Conference PaperPDF Available

"Blissfully happy" or "ready to fight": Varying Interpretations of Emoji

Authors:

Abstract and Figures

Emoji are commonly used in modern text communication. However, as graphics with nuanced details, emoji may be open to interpretation. Emoji also render differently on different viewing platforms (e.g., Apple's iPhone vs. Google's Nexus phone), potentially leading to communication errors. We explore whether emoji renderings or differences across platforms give rise to diverse interpretations of emoji. Through an online survey, we solicit people's interpretations of a sample of the most popular emoji characters, each rendered for multiple platforms. Both in terms of sentiment and semantics, we analyze the variance in interpretation of the emoji, quantifying which emoji are most (and least) likely to be misinterpreted. In cases in which participants rated the same emoji rendering, they disagreed on whether the sentiment was positive, neutral, or negative 25% of the time. When considering renderings across platforms, these disagreements only increase. Overall, we find significant potential for miscommunication, both for individual emoji renderings and for different emoji renderings across platforms.
Content may be subject to copyright.
“Blissfully happy” or “ready to fight”: Varying Interpretations of Emoji
Hannah Miller, Jacob Thebault-Spieker, Shuo Chang, Isaac Johnson,
Loren Terveen, Brent Hecht
GroupLens Research, University of Minnesota
Minneapolis, MN 55455, USA
{hmiller, thebault, schang, ijohnson, terveen, bhecht}@cs.umn.edu
Abstract
Emoji are commonly used in modern text communication.
However, as graphics with nuanced details, emoji may be
open to interpretation. Emoji also render differently on dif-
ferent viewing platforms (e.g., Apple’s iPhone vs. Google’s
Nexus phone), potentially leading to communication errors.
We explore whether emoji renderings or differences across
platforms give rise to diverse interpretations of emoji.
Through an online survey, we solicit people’s interpreta-
tions of a sample of the most popular emoji characters, each
rendered for multiple platforms. Both in terms of sentiment
and semantics, we analyze the variance in interpretation of
the emoji, quantifying which emoji are most (and least) like-
ly to be misinterpreted. In cases in which participants rated
the same emoji rendering, they disagreed on whether the
sentiment was positive, neutral, or negative 25% of the time.
When considering renderings across platforms, these disa-
greements only increase. Overall, we find significant poten-
tial for miscommunication, both for individual emoji render-
ings and for different emoji renderings across platforms.
Introduction
Emoji are “picture characters” or pictographs that are pop-
ular in text-based communication. They are commonly
used in smartphone texting, social media sharing (e.g.,
nearly half of all text on Instagram contains emoji (Dimson
2015)), advertising (e.g., Chevy’s press release written
entirely in emoji #ChevyGoesEmoji1), and more. Oxford
Dictionaries declared the emoji or “face with tears of
joy” to be the 2015 “word of the year.” As this is the first
time that they had selected an emoji, they noted that “emoji
have come to embody a core aspect of living in a digital
world that is visually driven, emotionally expressive, and
obsessively immediate.”2
Copyright © 2016, Association for the Advancement of Artificial Intelli-
gence (www.aaai.org). All rights reserved.
1 http://www.chevrolet.com/crack-the-emoji-code.html
2 http://time.com/4114886/oxford-word-of-the-year-2015-emoji/
Most commonly-used emoji are encoded in the Unicode
standard for indexing characters3. There are currently 1,282
emoji in the Unicode standard, and for each of these, the
Unicode Consortium provides a code and name (e.g.,
U+1F600 for “grinning face”) but not the actual graphic.
This is the same as is the case for Unicode text characters:
for example, the Unicode character U+0041 indexes the
Latin capital letter ‘A’, but it does not indicate specifically
how the ‘A’ should look. Instead, a font renders the
Unicode characters a particular way: the appearance of this
text that you are reading is dictated by the Times New
Roman font.
Similarly, individual platform vendors such as Apple
and Google create their own rendering for each emoji
character they support. This means that the Grinning
Face” emoji character has a different appearance when
viewed on an Apple device (e.g., an iPhone) than on a
Google device (e.g., a Nexus phone). This is just one ex-
ample of two different platform renderings; there are many
platforms that each have their own unique set of emoji ren-
derings. Emojipediaa website serving as an “encyclope-
dia for emoji”lists 17 such platforms4, which means that
there may be at least 17 different renderings for a given
Unicode emoji character.
An emoji conveys its meaning through its graphic re-
semblance to a physical object (e.g., a smiling face), but it
is not well understood how people interpret the meaning of
emoji. Words have a dictionary definition, but emoji are
nuanced, visually-detailed graphics that may be more open
to interpretation. Furthermore, since emoji render different-
ly on different platforms, the emoji graphic that is sent by
one person on one device may be quite different than what
is seen by the recipient using a different device.
