiation, there are factors that may influenced the detec-
tion ability in this study design. First, we used a flat classification scheme, because we
were originally interested in the emotions proposed in the model from Morris and Kelt-
ner (2000a, b). However, Ghazi et al. (2010) show that using hierarchical classification
in text analysis leads to significantly higher accuracy of emotion than flat classification.
Consequently, in the second study, we will use the hierarchical classification scheme
to address this potential limitation.
The fit between the human coders using a hierarchical classification scheme—where
they specially code for positive and negative emotions on their way to narrow down
the discrete emotion—should be better with the LIWC coding. The very low Cohen’s
kappa between LIWC and the human coders indicate that the applied coding scheme
does not work with the computerized text analysis.
Using a hierarchical classification scheme should also tackle a problem that more
than one discrete emotion is present in a unit. While this can be the case, with a
hierarchical classification scheme we force the coders to think about the most salient
emotion on each level in a unit. If there is a dominant emotion, the agreement rate

123
390 C. Laubert

should increase in the next study. Otherwise, there will be an interesting question as
to why different subjects find and/or focus on different salient emotions in a unit.
Another aspect to consider is the scenario used in this study. The Texoil case study
is difficult to negotiate because it has a negative bargaining zone. Furthermore, the
case study is about an emotional topic: a gas station owner wants to sell his business
to be able to pursue a lifetime dream and attend to the health of his partner. Thus, in
this case study many different emotions may be evoked and are especially difficult
to be differentiated and detected. A scenario with a positive bargaining zone and less
potential for emotional distress potentially leads to less noise in emotion coding. We
address this problem in the next study, too.

4 Study 2

4.1 Method

To address the concerns from our first study, and to specifically rule out emotional
complexity as an alternate explanation for emotion detection inconsistency, we con-
ducted a second coding study. This time we used email transcripts from three different
case studies: Texoil, Websurf (Notini 2008), and Lattitude (Manwaring 2003). The
case studies differed with respect to their bargaining zone: Texoil has a negative bar-
gaining zone, Websurf has one single point of agreement, and Lattitude has a positive
bargaining zone. Hence, we wanted to assess the impact of the bargaining zone on the
intercoder agreement rate. We used the same Texoil data as in study 1. We collected
data from students (25 dyads) from two West Coast Universities who negotiated the
Websurf case study via email as part of a class called “Negotiation and Bargaining
Strategy”. Websurf is a negotiation between an owner of a website called websurf.net
and a potential buyer who is a CEO of a company about to go public with a website
called websurf.com. To avoid customer confusion the CEO wants to buy the similar
website address. For the Lattitude case study, we collected 27 dyads consisting of stu-
dents from two different West Coast Universities. Negotiating the Lattitude case study
via email was part of their class called “Negotiation and Bargaining Strategy”. In the
Lattitude negotiation a teacher is selling a URL to an interested buyer. The teacher has
mistakenly chosen the URL and does not need it anymore. The buyer owns a French
boutique called L’Attitude and wants to expand his growing business to the online mar-
ket. The boutique owner wants to open the online shop with the address Lattitude.com.
Two research assistants (different from those assisting in study 1) who were unfa-
miliar with the subject of this research were trained to apply the coding scheme. We
randomly selected five email transcripts from each of the three different scenarios. This
is similar to Griessmair and Koeszegi’s (2009) procedure where they had randomly
selected 6 transcripts from a total of 75. Thus, the coders got data with five transcripts
comprising 202 thought units from Texoil case study, five transcripts comprising 348
thought units from the Websurf case study, and five transcripts comprising 168 thought
units from the Lattitude case study. Due to the high consistency in unitizing using two
coders in study 1, one author was responsible for the unitizing process in this study.
All in all, 718 thought units were coded.

123
Are You Angry (Happy, Sad) or Aren’t You? Emotion… 391

Fig. 1 Emotion structure based on Shaver et al. (1987, p. 1067)

For the coding procedure, we implemented a hierarchical coding scheme. We used


the basic emotion model from Shaver et al. (1987) because this model offers a three
level tree structure of emotions and is widely acknowledged and discussed in emotion
research (e.g. Ekman 1992; Smith and Lazarus 1990). The first level is positive or
negative emotion valence, followed by the five basic emotions (love, joy, anger, sad-
ness, and fear) on the second level. We excluded surprise, because it cannot be clearly
assigned to a positive or negative valence. On the third level, Shaver et al. (1987)
named 24 discrete emotions. Each of the emotions is assigned to one of the five basic
emotions. Figure 1 gives an overview of the hierarchical structure.
The coders received all data in one Excel sheet. They also got detailed instructions
about the hierarchy of the emotions and associated descriptions of each discrete emo-
tion on the third level (see Appendix 3). We added a “no emotion” category on the
first level to alleviate concerns that coders would code a unit as emotional even though
it was not. Compared to study 1, we did not offer a “not classifiable” category, but
instructed the coders to make up their mind about the most distinct emotion in one unit
and then code for this one. Further, the coders were equipped in the Excel sheet with
connected drop-down lists to simplify the coding process. When a coder decided on
the first level about positive, negative, or neutral valence, then the related drop-down
list became available for the second level with the two positive or three negative basic
emotions. A similar procedure was used when transitioning to the third level. If the
coder chose “no emotion” on the first level, no further decisions had to be made for the
other two levels. We instructed the coders to evaluate all thought units individually.

4.2 Results

Similar to study 1, Cohen’s kappa values indicate low intercoder reliability as shown
in Table 3. All values are considered as slight or fair (Landis and Koch 1977). Further,
Cohen’s kappa decreases across the three levels in all scenarios. The scores (0.16,
0.30, 0.15) for the discrete emotions on the third level mark low intercoder reliability
and thus replicate the results from study 1. The reliability scores for the comparison
between LIWC and the human coders paint a similar picture with very low scores even
on the first level (emotion valence).

123
392 C. Laubert

Table 3 Cohen’s kappa of study Case study Human coders LIWC


2
1st level

1st level 2nd level 3rd level Coder 1 Coder 2

Texoil .30 .23 .16 .19 .11


Websurf .40 .39 .30 .08 .07
Lattitude .24 .24 .15 .15 .19

Table 4 Emotion valence distribution (in %) of study 2

Emotion valence Texoil Websurf Lattitude

Coder 1 Coder 2 Coder 1 Coder 2 Coder 1 Coder 2

Positive 23.8 56.9 21.8 20.7 30.3 25.6


Negative 34.1 24.3 2.3 6.6 4.2 7.7
Neutral 42.1 18.8 75.9 72.7 65.5 66.7

Further insights provide the distribution of valence across the three scenarios
(Table 4). We find a relatively similar distribution in the Websurf and Lattitude data. In
this data, very few negative emotions are present, some positive emotions, but one out
of three units contain no emotion. This distribution is consistent between both coders.
The Texoil case study with its negative bargaining zone shows a different valence dis-
tribution. More negative and less neutral units are present. Both coders have a stronger
disagreement regarding the distribution of the valence in this scenario.
In total, 37 units (5.2%) were coded as opposite emotions by both coders. This is less
than in study 1, but it is still surprising that, for example, four times coder 1 thought it
is “enthrallment” and coder 2 judged it as “suffering” (see Appendix 2 for examples).

4.3 Discussion

The results from study 2 show that a hierarchical coding scheme does not substantially
improve coding outcome. Intercoder reliability is low, not only on the second or third,
but even on the first level (emotional valence). No consistent increase in intercoder
agreement from a negative (Texoil) towards a possible (Websurf) and then positive
(Lattitude) bargaining zone scenario was observed. Agreement between the coders
was highest for the case study with one single point of agreement (Websurf). This
is surprising because a single point of agreement leaves more room for arguments
and mixed emotions. A larger bargaining zone gives negotiators more opportunities
to agree and thus potentially reduces the need for mixed emotions, leading to higher
agreement scores.
There is one interesting demarcation between the single point and positive bar-
gaining zone cases towards the negative bargaining zone case: the amount of units
coded as “no emotion” increases substantially (from 42.1% to at least 65.5%) when

123
Are You Angry (Happy, Sad) or Aren’t You? Emotion… 393

an agreement is possible. Thus, even though we do not find a higher agreement rate
in the scenarios with a zone of possible agreement, we see that in such negotiations
fewer emotions are perceived overall. This might be due to less divergent interests and
preferences between the negotiators about the issues in the scenario. It is remarkable
that in the few units with mostly positive emotions the intercoder agreement about
the basic and specific emotions is still low. One reason for this finding could be the
level of analysis. Possibly, a thought unit is too narrow to contain all of the relevant
emotional information that is necessary for coders to form a precise judgment about
the emotions conveyed through the message. When only a few words are present to
make an assessment about the unit, coders have to rely on a single word or a single
sentence. Hence, in the next study we test if a broader level of analysis, e.g. using
message-level units, results in higher agreement scores.
This argument might also explain the low agreement rate between the human coders
and the LIWC scores. All six points of comparison across coders and scenarios do not
reach Cohen’s kappa values of 0.20. Thereby, these low scores indicate the agreement
on the first level, which focuses only on the valence of the units. As mentioned before,
this result might be due to the low data basis per unit LIWC can use for their analysis.
A thought unit contains between a few words and up to a few sentences. Thus, one
or a few strong words can bias the judgment about a whole unit if the context is not
considered. By using more data per unit, we can test to see if a bias exists and if so
we will see higher precision in decoding emotion valence between LIWC and coders
reached when using the message as the unit of analysis.

5 Study 3

5.1 Method

In study 3, we addressed the problem that a thought unit as unit of analysis may be
too narrow to allow for accurate emotion detection. Similar to previous research in a
CMC negotiation context we take a whole message as unit of analysis (Griessmair and
Koezsegi 2009). We want to exclude the possibility that an emotion is only difficult to
detect because it is nested in an isolated, short thought unit, but not in an entire email.
Reading and then coding a whole message may give a better, more accurate picture
of the feelings of an email sender. As basis, we use the transcripts from the previous
study and randomly draw three new transcripts from each scenario. These included 40
messages from Texoil, 44 messages from Websurf, and 32 messages from Lattitude.
All in all, 116 units are coded in this study 3.
These units were coded independently by two new research assistants.3 Similar
to study 2 a hierarchical classification scheme was applied, because it resulted in an
at least slightly higher consistency than flat classification (study 1) on the level of
discrete emotion. Furthermore, we instructed the coders to mark the words that were
3 We trained new research assistant coders for each new study as a way to begin to account for the possible
explanation that low consistency could be due to idiosyncratic mismatch of coders. While still recognizing
the limitations of the small sample of coders, we attempt to marginally address this limitation with new
coders for each study and then in our final study increasing our coders to eight total.

123
394 C. Laubert

most decisive for their coding decision. Coders again received an instruction sheet
with information about the coding process and detailed descriptions of each discrete
emotion. The units were also analyzed with LIWC. We expected the results to improve
because the program can base its decision on more words, i.e. more data per unit was
available. This should eliminate outliers that were created through single words in a
short thought unit. Such words may bias the valence of the unit towards this word,
even though the words do not reflect the overall emotional content of this thought unit.

5.2 Results

As Table 5 shows, Cohen’s kappa values are distributed similarly to study 2 on the first
level for the negative bargaining scenario (Texoil: 0.30). Reliability increases slightly
for the single point scenario (Websurf: 0.40 to 0.45) and decreases for the positive
bargaining scenario (Lattitude: 0.24 to 0.15). These Cohen’s kappa scores are judged
for: (a) Texoil as fair, (b) Websurf as moderate, and (c) Lattitude as slight., In all three
scenarios, scores for the second level reflect a slight strength of agreement.
The LIWC scores show low Cohen’s kappa for coder 1 and 2. In one case, (Texoil)
the Cohen’s kappa even indicates that the agreement by chance would have been higher
than the agreement reached between LIWC and coder 2. Results for the Lattitude data
are equal to chance level. The only barely acceptable scores are reached in the Web-
surf scenario on the first level between the human coders. In sum, we replicated the
inability to detect discrete emotions in email negotiations as found in study 1 and 2.
In Table 6, the intercoder matrix displays coding for the five basic emotion categories,
whereby this table also indicates the distribution across the three scenarios in paren-
theses (Texoil/Websurf/Lattitude). Thus, the numbers in the cell which indicates 9
agreements for ‘love’ shows that 5 of the agreements happened in the data from the
case study of Texoil, 3 in the Websurf, and 1 in the Lattitude. The percentages at the
right and bottom always refer to the total agreement rate including the data from all
three case studies.
The intercoder matrix shows that a relativly large number of 17 units (14.7%) are
rated contrary regarding emotional valence from both coders. Furthermore, the matrix
also gives information about the distribution on the first level of emotion: Coder 1
judged 56% as positive, 25% as negative, and 19% as no emotion; coder 2 judged
74% as positive, 17% as negative, and 9% as no emotion. Only in four out of 116 units
(3.5%), both coders agreed that there is no emotion in a message at all.

Table 5 Cohen’s kappa of study Case study Human coders LIWC


3
1st level

1st level 2nd level 3rd level Coder 1 Coder 2

Texoil 0.30 0.15 0.25 0.04 − 0.03


Websurf 0.45 0.15 0.04 0.17 0.28
Lattitude 0.15 0.13 0.10 0.00 0.00

123
Are You Angry (Happy, Sad) or Aren’t You? Emotion… 395

Table 6 Intercoder consistency-matrix of study 3


Coder 1 Coder 2

Love Joy Anger Sadness Fear No emotion Total Agreement (%)

Love 9 (5/3/1) 17 (3/9/5) 2 (1/1/0) 6 (3/2/1) – 2 (1/0/1) 36 (13/15/8) 25.0


Joy 8 (3/3/2) 23 (5/9/9) 3 (1/2/0) 1 (1/0/0) 1 (0/1/0) 14 (3/4/7) 50 (13/19/18) 46.0
Anger – 2 (1/1/0) 6 (2/4/0) 2 (0/2/0) 1 (1/0/0) 1 (1/0/0) 12 (5/7/0) 50.0
Sadness 1 (1/0/0) 1 (0/0/1) 1 (1/0/0) 2 (1/0/1) 1 (1/0/0) – 6 (4/0/2) 33.3
Fear – – 1 (0/1/0) – – 1 (0/0/1) 2 (0/1/1) 0.0
No emotion – 4 (3/0/1) 1 (0/1/0) – 1 (0/0/1) 4 (2/1/1) 10 (5/2/3) 40.0
Total 18 (9/6/3) 47 (12/19/16) 14 (5/9/0) 11 (5/4/2) 4 (2/1/1) 22 (7/5/10) 116
Agreement (%) 50.0 48.9 42.9 18.2 0.0 18.2

A bold figure indicates the number of agreed units between both coders

5.3 Discussion

Study 3 replicates results from the previous studies and provides new insights. First,
we find that it is possible to detect emotions on the first level on moderate acceptable
consistency score in the Websurf scenario, which corresponds with the result from
study 2. Furthermore, the results show that in a negative bargaining zone scenario the
text-based emotion detection consistency is low.
The results also reflect that the detection consistency is impaired on the second
and third level of emotions across all three scenarios (except Texoil 3rd level) as in
study 2. Thus, discrete emotions are also not perceived on an acceptable level when
the coders evaluated the emotional content of a whole message instead of a thought
unit. Email receivers can evaluate message valence to a certain degree (in the Websurf
scenario), but we argue that if the sender wants to use basic or discrete emotion as
a strategic negotiation tool, it seems to be advisable to write clearly in the email
(e.g., “I am angry about that offer”) to make sure its emotional content is conveyed
correctly.
Interestingly, the number of no emotion units is very low, indicating that emo-
tions are conveyed in a message, but not necessarily the ones intended by the
sender. Thus, the aforementioned problem of misinterpretation of emotions in email
negotiation is also present in this study. While coder 1 has a relatively balanced
perception of positive, negative and neutral emotions, coder 2 shows a positiv-
ity effect (74% of all units are coded positive). This effect has been identified
for a different coder in study 2, too. This points towards a possible recurring
effect.
The recurring low scores of the human coder/LIWC comparison requires further
explanation. Why are the scores on the first level exceptionally low or even below
chance level? It has to be questioned if a text analysis program can be applied to
detect emotions in a task with a lot of context such as a negotiation. It could be
that the specific communication task of a negotiation impedes the precise detec-