3 http://unicode.org/emoji/charts/full-emoji-list.html
4 http://emojipedia.org/
We contextualize our analysis in Herbert Clark’s psy-
cholinguistic theory of language use (Clark 1996). In social
psychology, a construal is the way that an individual inter-
prets communication. That is, when a speaker communi-
cates something, the addressee interprets or construes what
s/he believes the speaker to mean. When the addressee’s
interpretation differs from what the speaker intended, a
misconstrual occurs. In the context of emoji, a speaker is
sending emoji to an addressee through a mobile or desktop
platform. Likewise, the addressee is receiving the emoji via
a mobile or desktop platform. In this exchange, miscon-
strual can arise from differing interpretations derived from
either (1) the same rendering, if they each see the same
rendering or (2) different renderings, if they each see a
different rendering.
We explore the potential for misconstrual when using
emoji in communication by evaluating variation in emoji
interpretation. Using an online survey, we solicit people’s
interpretations of a sample of the most popular emoji
Unicode characters. In order to analyze how emoji inter-
pretations vary for renderings across platforms, the survey
included renderings of each emoji from five major mobile
platforms: Apple, Google, Microsoft, Samsung, and LG.
We identify the variance of interpretation in terms of sen-
timent (i.e., how positive is this emoji?) and semantics
(i.e., what does this emoji mean?).
We find that only 4.5% of emoji symbols we examined
have consistently low variance in their sentiment interpre-
tations. Conversely, in 25% of the cases where participants
rated the same rendering, they did not agree on whether the
sentiment was positive, neutral, or negative. When consid-
ering renderings across platforms, these disagreements
only increase. For U+1F601 (“grinning face with smiling
eyes” according to the Unicode Standard), participants
described the Google rendering as “blissfully happy”
while the exact same Unicode character, but rendered for
Apple , was described as “ready to fight.” We conclude
that emoji usage may be ripe for misconstrued communica-
tion and provide implications for design to manage the
likelihood of misinterpretation when using emoji.
Related Work
We begin this section with a discussion of the role of emot-
icons (e.g., :-) or smiley face’ – a precursor to emoji) in
the interpretation of text-based communication and how
emoji relate to emoticons. We then discuss what we know
about the consistency of interpretation for emoticons and
emoji.
Emoticons
Emoticons, or “typographic symbols that appear sideways
as resembling facial expressions,” (Walther and D’Addario
2001) such as :), have been in use in text-based communi-
cation since at least the early 1980s, with numerous studies
documenting their prevalence in SMS texts (Tossell et al.
2012), blogs (Huffaker and Calvert 2006), and, more re-
cently, Twitter (Park et al. 2013). Much research has fo-
cused on the role that emoticons can play in complement-
ing traditional text-based computer-mediated communica-
tion (CMC). Notably, Walther and D’Addario (2001)
found that while the emotional valence of text (e.g., “I am
happy”) tends to be more important than any accompany-
ing emoticons with respect to interpretation, a negative
emoticon (e.g., :( frowny face) can significantly change
the interpretation of the message. Lo (2008) provided addi-
tional evidence that emoticons affect interpretation, show-
ing that the same text can be perceived as either happy or
sad depending on which emoticon accompanies it. Derks,
Fischer, and Bos (2008) concluded in a survey of emotion
in CMC that emoticons largely function as non-verbal cues
do in face-to-face communication. Going beyond interpre-
tation of individual messages, Liebman and Gergle (2016)
demonstrated that emoticons (along with punctuation) are
important in interpersonal relationship development over
text-based communication. Together, this work emphasizes
the importance of emoticons in text-based communication.
The Rise of Emoji
Emoji were first created in the late 1990s in Japan but were
not officially added to the Unicode Standard until 2009
(Davis and Edberg 2015). They have become quite popular
since then, with, for example, over 2% of tweets (Novak et
al. 2015) and nearly half of text on Instagram (Dimson
2015) containing emoji. Emoji are often described as a
successor to emoticons (e.g., Novak et al. 2015), and
Pavalanathan and Eisenstein (2016) found that while emot-
icons are decreasing in popularity on Twitter, emoji are
increasing in popularity and seem to be replacing, not
complementing, emoticons.
While the large body of work on the role of emoticons in
text-based communication has largely not been replicated
for emoji, early work indicates that emoji do fulfill much
the same role. Kelly and Watts (2015) interviewed a cul-
turally diverse group of people and found that they did use
emoji in text-based communication to convey and modify
the meaning and emotional valence of their words.
Consistency of Emoticon and Emoji Interpretation
Whereas the display of emoji is platform-dependent, emot-
icons, as text, are displayed relatively consistently. Walther
and D’Addario (2001) found high agreement across their
participants (226 mostly male students) around sentiment
interpretations of the three emoticons that they studied, :-)
and :-( and ;-). In research on using emoticons in sentiment
analysis, Davidov, Tsur, and Rappoport (2010) found that
when Amazon Mechanical Turk participants were present-
ed with tweets in which emoticons had been removed, they
were able to identify with high precision the original emot-
icon that had been in the tweet.