123
396 C. Laubert

tion of emotion valence in an email. On the other hand, if we consider the LIWC
scores as true because of its technical properties, why are individuals not able to
detect these emotions? What is the use of a text analysis program for emotion
detection in negotiation, if it does not coincide with the human perception of emo-
tion?
After conducting the three studies, in which we systematically tested variables
(classification scheme, level of analysis, negotiation scenario, computerized text anal-
ysis) to reach higher intercoder agreement, one aspect cannot be neglected, which
may negatively influence the emotion recognition consistency in email negotiations.
We only compared the coding of two individuals (using different coder pairs in each
study), who were external to the negotiation. Hence, we do not have an alignment
of which emotion the sender intended to convey and how it is perceived by the
receiver. Comparing a true value for an intended emotion with LIWC should result
in substantial increase of agreement rates. Because a person involved in a negotiation
knows how she felt while reading the message during the negotiation process and
how the negotiation evolved emotionally. Next, we address this issue by comparing
a negotiator previously involved in the negotiation with the coding of an indepen-
dent rater. Our aim is to get a better idea as to how “accurate” the emotion coding
is.

6 Study 4

6.1 Method

In the last study, we tested if satisfactory reliability scores can be reached if we use
the Websurf scenario and a message as a unit of analysis (which offered the highest
agreement rate), but also include the respective negotiators as part of the coding task.
Due to organizational/university constraints, we could only include one negotiator in
the coding process. We are therefore able to get a “true” value of the intended and
perceived emotion for at least one side. Furthermore, the negotiator will be able to rate
the emotions as they experienced them when receiving the messages. Comparing the
coding from somebody internal to the negotiation with the LIWC text analysis program
should result in a higher agreement rate or add further support for the difficulty and
discrepancies when detecting emotions in text-based communication.
The coding procedure was similar to study 3. We gathered new data using the
Websurf scenario as part of a class in Negotiation and Bargaining at two American
West Coast universities. The data was provided for the negotiator and external coder in
an Excel sheet and they also received detailed instructions about the emotion hierarchy
scheme and description about each discrete emotion (see Sect. 4.1 and 5.1 from study
2 and 3 in this paper for more detail). Doing the emotion coding as a negotiator was
voluntary. Seven student negotiators participated and coded a total of 107 messages.
A new research assistant was trained and coded all 107 messages as second coder.

123
Are You Angry (Happy, Sad) or Aren’t You? Emotion… 397

6.2 Results

Table 7 gives an overview of the results of study 4. There is no increase in agreement


on the first level compared to study 3. There is a slight increase in agreement from
study 3 to study 4 on the second and third level. However, as in all three studies before,
the Cohen’s kappa is low and far from an acceptable level following any benchmark
(Green 1997; Krippendorff 1980; Landis and Koch 1977). The LIWC comparison
does score higher when compared to the coding of the involved negotiators than the
external coder, but here, too, the score is close to the chance level and the difference
is minimal (0.04 vs 0.01). Hence, our results from the previous studies are once again
replicated and confirmed. Even if an individual who was involved in the negotiation
coded the data, there is no satisfactory agreement rate either with another external
coder or with the text analysis program.
The valence distribution is similar between the coder and negotiators. The coder
rated 55.1% as positive, 21.5% as negative and 23.4% as no emotion. The negotiators in
sum evaluated 53.3% as positive, 27.1% as negative and 19.6% as no emotion. Again,
a notable amount of units (17; 15.9%) are rated opposite regarding emotional valence
by the coder and the respective negotiator. That supports the previous finding and
shows the strong misfit between the perception of emotions in text-based negotiations
by individuals. In only 9.3% (9 units) of all units, both coders agreed that there is
no emotional content in a message. In all other units, at least one coder detected an
emotion. This finding was also present in study 1 and 2.

6.3 Discussion

With the results of the last study, another variable can be excluded as an alternate
explanation of the lack of agreement. The consideration of the negotiators participating
in the emotion coding process does not improve the intercoder agreement rate in any
substantial way. The low Cohen’s kappa between the negotiators and LIWC also raise
questions concerning the applicability of this program to the specific context of natural
communication in email negotiation.
Furthermore, the findings from this study confirm again the results from the previ-
ous three studies and provide further support for our conclusion that discrete emotions
cannot be reliably decoded in email negotiation. More importantly, we find that even
broader, basic emotions or, at the surface level, the emotion valence is difficult to
access correctly. Our systematic approach consistently reveals that there is a high
level of variance in both valence of emotions and their categorization into emotional
categories. To put our findings into context, other studies that investigate emotion

Table 7 Cohen’s kappa of study Cohen’s Negotiators versus coder LIWC 1st level
4 kappa
1st level 2nd level 3rd level Versus Versus
negotia- coder
tors

Websurf 0.33 0.21 0.21 0.04 0.01

123
398 C. Laubert

decoding in voice, photograph or video communication (e.g., Elfenbein and Ambady


2002: 58%) or affect vocalization (e.g., Schröder 2003: 81%) reached much higher
agreement rates in percent than in our study within text-based context. Even though
these studies relied on agreement rates as opposed to Cohen’s kappa, they reveal
that the extremely low agreement rates seem to be specific to text-based communica-
tion.
Last, the combination of a high amount of units with emotional information and
the relatively high amount of units with contrary coding by two independent coders
is a dangerous mixture. Because it implicates that often emotions are conveyed and
notably often two individuals extract contrary emotions from an email in negotiation.
This circumstance can negatively influence the negotiation process and outcome to a
substantial degree.

7 General Discussion

The findings of this research challenge the assumption that emotions can be decoded
above chance level in email negotiation. Not only are discrete emotions difficult to
be conveyed and perceived consistently, but also broader basic emotions and emotion
valence seem to be much more complex to understand correctly. Transferring such
emotional messages via a text-based communication medium seems inefficient, espe-
cially when the sender is keen on making the receiver aware of his or her emotions.
This insight has important implications, especially when considering the amount of
emails exchanged per day (see Radicati Group 2016). We want to further elaborate
on this argument: It needs to be addressed if, and to what extent, the social function
of emotion approach (Keltner and Haidt 1999) can be applied to email negotiation.
Based on our findings, it is not only difficult to convey the right emotion and thus
elicit a desired response from the counterpart, but there is also a high chance that
a different, undesired, and possibly even counterproductive emotion is perceived by
the receiver of an email. For example, imagine a negotiator considers accepting an
offer, but before doing that she wants to communicate sadness about an issue to
claim a bit more value (Sinaceur et al. 2015). However, the counterpart reads the
email and perceives anger, which can hinder a settlement if she also expresses anger
as an answer to that request (Friedman et al. 2004), quickly leading to an escala-
tion of arguments that is unintended by both sender and receiver and then leads to
impasse. This suggests an important contribution: the assumption that emotions are
reliably detectable in email is in question. Communication of emotion in email can
be ambiguous for many reasons (intentional subterfuge or lack of non-verbal or con-
text clues) and complex (intended emotions are masked but leak out). Many models
or frameworks assume accurate communication of emotion (see EASI model, Van
Kleef 2009); however, our research suggests that this notion needs further attention
and exploration.
Therefore, this research offers a new consideration to the social function of emo-
tions theory. Most of the writing in the area focuses on emotional expression serving
social functions conceptualized in a face-to-face environment where expressions like
facial, postural, gestural or speech cues communicate the emotion. And based on these