Less is known about the consistency of emoji interpreta-
tion. Researchers such as Liu, Li, and Guo (2012) and No-
vak et al. (2015) have developed classifiers of emoji senti-
ment by labeling emoji with the sentiment of the surround-
ing text. While this has proven largely effective, both pa-
pers mentioned instances of emoji being associated with
different, and occasionally opposite, sentiment labels. We
know of no work, however, that has investigated how the
interpretation of emoji varies. We seek to address this gap
in the literature and also to understand how the platform-
dependence of emoji implementation might further com-
plicate interpretation.
Research Questions
As noted above, each platform has its own unique render-
ing of emoji Unicode characters (e.g., see Figure 1). Com-
munication can take place within platform or across plat-
form. If the sender and the receiver are both using the same
platform, then they are communicating within platform and
they see the same emoji rendering. If they are using differ-
ent platforms, then they are communicating across plat-
form and see different renderings of emoji. We break down
the goal of learning whether people interpret emoji the
same way or not into two research questions based on
within- and across-platform communication:
RQ1 (Within Platform): Do people look at the exact
same rendering of a given emoji and interpret it the
same way? For each platform, which emoji are
most/least likely to be misinterpreted in communica-
tion within platform?
RQ2 (Across Platform): Do people interpret one plat-
form’s rendering of an emoji character the same way
that they interpret a different platform’s rendering?
Which emoji are most/least likely to be misinterpreted
in communication across platforms?
We examine interpretation agreement and disagreement
along two dimensions: sentiment and semantics. Sentiment
analysis involves classifying the polarity of a given text.”5
For our purposes, this means determining whether the ex-
pression of a given emoji is positive, negative, or neutral.
In our context, semantics refers to what people think a giv-
en emoji means. For each of our research questions, we
explore how people’s interpretations manifest (a) senti-
ment and (b) semantic differences.
5 https://en.wikipedia.org/wiki/Sentiment_analysis
Survey
We created an online survey to solicit people’s interpreta-
tions of a sample of emoji Unicode characters, each ren-
dered for multiple platforms.
Emoji Unicode Character Sample
We selected a sample of Unicode characters from the most
popular emoji. To determine their popularity, we identified
emoji present in a dataset of approximately 100 million
random tweets collected between August and September
2015. This dataset provides a recent ranking of how often
each emoji is used.
We restricted our sampling to anthropomorphic emoji,
or those that represent faces or people, because (1) they are
very common and (2) we hypothesized that misconstrual
would be more likely among these emoji than those that
characterize “things (e.g., an airplane, a balloon, flowers,
flags, etc.). Anthropomorphic emoji account for approxi-
mately 50% of emoji use in our Twitter dataset, and
SwiftKey (2015) reports that faces or smileys comprise
59% of emoji characters typed with their smartphone key-
board app. We selected the top 25 most popular anthropo-
morphic emoji Unicode characters for our sample.
Platform Selection
To investigate how people interpret renderings from differ-
ent platforms, we solicited people’s interpretations of mul-
tiple platform renderings of each emoji Unicode character
in our sample, focusing on smartphone platforms. Using
comScore reports from 20156, we picked the top three
smartphone platforms: Android, Apple, and Microsoft.
Since Android is fragmented by manufacturer, we selected
Googles rendering, as well as the renderings of the top
two Android hardware manufacturers: Samsung and LG7.
We used renderings for these five platforms for every
Unicode character in our study. To collect the graphics of
the emoji to use in our survey, we used data from Emoji-
pedia8.
Survey Design
With 5 platform renderings of 25 emoji Unicode charac-
ters, we gathered survey results for 125 total emoji render-
ings. We employed a purely random between-subjects de-
sign, and each participant received a random sample of 15
emoji renderings to interpret from the 125 total. We aimed
6 https://www.comscore.com/Insights/Market-Rankings/comScore-
Reports-July-2015-US-Smartphone-Subscriber-Market-Share
7 Google provides the pure Android rendering, but many smartphone
manufacturers using the Android operating system (e.g., Samsung and
LG) override this rendering with their own rendering.
8 http://emojipedia.org/
to collect approximately 40 interpretations per emoji ren-
dering. Thus for a total of 5000 interpretations, and 15 in-
terpretations per participant, we recruited 334 participants
to complete the survey.
The survey began with a section to solicit background
information about the participants such as their age, their
gender, the smartphone platform that they use, and their
frequency of emoji usage. Next, each emoji rendering was
displayed on its own survey page, which showed an image
of the emoji and asked:
1. In 10 words or less, say what you think this emoji
means:
2. If you had to use one or two words to describe this
emoji, which would you use?
3. Judge the sentiment expressed by the emoji [on an or-
dinal scale from Strongly Negative (-5) to Strongly
Positive (5)]:
4. Fill in the blank: I would use this emoji [to / for /
when] _____________________
Questions one, two, and four elicited text responses and
were focused on semantic interpretations of emoji. Ques-
tion three elicited a numeric sentiment judgment, mirroring
the -5 to 5 sentiment scale used in Taboada et al. (2011).