123
Are You Angry (Happy, Sad) or Aren’t You? Emotion… 399

common actions other individuals can interpret the emotion. For example, if someone
is angry they might increase the volume of their voice or storm out of a room and
slam the door. Our research is first evidence that text cues do not seem to consistently
indicate an emotion and therefore, over email, emotions are not performing the ‘social
function’ that they have been understood to perform. Importantly, individuals may
not be aware of this limit and, if an emotion is expected to elicit a particular contin-
gent action and it does not, this might escalate conflict, limit performance and hinder
social interaction leading to a social dysfunction of emotion in email negotiation. If
consistent and accurate emotion decoding is required to reliably know the thoughts
and feelings of others and then to regulate behavior, our research suggests that email
has quite a long way to go before it can be used in a way where emotions function as
an extra channel for solving problems.
Morris and Keltner (2000a, b) discuss cultural considerations as posing a chal-
lenge to the social function of emotions as different “codes” of sender and receiver
could lead to an emotional communication breakdown. This is precisely what our
research suggests but substituting culture for communication mode. We would sug-
gest that the communication using a text-based mode confuses the “codes” that
would normally clearly communicate one emotion or another. Too much research
assumes that the emotional codes in email are being interpreted correctly and consis-
tently.
With this issue—communicating accurately via email—we address a topic that has
been investigated in a different context before. The low accuracy of decoding emotions
in email communication can be due to egocentrism (Krueger et al. 2005) or lead to
negativity or neutrality effects (Byron 2008). In our research, we can add some support
to Byron’s (2008) argument for a neutrality effect. Further, we tentatively suggest
considering a positivity effect in a negotiation context, thereby extending Byron’s
model. Our data shows that multiple coders tend to evaluate a majority of units in
a negotiation as positive emotions. Though we could not investigate the reasons for
this behavior in this research, we think it is worthwhile to study factors leading to
this effect. According to Byron (2008) it could be due to specific sender, receiver or
contextual factors. Future research may address this issue. One possible explanation
for this positivity effect could be that coders assume that both negotiators are keen
to settle on some form of agreement, naturally expecting positive emotions along the
road to agreement. However, skilled negotiators rely on a full arsenal of emotions to
convey their message effectively.
In hand with the positivity effect, the observed neutrality effect also deserves atten-
tion. Do some senders actively try to mask their emotion in communication since
email gives them time to cool down before sending the communication (Pesendorfer
and Koeszegi 2006)? If sometimes emotions are less present in email negotiation,
are the few conveyed emotions more salient and hence more important for the pro-
cess? Especially then, the negotiator should avoid the chance of misinterpretation
(Kruger et al. 2005) and assure that the intended emotion is accurately sent and
perceived. In this case, clear, specific formulation of the text is necessary. Indi-
viduals need to keep in their mind that non-verbal signs (e.g., body posture, eye
contact, facial expression) are absent and will not help to convey and decode emo-
tions.

123
400 C. Laubert

With having fewer cues to convey emotions (Hatta and Ken-ichi 2008; Walther and
Parks 2002), it is important to ask how negotiators can get better at emotion detection
in text-based interactions. Similar to other research (Griessmair and Koeszegi 2009),
we find that if emotions are present in text-based negotiation, the only chance for
individuals is to be more sensitive to the few cues still left in the email medium.
Factors that increase decoding accuracy and consistency may be the familiarity with
the communication medium, with the negotiation task, with the counterpart, or with
the issues being negotiated. Additionally, emotional intelligence may be a predictor
for emotion detection in a text-based context and could be examined in future research.
Because with higher emotional intelligence, the correct recognition of emotions should
increase (Salovey and Mayer 1990).
Our findings also add to the theory of emotion recognition accuracy (Elfenbein and
Ambady 2002; Hawk et al. 2009). The data provides further support that hierarchical
classification schemes are more useful than flat classification (Ghazi et al. 2010).
Additionally, the highest score was reached when coding whole messages. In the
context of emotion, a small unit such as a thought unit seems to provide too little
context to accurately evaluate discrete emotions. Even though the number of coders
involved across all studies is comparatively low (n  14), the overall number of
coded units (n  2258) should substantively corroborate the conclusions made in this
exploratory research.
Another interesting observation is that only four times in 2258 thought units
a negotiator used a smiley/emoticon. It is possible that the paucity of emoti-
cons to express emotions may have lead to a lower accuracy of emotion detec-
tion. However, current research (Glikson et al. 2018) indicates that using smileys
in work-related context is not recommended. These researchers found that peo-
ple using smileys were perceived as less competent. This finding could be an
explanation for why our study participants did not use emoticons in their negotia-
tion.
Surprisingly, the results comparing human coders to the text analysis program
of LIWC show very low reliability scores. Even though LIWC has been shown to
be a valid instrument for such classification tasks (Tausczik and Pennebaker 2010),
its application to the context of emotion decoding in email negotiation has to be
questioned. Across four studies, including comparisons of 14 different coders and
14 different data sets, Cohen’s kappa values never surpassed 0.28 on the most
abstract level of emotion valence. Our study is the first to compare LIWC scores
to human coders in a negotiation context and hence requires further research. It
is disputable if computerized text analysis can be successfully applied to extract
complex emotions. In our studies, LIWC performed not as good as the human detec-
tion of emotion in the email negotiation context. Even one of the developers of the
LIWC software states that such programs are prone to neglecting important aspects
such as context, irony, or idioms (Tausczik and Pennebaker 2010). Further, it could
be interesting to compare other categories of LIWC with human coding and see if
agreement rates will be better in these categories. In addition, future research could
consider using more robust text analysis software (e.g., NVivo) or dictionaries that
have more complexity in emotional content (e.g., WordNet Affect) (see Gupta et al.
2013).

123
Are You Angry (Happy, Sad) or Aren’t You? Emotion… 401

While we allowed coders in study 1 to code for “mixed emotions” and found low
agreement, there still exists the possibility that low emotion detection consistency in
studies 3 and 4 can be partially explained by the whole message containing more
than one emotion (see examples in Appendix 2). In our design for study 3 and 4, we
asked coders to always decide for only one emotion. Hence, the low agreement rate
might result from different individuals focusing on or finding one particular emotion
more salient than another emotion in email negotiation. This could be something
to investigate in future research. That said, regardless of whether the low detection
consistency results from multiple emotions or different interpretation, it seems clear
that the interpretation of emotional content in a text-based communication medium
is complicated. If discrete emotions are layered within each other, it can explain that
the emotion detection consistency decreases from first to third level. Further, layered
emotions would indicate that the sender has even less control over the emotion that
will be perceived by the receiver. As it appears that different emotions are more
salient for different individuals (Walther and Addario 2001). Thus, sending more
than one emotion in an email can have a disadvantage, because the sender does not
know how many and which emotion a receiver picks up and eventually will react
to.
Based on our findings, we would suggest that negotiators consider the likelihood
of misinterpretation of emotions in text-based negotiations. It might be helpful for
negotiators to directly address their emotional state by writing “I am angry/happy/sad
about that offer” instead of using masked or indirect formulations which gives leeway
for misinterpretation by the counterpart. In our email negotiation, we did not find many
words used by the individuals that directly describe emotions (e.g., angry/happy/sad).
It also might be helpful for a counterpart to ask for emotional clarification from
the other party. For example, “I’m sensing the most recent package offer is anger-
ing you. Is that correct?” or “Am I correctly understanding that you are happy
with the proposal?”. Additionally, we do not know if any negotiator wanted to use
emotions strategically. Nevertheless, using direct formulations will give the email
receiver an exact picture about the emotions of the sender, if the sender wishes to
share their emotions. Further, even if a sender does not want to convey an emo-
tion, they always needs to consider that the email may not be perceived neutral
by the receiver. Thus, if much is at stake, and emotions are already intense, more
deliberate consideration may be needed to avoid escalation or missed opportunities.
Negotiators might also want to consider if email is the most effective medium in
negotiation should emotional communication be key to reaching a mutual agree-
ment.