In addition to the survey pages for the emoji in our sam-
ple, we created the same page for Apple’s heart emoji ( ,
Unicode U+2764). We had each participant complete this
survey page twice, once at the beginning of the survey, and
once at the end (after being shown their random sample of
15). This allowed us to control for quality of responses by
assessing intra-rater agreement on each participant’s two
ratings of the heart emoji. We also assessed the variance of
participants’ overall ratings of the heart emoji, and find
that our participants are very consistent in their sentiment
evaluation: they vary, on average, by 0.54 (out of 10) sen-
timent points.
Participants
We recruited our survey participants via Amazon Mechan-
ical Turk. We required participants to be located in the
United States in order to minimize interpretation differ-
ences that may arise from geographic and cultural influ-
ence, although this is an interesting direction of future
work. In pilot testing our survey, we estimated that it
would take roughly 30 to 35 seconds to complete each
emoji survey page. Prorating from a minimum wage of $8
per hour, this equated to about $0.07 per emoji page. With
17 emoji pages per survey (random sample of 15 plus the
heart emoji page shown twice), we compensated partici-
pants $1.20 for completing the survey.
Our participants had a record of high quality work on
Mechanical Turk: they each had at least 97% of their work
approved with at least 1,000 approved tasks completed.
Still, we calculated intra-rater reliability to ensure con-
sistency within each participant’s ratings. We computed
the difference between each participant’s pair of sentiment
ratings for the heart emoji character. Out of the 334 partic-
ipants, 308 (92%) of the participants differed by zero or
one rating. We considered these participants to be con-
sistent in their ratings and excluded the remaining 26 par-
ticipant responses from our dataset. To identify any low-
quality participant responses that were not reflected
through sentiment rating inconsistency, we also read par-
ticipant responses for the heart emoji questions and ex-
cluded four more participants for problematic responses
(e.g., the participant used the word “devil” to describe the
heart emoji). After these quality control checks, we re-
tained the data of 304 participants for our analysis.
Of the 304 participants, 134 were male, 169 female, and
1 other. The average age was 38.6 (SD = 12; min = 19;
max = 74). With regard to smartphone platform, 35% of
the participants use Apple, 8% use Google/Android, 29%
Samsung, 10% LG, 1% Microsoft, and the remaining 17%
use others. Participants also reported their emoji usage on a
scale from Never to Always: 3% said they never use
emoji, 16% rarely, 45% sometimes, 27% most of the time,
and 9% indicated “always”.
Data for Analysis
With 304 participants each completing 15 emoji interpreta-
tions, we had a total of 4,560 emoji interpretations and
ended up with approximately 37 interpretations per emoji
rendering (median = 37, min = 30, max = 41).
In the midst of our analysis, we discovered an error in
our emoji sample. We cross-checked back with Emojipe-
dia, the site from which we downloaded our emoji images,
and discovered that some of the images in our set (auto-
matically labelled by Unicode and platform at the time of
download) had been incorrectly labeled at the time of
download. We accordingly examined and reorganized our
survey data to ensure that we were associating participants’
interpretations with the correct emoji rendering. We ended
up with incomplete data for 3 of the 25 Unicode emoji
characters we sampled, so we excluded them from our
analysis (U+1F614 pensive face,” U+1F633 “flushed
face,” and U+1F604 “smiling face with open mouth and
smiling eyes”).
Analyses and Results
We conducted two separate analyses of the participants’
interpretations: one for sentiment judgments and one for
semantics, as indicated in the open-text questions. We next
detail our methods and results for each analysis.
Sentiment Analysis
In this section, we explore the role that sentiment may play
in emoji misconstrual. We describe our methods and rele-
vant results for each of our research questions.
Methods
For each emoji rendering, we have 30 to 41 sentiment
scores that are between -5 (most negative) and 5 (most
positive). In order to understand the degree to which indi-
vidual participants disagree on the sentiment of an emoji
rendering, we computed the pairwise differences (i.e., dis-
tances) of these sentiment scores. These values can range
from zero (perfect agreement) to 10 (perfect disagreement)
and describe the degree to which the participants disagree
on the sentiment of a given rendering.
To examine the variation in interpretation for specific
emoji renderings (RQ1), we calculated the average of these
distances to generate a within-platform sentiment miscon-
strual score for each emoji rendering. This reflects the av-
erage sentiment-based misconstrual between two people.
For instance, if a given symbol has a within-platform sen-
timent misconstrual score of 3, the sentiment ratings of this
symbol would differ by 3 points (e.g., 5 and 2), on average.
To examine variation in interpretation across platforms
(RQ2), we performed a similar calculation, but focused on
differences in rated sentiment across different platform
renderings of the same emoji Unicode character. For a giv-
en Unicode character (e.g., “face with tears of joy”), and a
pair of platforms (e.g., Apple and LG), we computed all
pairwise distances between the two sets of sentiment rat-
ings, and then took the average (e.g., an Apple-LG average
sentiment distance). We did this for all pairs of platforms,
and ended up with platform-pair average sentiment dis-
tances (e.g., one for Apple-LG, one for Apple-Microsoft,
one for LG-Microsoft, etc.). We then computed the grand-
mean (mean of these average sentiment distances), as the
across-platform sentiment misconstrual score.
Results
RQ1 (Within Platform) for Sentiment
To understand the extent to which interpretation of the
sentiment of each emoji rendering varies, we ranked each
rendering based on the within-platform sentiment miscon-
strual score in descending order for each platform. We pre-
sent the top three and bottom three of this ranking in Table
1. With an average sentiment distance of 4.40, Microsoft’s
rendering of “smiling face with open mouth and tightly
closed eyes” has the highest disagreement. For that emoji,
44% of participants labeled it as negative and 54% labeled
it as positive, indicating a clear lack of consensus. Because
Microsoft’s rendering has a within-platform sentiment
misconstrual score of 4.40, our participants differed by 4
sentiment points, on average. On the other end is the Apple
rendering of “sleeping face” with an average sentiment
distance of 0.45. For that emoji, 79% of participants con-
sidered it to be neutral (sentiment = 0) and all but one of
the other participants gave it a 1 or -1.
Overall, 44 of 110 renderings (40%) have a sentiment
misconstrual score larger than or equal to 2, meaning that
the average amount of sentiment disagreement between
two people for these emoji (even within a single platform)
is 2 or more. On the other hand, only five renderings
(4.5%) have a misconstrual score of 1 or less.
We also report the average sentiment misconstrual score
across all Unicode characters for each platform in Table 1.
Apple has the highest average within-platform sentiment
misconstrual (1.96); Google has the lowest (1.79).
Overall, we see that even when the emoji rendering se-
lected by the sender is exactly the same as what the recipi-
ent sees (because both sender and recipient are using the
same smartphone platform), there is still plenty of senti-
ment misconstrual. Indeed, if we select two participants
who have rated the exact same rendering, in 25% of those
cases, they did not agree on whether the sentiment was
positive, neutral, or negative. This reflects the most
straightforward form of within-platform communication,
and our results suggest that, even in this case, there are
clear opportunities for misconstrued communication.
RQ2 (Across Platform) for Sentiment
We now explore variance in sentiment for renderings
across platforms. In Figure 1, we show the distribution of
platform-pair sentiment misconstrual scores (i.e., average
sentiment distances of all possible sentiment rating pairs
between two platforms for a given character) for all
Unicode characters (each set of five renderings are shown
along the x-axis in Figure 1). We find that approximately
41% (9 of 22) of the Unicode characters have a range wid-
er than one sentiment unit, suggesting that at least one plat-
Most/Least Within-Platform
Sentiment Misconstrual
Apple
Microsoft
Samsung
Top 3
3.64
3.26
4.40
3.69
2.59
3.50
2.66
2.94
2.36
2.53
2.72
2.61
2.35
2.29
2.51
•••
•••
Bottom 3
1.25
1.13
1.12
1.23
1.30
0.65
1.06
1.08
1.09
1.26
0.45
0.62
0.66
1.08
0.63
Average
(SD)
1.96
(0.77)
1.79
(0.62)
1.90
(0.54)
1.84
(0.78)
1.84
(0.59)
Table 1. Top-3 and bottom-3 most different in terms of sentiment.
Higher values indicate greater response variation.
form’s rendering of these Unicode characters is different
from the other platforms. For instance, the large range for
grinning face with smiling eyes” (U+1F601) reflects the
very wide disagreement between the Apple platform and
the four others (platform-pair sentiment misconstrual
scores larger than 4.7), whereas the other platforms tend to
agree much more among themselves (platform-pair mis-
construal scores below 2). Similarly, for “sleeping face”
(U+1F634), the poor agreement arises from the fact that
while 91% of participants agreed that the Microsoft render-
ing was negative, there was a 68% chance that Samsung’s
rendering would be viewed as positive or neutral. It is also
worth noting here that we find person raising both hands
in celebration” (U+1F64C) in the top three most different
renderings for four of our five platforms, suggesting some
Unicode characters are simply more ambiguous than oth-
ers, leading to within- and across-platform differences.
The results from RQ1 and RQ2 regarding interpretation
of sentiment suggest that there are opportunities for mis-
construal for both within-platform and across-platform
renderings of emoji.
Semantic Analysis
Along with the perceived sentiment, differences in seman-
tic interpretations of emoji renderings could also contribute
to misconstrual.