8 Limitations and Future Research

Despite our efforts to systematically vary a number of factors to exclude several


alternative explanations regarding why unanimous emotion detection was low, our
studies have some limitations. First, the first three studies only used two independent
coders each for the coding task. Though considering the seven negotiators in the
fourth study and thus having a total of 14 human coders participating across all our

123
402 C. Laubert

studies, we cannot draw statistical conclusions from the data. Future research may
address this methodological issue by letting a high number of coders evaluate one
email negotiation, resulting in “crowd-coding” of the negotiation messages (Benoit
et al. 2012). In this case, a high number of subjects should evaluate and code for
emotions of a few email negotiation transcripts. In this study, an overall and pairwise
comparison helps to gain deeper insight into the level of emotion detection consistency.
We suggest to use a hierarchical classification scheme, a whole message unit and a
negotiation with a possible zone of agreement, because it provided the highest emotion
detection consistency across all our studies. Such a quantitative future study would
lend more robust evidence that would either corroborate or contradict the findings from
our exploratory research investigating the reliability of text-based emotion detection
consistency. Simultaneously, data about the coders could be collected and give insights
if the impact variables such as gender, age, or emotional intelligence have an influence
on the intercoder agreement rate. It could also be interesting to incorporate a scale
about the attitude towards the communication medium (e.g., Geiger and Parlamis
2014 or Kelly and Keaten 2007) in such a study. If individual differences between
a sender and receiver of an email can explain the problem of decoding emotions
accurately, it will have an influence in how we should communicate with each other
in a negotiation.
Another limitation of our study pertains to the circumstance that we did not get
both negotiators to code for emotion during the negotiation process. Because this is
an exploratory study about emotion decoding in email negotiation, we think that we
could get first insights by analyzing the coding which only included one negotiator.
However, future studies should incorporate data from both involved negotiators for
further validation of these findings.
For future research, it would also be interesting to look at relationships between
the level of emotion detection and the outcome of the negotiation. Can more value be
created if more emotions are correctly perceived? If there is no relationship between
more correctly perceived emotions in email and a successful negotiation, are discrete
emotions then important in email negotiations at a significant level? Or is conveying
a general valence of emotion sufficient to more successfully negotiate as shown by
previous research (Hine et al. 2009)? Last, future research can also address the issue
that more than one discrete emotion is present in a coding unit and thus allow to code
for multiple emotions.

9 Conclusion

Two implications of this research make significant contributions to the literature on


email negotiation specifically as well as email exchanges more generally and should
be underscored. First, with the finding that emotion detection in email negotiations
is unreliable and inconsistent, it is suggested that emotions cannot be used in this
context for strategic purposes or to serve the critical social function of influenc-
ing the behavior of another (see Morris and Keltner 2000a, b; Van Kleef et al.
2010 for more on the social function of emotion). This understanding of the func-
tional limits of email, from an emotional perspective, sheds greater light on the

123
Are You Angry (Happy, Sad) or Aren’t You? Emotion… 403

perennial problems and challenges faced in email exchanges. Second, in practice,


more explicit emotional articulation could be a simple and effective prescription
for email negotiations and email exchanges in general. Clearer emotional com-
munication may be a key to email negotiation success. The implications of these
findings to the real world of business communication and negotiation are criti-
cal because of the billions of emails sent and received each day (Radicati Group
2016).
The next step in the research stream is to replicate these findings with a larger
sample, identify and test moderators and interventions that might improve the emotion
communication and detection in email negotiations, and then link these with subjective
and objective outcomes.

Acknowledgements We thank Ingmar Geiger, Sabrina Gottschalk, and Alexander Mafael for commenting
on earlier versions of this article.

Appendix 1: Coding Manual Study 1

Dear Coder,
We highly appreciate your support for coding our data. This instruction sheet helps
you to conduct the coding. You have received two files. This Word document and an
Excel file. This Word document offers you an instruction how to code the data in the
Excel file and offers a table with emotions and their characteristics. It may be helpful
to print out this table and put it next to you on your desk while you are doing the
coding in the Excel sheet.
To start coding please open the Excel file. This Excel file contains data in the
column A (row 2 to 1372). In this column, you find text pieces from 23 email negoti-
ations. Each negotiation is divided by a yellow cell. Within each negotiation, different
emails are separated by a black line. Between two black lines, the email text is divided
into text pieces (e.g. thought units) by using different cells. You do not need to pay
any attention to the red numbers. In row 1, you find in column B to I the names of
eight emotions we are interested in. These are the emotions you also find in the
table below. Column J asked you to put in a signal (key words, brief sentence),
why you decided to code one thought unit as a specific emotion. Column K offers
you the opportunity to code a thought unit as “containing an emotion but not being
classifiable into the categories (B to I)”; Column L is the category “no emotion pre-
sent”.

Step 1

Please have a look at each thought unit in column A separately. You have to decide
if you can categorize one thought unit into one of the eight emotions (B to I). Please
have a look at the table for a description of each emotion. If you can classify a thought
unit to one emotion, put an “x” into the corresponding cell. For example, if you say
the thought unit in row 4 contains the emotion “contempt”, please put an “x” in cell

123
404 C. Laubert

“E4”. Please rate every of the 1317 thought units. Of course, it may happen regularly
that one unit does not contain any emotion. In this case, please put an “x” in column L.
It is important that you evaluate all thought units individually and do not connect
your decision for one unit with the unit before that one. Each thought unit should be
evaluated isolated.
Step 2

When you have decided to code one thought unit with one emotion, we ask you to put a
short explanation into column J in the same row. Here, please write down the aspect that
has influenced you to decide to put a thought unit into that specific emotion category.
It might be a certain word in the unit, a punctuation character, length of a sentence, use
of adjectives, or any other aspect that affected your decision. Please consider there is
no right or wrong in this coding. We are interested in your well-thought-out subjective
evaluation.
We know there are many thought units to code. Therefore, please take a break every
now and then to refresh and regularly check with the table below if your evaluation
still fits to the characteristics below. If you have any questions regarding the procedure,
please do not hesitate to contact us.
Best regards,

Emotion Description from Possible characteristic in Example


dictionaries negotiation

Liking (1) The feeling of liking (1) Enter in cooperative (1) “I agree to have
or enjoying behavior performance metrics, I
something/someone (2) Express affiliation think you have a good
(2) A feeling of being towards the counterpart idea here.”
pleased with someone (2)”Dear Mr. Smith, We
or something are delighted to hear
that you plan to
continue to serve our
community in case of
purchase”
Interest (1) A feeling of wanting (1) Ask personal (1) “I am limited to the
to learn more about questions $400K range for the
something or to be (2) Flatter the other price. Is there anything
involved in something person else we can do to make
(2) A quality that attracts the deal work? Can you
your attention and tell more about your
makes you want to learn costs?”
more about (2) “Are you and your
prestigious firm
interested in continuing
the service?”

123
Are You Angry (Happy, Sad) or Aren’t You? Emotion… 405

Emotion Description from Possible characteristic in Example


dictionaries negotiation

Anger (1) A strong feeling of (1) Hostile criticism of (1) “Unfortunately I am


being upset or annoyed other’s point starting to feel that
because of something (2) Talk in idioms of we’re not on the same
wrong or bad power and dominance page. Your offer is not
(2) The feeling that makes competitive and we are
someone want to hurt still better off by doing
other people, to shout, business with either BP
etc or Shell”
(2) “However, to be frank,
the offers you are
providing me are
borderline ridiculous”
Contempt (1) A feeling that (1) Talk with a dismissive (1) “That really surprises
someone or something tone me as the value of a
is not worthy of any (2) Tell the counterpart business is not
respect or approval that he is not worthy to determined by its
(2) A lack of respect for negotiate with physical replacement
or fear of something cost, but also all of the
that is usually respected intangibles. In addition,
or feared our station meets all
environmental and other
regulations meaning
that an equipment
upgrade is not even
necessary”
(2) “I think it is an
amazing opportunity
given our growth here,
but considering how low
the offer is, I believe I
should look elsewhere”
Embarrassment (1) The state of feeling (1) Apologize or signal (1) “I apologize if there
foolish in front of others regret was any
(2) Something or (2) Give a brief miscommunication. I
someone that causes a expression of had meant “at least” to
person or group to look embarrassment indicate less than
or feel foolish $500,000”
(2) “I’m afraid you have
lost me a bit. I am just
small businessman so I
am not sure about all
the terms you have
mentioned”

123
406 C. Laubert

Emotion Description from Possible characteristic in Example


dictionaries negotiation

Empathy (1) The feeling that you (1) Create mutual (1) “Thank you and we
understand and share positivity and interest appreciate your
another person’s (2) Agree with the generosity and great
experiences and counterpart doing business with
emotions you”
(2) That ability to share (2)”If you feel that a deal
someone else‘s feelings between us does not
make sense I completely
understand and would
wish you the best of
luck in constructing
your new site”
Pain (1) Mental or emotional (1) Assault the (1) “This offer is an insult
suffering : sadness counterpart and hurts my feelings”
caused by some (2) Express that you are (2) “I am a Texoil
emotional or mental hurt representative. I was
problem extremely saddened to
(2) Someone or hear that you are
something that causes planning on leaving the
trouble or makes you Texoil family of service
feel annoyed or angry station owners: (“
Exasperation (1)The state of being very (1) Show surprise and/or (1) “Please advise as soon
annoyed or upset strong anger as possible whether you
(2) The state of being (2) Express indignation can meet my requested
exasperated; frustrated price or if we should
annoyance just terminate this
discussion”
(2) “I can’t believe it. I
really don’t understand
how you can demand
that much. That makes
me mad”
Mixed emotion Please evaluate a thought unit as mixed emotion if: (1) more than one of the eight
above mentioned emotions is present or (2) an emotion is present that is not listed
above (e.g., joys, fear)

Appendix 2: Quotes of Contrary Coding for Studies 1–4

Quote Coder 1/Coder 2

Study 1 (thought units)


“We think you have done a tremendous job in managing it and would have Liking/pain
loved to keep you as a business partner going forward, therefore we are
sorry to hear you will be selling it”

123
Are You Angry (Happy, Sad) or Aren’t You? Emotion… 407

Quote Coder 1/Coder 2

“Please let me know what is the rationale for this number as I’ve explained Anger/interest
mine and I would like to understand why a multinational corporation like
yours is “limited” to bid $450k when their best alternative is paying
$675k for a new station”
Study 2 (thought units)
“Me and my wife can only take advantage of some its value really because Optimism/sadness
we don’t have the capital or energy required to maintain the station open
24/7 and open a minimart. But a company like Texoil certainly can do it
and reap all those benefits”
“Unfortunately I can’t pay that much and will have to walk away” Pride/disappointment
Study 3 (whole message units)
“S., Affection/shame
After some thought and advising with my team, I have decided to accept
the $500,000 price tag.
I also want to offer my apologies if any of my correspondence came off as
abrasive. I was merely working with the advice given to me and the
limited information I had about your business. I’m sure you were as well.
M.”
“Dear Ms. G., Disappointment/affection
It is unfortunate that a deal could not be done at this moment.
In spite of this, we would still like to thank you and your husband for your
continued dedication and hard work. Hopefully, we will meet again when
new opportunities arise.
Again, thank you, and best of luck to your endeavors.
Best,
E.L.”
“Dear T., Irritation/cheerfulness
I understand your concern over the potential costs to your business should
you change your URL. May I suggest that we reach a compromise, where
the old URL may remain as a redirect page for consumers to clarify
whether they would like to proceed to your new website domain, or
whether they were intending to visit Websurf.com? This seems a
reasonable solution to ensure that during the transitional period, perhaps
6-12 months, our respective customers are able to find their particular
website?
With regards to the compensation, I must say that I am a little disappointed
with the figure that you quoted. Per our research, this sum would amount
to the income generated by your website over 16 years. Further we are
simply incapable of being able to compensate you at such a high sum. I
understand that you have a significant investment of time and energy in
the domain and in your establishment, however, and would be willing to
double our original offer to $100,000, and create a transition period as
outlined above.
I am eager to settle this matter in a productive way, but am concerned that
you are not interested in taking these negotiations seriously. It really
should serve both of our interests to guarantee that our respective
customers are not confused by the striking similarity between our website
domains. My proposal would guarantee you a large cash advance to
expand your business, and a transition period to guarantee neither of us
lose our clients.
I look forward to your response.
Sincerely,
D.”

123
408 C. Laubert

Quote Coder 1/Coder 2

Study 4 (whole message unit)


“Hi A., Disgust/pride
Thank you for your understanding! However, I would not want to sell my
domain name as changing domain for the web might affect my customer
notification and customer engagement toward my website. This action
will cause a lot of harm to my business. Why don’t you consider
changing your website name instead? So, our website will not have a
duplicate name.
Warm Regards,
J.”
“S., Optimism/irritation
I like your idea of supporting small businesses with your service for free.
And of course your company would be named partner in this venture. I
originally said we could settle on 3 years now and then re-evaluate if we
think it is a working and worthwhile relationship. How much business do
you think your team could handle per year?
I can work with a banner popping up for 5 s when someone lands on the
homepage. However, I don’t think “Click here for surfing in Hawaii”
would speak to my brand. We offer more than just surfing and I want to
make sure that my customers recognize the brand and don’t quickly
dismiss it for some other ad, especially if it will only be on the screen for
5 s and there is no way to find that information again.
Although we seem to be working toward an agreement for the banner, i will
still incur losses from customers that may not visit the site on a regular
basis. Are you thinking that this banner would be up indefinitely? If any
of my customers don’t see the notification I could potentially lose them
altogether. Additionally, I’ll still have to make other marketing efforts to
promote the change. Moreover, I will have to work extra hours along with
my team to make sure a switch to a new domain goes smoothly. My
company typically makes $200,000 in profits per year. But given the
overtime I’m expecting my team to work, marketing efforts, potential
customer losses altogether (even with the help of the flash ad), sunk costs
of inventory, I expect that I will suffer loses of $600,000 that I otherwise
would not have incurred by selling my domain. I’m not trying to give you
a run for your money, just simply trying to make sure that this domain
switch does not severely affect my business in the long term.
I hope this has shed some light into my thought process. I look forward to
hearing from you
Sincerely,
M.”
“Good Morning, Nervousness/optimism
My name is D.S. and I recently received some interesting news from your
lawyer. My understanding is that you would like to purchase my current
domain name. I have had this name for quite some time now, so my
question to you would be why do you want to purchase my domain and
what do you intend to use it for?
If you are truly interested in buying my domain name, I would be willing to
sell it for $3 million.
I look forward to hearing from you.
Sincerely,
D.S”