Methods
We analyzed the free-text responses to Questions 1, 2, and
4 from our survey, which focused on the perceived mean-
ing and use cases for the emoji. Here, we use a very similar
technique to that presented above, adapted for text re-
sponses. For each participant’s answer for each rendering,
we aggregated their text responses to all three questions,
removed stop words and stemmed word tokens (using the
snowball stemmer implemented in the Scikit-Learn Python
library) and then converted the text to word vectors using a
standard bag-of-words model. For each rendering, we end-
ed up with 30 to 41 word vectors representing the respons-
es of different participants. We applied a TF-IDF transfor-
mation to all of the word vectors to reduce the importance
of common words that appear in all responses, e.g., “face,”
“something,and “etc.” We compute overall difference in
responses for a given emoji rendering as the average pair-
wise cosine distances of corresponding word vectors. This
is similar to our within-platform sentiment misconstrual
score above, so we will refer to this as our within-platform
semantic misconstrual score. These values range from zero
to one, increasing as participants use a greater variety of
words in their responses, and are insensitive to the number
of word vectors for each rendering.
To illustrate how the differences in word usage map to
the values of average text distance, we present samples of
aggregated responses in Table 2. The emoji rendering with
smallest within-platform semantic misconstrual (0.52) was
Apple’s rendering of “smiling face with heart-shaped
eyes.The responses for this rendering all focus heavily on
the concept of “love.” On the other hand, the emoji render-
ing with the largest within-platform semantic misconstrual
(0.97) was Apple’s rendering of “unamused face.” The
responses for this rendering show several different inter-
pretations “disappointment, “depressing, “unim-
pressed” and “suspicious.”
To answer our two research questions with regard to se-
mantic interpretation, we ran a similar analysis as the one
we did for sentiment. We first use the within-platform se-
Figure 1. Across-platform sentiment misconstrual scores grouped by Unicode. Each boxplot shows the range of sentiment misconstrual
scores a cross the five platforms. They are ordered by decreasing m edian platfo rm-pair sentiment misconstrual, from left to right.
mantic misconstrual score described above to answer RQ1.
We also computed across-platform semantic misconstrual
scores of each Unicode character, mirroring the computa-
tion for our sentiment analysis. For each Unicode character
(e.g., “face with tears of joy”) and each pair of platforms
(e.g., Apple and LG), we compute the pairwise word vec-
tor distances between the two sets of word vectors, and
then take the average (e.g., an Apple-LG average word
vector distance for the “face with tears of joy” emoji). We
then computed the grand-mean (mean of these platform-
pair average word-vector distances) to get the across-
platform semantic misconstrual score for each Unicode
character.
Results
RQ1 (Within Platform) for Semantics
Shown in Table 3, we observe significant variation in the
within-platform semantic misconstrual scores of all emoji
renderings. For all five platforms, the top three renderings
have a semantic misconstrual score (or average description
text distance) of nearly one, indicating significantly differ-
ent responses from the participants for a given rendering.
Though the emoji with the largest misconstrual scores vary
across platforms, the “smirking face” emoji (U+1F60F)
appears in the top three for all platforms except Google.
Only a few of the renderings (largely from Apple and Mi-
crosoft) were relatively similar, with average text distances
around 0.6. These results suggest that, as with sentiment,
many emoji evoke different interpretations from people.
RQ2 (Across Platform) for Semantics
Figure 2 shows the distribution of across-platform seman-
tic misconstrual scores for all platform pairs (e.g., Google
and Apple, Apple and Microsoft, etc.) for all emoji
Unicode characters. We conducted a Kruskal-Wallis test (a
non-parametric version of a one-way ANOVA, because the
word vectors are not normally distributed) to explore
whether the platform-specific word vectors differed from
one another, for each Unicode character. Indeed, we ob-
serve that there are statistically significant differences in
the platform interpretations of Unicode characters (Krus-
kal-Wallis test, p<0.001). For example, “person raising
both hands in celebration” (U+1F64C) is interpreted most
diversely across platforms: the top words used to describe
the Apple rendering are “hand, celebrate,” “stop, clap”
for the Google rendering , “praise, hand” for the LG
rendering , “exciting, highfor the Microsoft rendering
, and exciting, happy” for the Samsung rendering .
On the other hand, for smiling face with heart-shaped
eyes” (U+1F60D), people on all five platforms use words
like “love something/someone.
It is worth pointing out that the distributions of some
Unicode characters have much wider variances because
interpretation of a rendering for one platform largely dif-
fers from the interpretation of the renderings for the other
platforms. For example, all renderings of sleeping face”
(U+1F634) except the Microsoft rendering are clearly
interpreted as a sleeping face.” In comparison, renderings
Most/Least Within Platform
Semantic Misconstrual
Apple
Google
Microsoft
Samsung
LG
Top 3
0.97
0.97
0.96
0.96
0.96
0.96
0.95
0.95
0.95
0.96
0.95
0.94
0.95
0.95
0.93
•••
•••
Bottom 3
0.73
0.75
0.64
0.72
0.73
0.63
0.73
0.63
0.72
0.69
0.52
0.72
0.54
0.71
0.69
Average
(SD)
0.841
(0.111)
0.844
(0.078)
0.823
(0.115)
0.844
(0.080)
0.845
(0.087)
Table 3. Top-3 and bottom-3 most differently described render-
ings. Higher values indicate greater response variation.