123
Are You Angry (Happy, Sad) or Aren’t You? Emotion… 409

Appendix 3: Coding Manual Study 2

Dear Coder,
We greatly appreciate your support in coding our data. This Word document offers
you instructions on how to code the data in the Excel file and provides a table with
descriptions of the emotions we are interested in. It may be helpful to print out this
table and put it next to you on your desk while you are coding.
Please open the Excel file to start the coding. This Excel file contains data in the
column A. In this column, you find one single messages from fifteen email negotiations.
Each negotiation is divided by a yellow row. Within each negotiation, each cell includes
one whole email message.
In column B to D you will find drop-down lists for three different levels of emotions.
These emotions are described in the table below. On the first level, we ask you to judge
the text as “positive emotion”, “negative emotion”, or “no emotion”. On the second
level, you can chose between two positive and three negative basic emotions. On the
third level, we ask you to code for a more specific emotion of the prior chosen basic
emotion. Even if no option on the 2nd level seems to fit 100% to your perception of the
emotion, please chose one of the five basic emotions, which you think is the closest to
your opinion. It is important that you make a decision on all three levels (exception:
you chose “no emotion” on the 1st level). In other words, if you choose “no emotion”
on the first level, you do not need to make a 2nd or 3rd level decision.
Step 1 Please get familiar with the descriptions of the emotions below. Take your
time and make yourself aware of the differences among the emotions.
Step 2 Have a look at each message in column A in the Excel sheet. You have to
decide if you can categorize one message regarding its emotional direction (1st level:
positive/negative), its basic emotion (2nd level: love, joy, anger, sadness, fear), and its
specific form (3rd level).When you click on a cell, a drop-down list will appear. Only
if you chose an option from the first level, the drop-down list for the second level will
be activated and so on. Please rate every of the messages. It may happen regularly
that one message does not contain any emotion. Then chose “no emotion” on the first
level. If you think a message includes more than one emotion, please decide which
emotion is most distinct and code for this one. It is important that you evaluate all
message individually, but also consider the context (i.e., the general atmosphere in the
conversation) that is present. Code the emotions according to how you as a receiver
of the message feel about the emotion.
Step 3 After you coded a message on all three levels, we ask you to mark the
words red, which you identified as a signal for the (most distinct) emotion. It might
be a certain word in the message, a combination of words, or two words that are not
directly next to each other. Please consider there is no right or wrong. We are interested
in your well-thought-out subjective evaluation.
Please regularly check with the table below if your evaluation still fits to the emotion
descriptions. If you have any questions regarding the procedure, please do not hesitate
to contact us.
Best regards,

123
410 C. Laubert

1st level 2nd level 3rd level Description

Positive emotion Love Affection/liking A feeling of liking and caring for someone or
something
Lust A strong feeling of sexual desire
Longing A strong desire for something or someone
Joy Cheerfulness Feeling or showing happiness
Zest Lively excitement : a feeling of enjoyment and
enthusiasm
Contentment The state of being happy and satisfied
Pride A feeling that you respect yourself and deserve to be
respected by other people
Optimism A feeling or belief that good things will happen in the
future
Enthrallment To hold the attention of someone by being very
exciting, interesting, or beautiful
Relief A pleasant and relaxed feeling that someone has when
something unpleasant stops or does not happen
Negative emotion Anger Irritation To make someone impatient, angry, or annoyed
Exasperation The state of being very annoyed or upset
Rage A strong feeling of anger that is difficult to control
Disgust A strong feeling of dislike for something that has a
very unpleasant appearance, taste, smell, etc
Envy The feeling of wanting to have what someone else has
Torment Extreme physical or mental pain
Sadness Suffering Pain that is caused by injury, illness, etc. : physical,
mental, or emotional pain
Sadness Affected with or expressive of grief or unhappiness
Disappointment Failing to meet expectations
Shame A feeling of guilt, regret, or sadness that you have
because you know you have done something wrong
Neglect To fail to take care of or to give attention to (someone
or something)
Sympathy The feeling that you care about and are sorry about
someone else’s trouble, grief, misfortune
Fear Horror A very strong feeling of fear, dread, and shock
Nervousness Having or showing feelings of being worried and
afraid about what might happen
No emotion

References
Allred KG, Mallozzi JS, Matsui F, Raia CP (1997) The influence of anger and compassion on negotiation
performance. Organ Behav Hum Dec Process 70(3):175–187
Barry B, Oliver RL (1996) Affect in dyadic negotiation: a model and propositions. Organ Behav Hum Dec
Process 67(2):127–143

123
Are You Angry (Happy, Sad) or Aren’t You? Emotion… 411

Barry B, Fulmer IS, Van Kleef GA (2004) I laughed, I cried, I settled: the role of emotion in negotiation.
The handbook of negotiation and culture, pp 71–94
Bazerman MH, Curhan JR, Moore DA, Valley KL (2000) Negotiation. Annu Rev Psychol 51(1):279–314
Benoit K, Conway D, Laver M, Mikhaylov S (2012). Crowd-sourced data coding for the social sciences:
massive non-expert human coding of political texts. In: Presentation at the 3rd annual new directions
in analyzing text as data conference, Harvard University
Brett JM, Olekalns M, Friedman R, Goates N, Anderson C, Lisco CC (2007) Sticks and stones: language,
face, and online dispute resolution. Acad Manage J 50(1):85–99
Byron K (2007) Male and female managers’ ability to read emotions: relationships with supervisor’s per-
formance ratings and subordinates’ satisfaction ratings. J Occup Organ Psychol 80(4):713–733
Byron K (2008) Carrying too heavy a load? The communication and miscommunication of emotion by
email. Acad Manage Rev 33(2):309
Byron K, Baldridge DC (2005) Toward a model of nonverbal cues and emotion in email. In: Academy of
management proceedings, vol 2005, No. 1, pp B1–B6. Academy of Management
Calvo RA, Mac Kim S (2013) Emotions in text: dimensional and categorical models. Comput Intell
29(3):527–543
Cheshin A, Rafaeli A, Bos N (2011) Anger and happiness in virtual teams: emotional influences of text
and behavior on others’ affect in the absence of non-verbal cues. Organ Behav Hum Dec Process
116(1):2–16
Derks D, Fischer AH, Bos AE (2008) The role of emotion in computer-mediated communication: a review.
Comput Hum Behav 24(3):766–785
Druckman D, Olekalns M (2008) Emotions in negotiation. Group Decis Negot 17(1):1–11
Ebner N (2011) Negotiating Via Email. In: Benoliel M (ed) Negotiation excellence: successful deal making.
World Scientific Publishing, Singapore, pp 397–415
Ekman P (1992) An argument for basic emotions. Cogn Emot 6(3/4):169–200
Elfenbein HA, Ambady N (2002) On the universality and cultural specificity of emotion recognition: a
meta-analysis. Psychol Bull 128(2):203
Elfenbein HA, Der Foo M, White J, Tan HH, Aik VC (2007) Reading your counterpart: the benefit of
emotion recognition accuracy for effectiveness in negotiation. J Nonverbal Behav 31(4):205–223
Filzmoser M, Hippmann P, Vetschera R (2016) analyzing the multiple dimensions of negotiation processes.
Group Decis Negot 25(6):1169–1188
Fleiss JL (1981) Statistical methods for rates and proportions, 2nd edn. Wiley, New York
Folger J, Hewes DE, Poole MS (1984) Coding social interaction. In: Dervin B, Voigt M (eds) Progress in
communication sciences. ABLEX, Norwood, pp 115–161
Friedman RA, Currall SC (2003) Conflict escalation: dispute exacerbating elements of e-mail communica-
tion. Hum Relat 56(11):1325–1347
Friedman R, Anderson C, Brett J, Olekalns M, Goates N, Lisco CC (2004) The positive and negative
effects of anger on dispute resolution: evidence from electronically mediated disputes. J Appl Psychol
89(2):369–376
Geiger I, Parlamis J (2014) Is there more to email negotiation than email? The role of email affinity.
Computer Hum Behav 32:67–78
Ghazi D, Inkpen D, Szpakowicz S (2010) Hierarchical versus flat classification of emotions in text. In:
Proceedings of the NAACL HLT 2010, Association for computational linguistics, pp 140–146
Glikson E, Cheshin A, Kleef GAV (2018) The dark side of a smiley: effects of smiling emoticons on virtual
first impressions. Soc Psychol Personal Sci 9(5):614–625
Goldberg SB (1998) Texoil. Dispute Resolution Research Center Negotiation Exercises. https://www.
negotiationexercises.com/Details.aspx?ItemID=76 [02/28/17]
Green AM (1997) Kappa statistics for multiple raters using categorical classifications. Proc 22nd Annu
SAS User Group Int Conf 2(3): 4
Griessmair M, Koeszegi ST (2009) Exploring the cognitive-emotional fugue in electronic negotiations.
Group Decis Negot 18(3):213–234
Griessmair M, Hippmann P, Gettinger J (2015) Emotions in e-negotiations. In: Emotion in group decision
and negotiation, Springer Netherlands, pp 101–135
Gupta N, Gilbert M, Di Fabbrizio G (2013) emotion detection in email customer care. Comput Intell
29(3):527–543
Gwet KL (2014) Handbook of inter-rater reliability: the definitive guide to measuring the extent of agreement
among raters. Advanced Analytics, LLC, Arlington