Emoji
Avg. Text Distance
Randomly Selected Aggregated Responses for each Emoji
(Min) 0.52
a cool kind of love cool love for when I was feeling loving but also a little chill
I love you/this! love face I loved something someone else did or that I spotted.
that I love something love I wanted to show I loved an idea, photo or person
love something love something when i love something
(Max) 0.97
Dismay, disappointed Disappointed I am dismayed or disappointed
unimpressed unimpressed I saw, heard, or read something that I was indifferent towards
dissapointed dissapointed dissapointment
something depressing happened depression when something made me feel depressed
Table 2. Example participant responses about the semantic meaning of a given emoji rendering and their relationship to pairwise
word distance. The table includes emoji renderings with minimum and maximum average text distances in all emoji renderings.
of person raising both hands in celebration” (U+1F64C)
are confusing across all five platforms.
Results Summary
Stepping back slightly, we summarize insights from both
our sentiment and our semantic findings and triangulate the
degree to which both within-platform and across-platform
misconstrual may occur.
RQ1: We find that in many cases, when two people con-
sider the same emoji rendering, they may interpret both
the sentiment and semantic meaning differently. In other
words, there is opportunity for within-platform miscon-
strual. On our sentiment scale, only 4.5% of our render-
ings have an average misconstrual score below 1, and
40% have scores larger than 2, and our semantic analysis
finds very few renderings are described the same way.
RQ2: We find that for both sentiment and semantic in-
terpretations across platforms, there is disagreement. For
a given emoji Unicode character (five renderings, one for
each platform), there is clear opportunity for across-
platform misconstrual. 9 of the 22 (41%) Unicode char-
acters have sentiment distributions wider than one senti-
ment unit, and we see similar distributions of disagree-
ment when considering how people describe renderings
across platforms.
Thus, it is natural to ask: is the potential for misconstrual
greater within or across platform? We found that miscon-
strual was incrementally larger across-platform than with-
in-platform. More specifically, the average across-platform
sentiment and semantic misconstrual scores were 2.03 and
0.86, respectively (considering all across-platform pairs of
judgments). This is in contrast to the average within-
platform sentiment and semantic misconstrual scores,
which were 1.86 and 0.84, respectively (considering all
within-platform pairs of judgments).
Discussion and Implications
Emoji are very popular in text communication, but we have
shown that people do not interpret them in the same way.
Below, we tie our results back to Clark’s psycholinguistic
theory of communication, presenting additional qualitative
results in support of this discussion. Following that, we
highlight several implications for design.
Contextualizing Our Results in Theory
In the context of Clark’s psycholinguistic theory of com-
munication discussed above (Clark 1996), let us consider
the use of emoji in a hypothetical smartphone text conver-
sation: When Abby sends an emoji, she intends a particular
meaning. When Bill views the emoji, he interprets what he
thinks it means, or develops his own construal. If Bill’s
interpretation differs from Abby’s intended meaning, then
Bill misconstrued Abby’s communication. Our results sug-
gest that people often interpret emoji in diverse fashions,
potentially leading to situations like that of Abby and Bill.
With discrepancy between the sender’s and receiver’s in-
Figure 2. Across-platform semantic misconstrual scores grouped by Unicode. Each boxplot shows the range of semantic misconstrual
scores across the five platforms. They are ordered by decreasing median platform-pair semantic misconstrual, from left to right.
terpretations, the sender’s intended meaning is not com-
monly understood by both of them, so the communication
suffers. From our results, we see that this applies to emoji
usage in its most simple form: within-platform communi-
cation, where the sender and the receiver see the same
emoji rendering in their exchange.
Communicating across platforms, however, adds addi-
tional potential for misconstrual. Clark discusses in detail
the cognition behind how people internalize communicated
information. One main way is through joint personal expe-
riences, which fall into joint perceptual experiences
perception of natural signs of thingsand joint actions
interpretation of intentional signals. Emoji usage falls into
both: in addition to intending to communicate meaning,
they also require perceptual interpretation to derive mean-
ing. Clark posits that in order for a perceptual experience
to be commonly understood, people must attend toor be
perceivingthe same things and become confident that
they have done so in the right way. Unlike plain text where
people view the same characters in their exchange, plat-
forms effectively translate emoji: the emoji that the sender
chose is translated to the receiver’s platform’s rendering.
As a result, people do not attend to the same things when
communicating with emoji across platform. In fact, our
results show that people’s interpretations for a given emoji
character vary more across multiple platform renderings
than for a single platform’s rendering. This implies that
communication across platform is even more prone to mis-
construal than within-platform.