123
412 C. Laubert

Hancock JT, Gee K, Ciaccio K, Lin JMH (2008) I’m sad you’re sad: emotional contagion in CMC. In:
Proceedings of the 2008 ACM conference on Computer supported cooperative work, pp 295–298
Harinck F, Van Kleef GA (2012) Be hard on the interests and soft on the values: conflict issue moderates
the effects of anger in negotiations. Br J Soc Psychol 51(4):741–752
Hatta T, Ken-ichi O (2008) Effects of visual cue and spatial distance on exitability in electronic negotiation.
Comput Hum Behav 24(4):1542–1551
Hawk ST, Van Kleef GA, Fischer AH, Van der Schalk J (2009) “Worth a thousand words”: absolute and
relative decoding of nonlinguistic affect vocalizations. Emot 9(3):293
Hine MJ, Murphy SA, Weber M, Kersten G (2009) The role of emotion and language in dyadic e-
negotiations. Group Decis Negot 18(3):193–211
Kahn JH, Tobin RM, Massey AE, Anderson JA (2007) Measuring emotional expression with the linguistic
inquiry and word count. Am J Psychol 120(2):263–286
Kelly L, Keaten JA (2007) Development of the affect for communication channels scale. J Commun
57(2):349–365
Keltner D, Haidt J (1999) Social functions of emotions at four levels of analysis. Cognit Emot 13(5):505–521
Keltner D, Kring AM (1998) Emotion, social function, and psychopathology. Rev Gen Psychol 2(3):320
Koeszegi ST, Pesendorfer E-M, Vetschera R (2011) Data-driven phase analysis of e-negotiations: an exem-
plary study of synchronous and asynchronous negotiations. Group Decis Negot 20(4):385–410
Kopelman S, Rosette AS, Thompson L (2006) The three faces of Eve: strategic displays of positive, negative,
and neutral emotions in negotiations. Organ Behav Hum Dec Process 99(1):81–101
Krippendorff K (1980) Content analysis. Sage, Beverly Hills
Kruger J, Epley N, Parker J, Ng ZW (2005) Egocentrism over e-mail: can we communicate as well as we
think? J Personal Soc Psychol 89(6):925
Landis JR, Koch GG (1977) The measurement of observer agreement for categorical data. Biom
33(1):159–174
Lelieveld GJ, Van Dijk E, Van Beest I, Van Kleef GA (2013) Does communicating disappointment in
negotiations help or hurt? Solving an apparent inconsistency in the social-functional approach to
emotions. J Personal Soc Psychol 105(4):605–620
Lewicki RJ, Barry B, Saunders DM (2011) Essentials of negotiation. McGraw-Hill/Irwin, Boston, MA
Li W, Xu H (2014) Text-based emotion classification using emotion cause extraction. Expert Syst Appl
41(4):1742–1749
Manwaring M (2003) Lattitude.com, Harvard Law School, Program of Negotiation, Cambridge
Mehl MR, Pennebaker JW (2003) The sounds of social life: a psychometric analysis of students’ daily
social environments and natural conversations. J Personal Soc Psychol 84(4):857
Morris MW, Keltner D (2000). How emotions work: the social functions of emotional expression in nego-
tiations. Res Organ Behav, 22:1–50
Morris MW, Keltner D (2000b) How emotions work: the social functions of emotional expression in
negotiations. Res Organ Behav 22:1–50
Morris M, Nadler J, Kurtzberg T, Thompson L (2002) Schmooze or lose: social friction and lubrication in
e-mail negotiations. Group Dyn: Theory Res Pract 6(1):89
Notini J (2008) Websurf negotiation. Berkeley Law School, Berkeley
Olekalns M, Druckman D (2014) With feeling: how emotions shape negotiation. Negot J 30(4):455–478
Overbeck JR, Neale MA, Govan CL (2010) I feel, therefore you act: intrapersonal and interpersonal effects of
emotion on negotiation as a function of social power. Organ Behav Hum Dec Process 112(2):126–139
Parlamis JD, Geiger I (2015) Mind the medium: a qualitative analysis of email negotiation. Group Decis
Negot 24(2):359–381
Pennebaker JW, Francis ME, Booth RJ (2001) Linguistic inquiry and word count: LIWC 2001. Lawrence
Erlbaum Associates, Mahway
Pesendorfer EM, Koeszegi ST (2006) Hot versus cool behavioural styles in electronic negotiations: the
impact of communication mode. Group Decis Negot 15(2):141–155
Pesendorfer EM, Graf A, Koeszegi ST (2007) Relationship in electronic negotiations: tracking behavior
over time. Zeitschrift für Betriebswirtschaft 77(12):1315–1338
Radicati Group (2016). Email statistics report, 2015–2019
Rice RE (1990) Computer-mediated communication system network data: theoretical concerns and empir-
ical examples. Int J Man-Mach Stud 32(6):627–647
Riordan MA, Trichtinger LA (2017) Overconfidence at the keyboard: confidence and accuracy in interpret-
ing affect in e-mail exchanges. Hum Commun Res 43:1–24

123
Are You Angry (Happy, Sad) or Aren’t You? Emotion… 413

Robinson RL, Navea R, Ickes W (2013) Predicting final course performance from students’ written self-
introductions: a LIWC analysis. J Lang Soc Psychol 32(4):469–479
Salovey P, Mayer JD (1990) Emotional intelligence. Imagin Cognit Personal 9(3):185–211
Schröder M (2003) Experimental study of affect bursts. Speech Commun 40(1):99–116
Shaver P, Schwartz J, Kirson D, O’connor C (1987) Emotion knowledge: further exploration of a prototype
approach. J Personal Soc Psychol 52(6):1061
Sinaceur M, Kopelman S, Vasiljevic D, Haag C (2015) Weep and get more: when and why sadness expression
is effective in negotiations. J Appl Psychol 100(6):1–25
Smith CA, Lazarus RS (1990) Emotion and adaptation. In: Pervin LA (ed) Handbook of personality: theory
and research. Guilford Press, New York, pp 609–637
Srnka KJ, Koeszegi ST (2007) From words to numbers: how to transform qualitative data into meaningful
quantitative results. Schmalenbach Bus Rev 59:29–57
Strapparava C, Mihalcea R (2008) Learning to identify emotions in text. In: Proceedings of the 2008 ACM
symposium on applied computing, ACM, New York, pp 1556–1560
Tausczik YR, Pennebaker JW (2010) The psychological meaning of words: LIWC and computerized text
analysis methods. J Lang Soc Psychol 29(1):24–54
Thompson L, Nadler J (2002) Negotiating via information technology: theory and application. J Soc Issues
58(1):109–124
Tov W, Ng KL, Lin H, Qiu L (2013) Detecting well-being via computerized content analysis of brief diary
entries. Psychol Assess 25(4):1069–1078
Trevino LK, Lengel RH, Daft RL (1987) Media symbolism, media richness, and media choice in organi-
zations a symbolic interactionist perspective. Commun Res 14(5):553–574
Van Kleef GA (2009) How emotions regulate social life: the emotions as social information (EASI) Model.
Curr DirPsychol Sci 18(3):184–188
Van Kleef GA, De Dreu CK, Manstead AS (2010) An interpersonal approach to emotion in social decision
making: the emotions as social information model. Adv Exp Soc Psychol 42:45–96
Walther JB (1992) Interpersonal effects in computer-mediated interaction a relational perspective. Commun
Res 19(1):52–90
Walther JB, D’Addario KP (2001) The impacts of emoticons on message interpretation in computer-
mediated communication. Soc Sci Computer Rev 19(3):324–347
Walther JB, Parks MR (2002) Cues filtered out, cues filtered in. Handb Interpers Commun 3:529–563

Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps
and institutional affiliations.

123
Group Decision & Negotiation is a copyright of Springer, 2019. All Rights Reserved.

You might also like