At the end of the survey, we asked our participants if
they had had any experiences with communication errors
around emoji. Many participants mentioned instances in
which emoji did not render on their phone (showing up as
black squares), which at least informs the recipient that
they are missing some meaning. However, some comments
were specifically about emoji being misinterpreted in an
exchange:
People have interpreted the emoji meaning some-
thing different than I intended and gotten upset.”
(P35)
Finally, some explicitly mention cases of miscommunica-
tion or confusion that arose from communicating across
platforms:
“When I use an emoji on an android and my iPhone
friend says that it was a sad face instead of a crying
excited face.” (P179)
“I downloaded the new iOS platform and I sent some
nice faces, and they came to my wife's phone as al-
iens.” (P22)
These cases provide further evidence that using emoji in
communication is prone to misinterpretation, although fur-
ther qualitative work would aid in understanding the
broader context of this phenomenon.
Implications for Design
Our results suggest that emoji users would benefit from
convergence of emoji design across platforms. The
Unicode Consortium succeeds at its goal of standardizing
emoji characters such that there is a character-level map-
ping between platforms. However, as we have shown, this
does not mean that interpretation is standardized across
platforms. Converging on emoji renderings across plat-
forms rather than diverging (e.g., to maintain distinctive
branding) may reduce the variation of interpretation and
thus lower the likelihood of miscommunication.
However, in addition to across-platform challenges, we
also observed that a great deal of the diversity in interpreta-
tions occurs within-platform, when people examine the
exact same emoji rendering. One hypothesis for the mech-
anisms behind these results is that there is a tradeoff when
it comes to “nuance” in emoji design, such as the color
shade of a cheek or the slant of an eyebrow. The graphic
nature of emoji affords nuanced expression, but this nu-
ance also potentially gives rise to a greater range of inter-
pretation. Exploring the relationship between detail and
misconstrual is an important direction of future work.
Besides the design of emoji themselves, there are con-
ceivably better ways to support emoji usage in communica-
tion. For example, when an emoji renders, smartphones
could indicate whether the particular rendering being
shown is the one the sender sent so the receiver can know
if she is viewing the intended rendering or not. If not,
smartphones could provide a way to look up the original
rendering to use for interpretation rather than a translated
rendering.
Future Work and Limitations
Though we studied 22 of the most popular anthropo-
morphic emoji, there are currently 1,282 total emoji
Unicode characters (including non-anthropomorphic ones).
Likewise, we studied 5 of the most popular mobile plat-
forms, but there are at least 17 platforms with their own
unique emoji renderings. We also only looked at one ver-
sion of each platform’s emoji even though people do not
consistently use the same version of operating systems. For
example, emoji in Android 4.4 look different from those in
Android 5.0, which look different from those in Android
6.1 (used in our study).
There are many different emoji renderings, and they all
may be subject to differing interpretation. It would be in-
feasible to survey all of them and new ones are constantly
emerging. Developing models to predict the sentiment and
consistency of a new (or unstudied) emoji is a line of re-
search that could prove fruitful for designers and support
applications that can provide feedback about the likelihood
of misconstrual for a given set of renderings.
One limitation of this work is that it considered emoji
out of context (i.e., not in the presence of a larger conver-
sation). While emoji are sometimes sent and received in-
dependently, they are often accompanied by surrounding
text (e.g., in a text message). Researchers have found that
emoticons can affect the interpretation of a message (Wal-
ther and D’Addario 2001; Lo 2008), but the parallel for
emoji has not yet been explored. Other researchers have
developed emoji sentiment classifiers based purely on the
sentiment of text they appear in (Liu, Li, and Guo 2012;
Novak et al. 2015), but this reflects interpretation solely of
context and not the emoji themselves. It is an important
direction of future work to explore people’s interpretations
of emoji with respect to the contexts in which they appear.
Another interesting avenue of future work lies in the
potential for cultural differences in interpretation of emoji.
Originating in Japan with global expansion, it is likely that
emoji usage and interpretation is culturally dependent. Ad-
ditionally, our approach to semantic analysis could be ex-
tended to use semantic relatedness measures, which would
address challenges associated with vocabulary mismatch.
Conclusion
Emoji are used alongside text in digital communication,
but their visual nature leaves them open to interpretation.
In addition, emoji render differently on different platforms,
so people may interpret one platform’s rendering different-
ly than they interpret another platform’s. Psycholinguistic
theory suggests that interpretation must be consistent be-
tween two people in order to avoid communication chal-
lenges. In this research, we explored whether emoji are
consistently interpreted as well as whether interpretation
remains consistent across renderings by different plat-
forms. For 5 different platform renderings of 22 emoji
Unicode characters, we find disagreement in terms of both
sentiment and semantics, and these disagreements only
increase when considering renderings across platforms.
Acknowledgments
We thank the Mechanical Turk workers who participated
in our survey. The National Science Foundation Graduate
Research Fellowship Program under Grant No. 00039202
and a University of Minnesota College of Science and En-
gineering Graduate Fellowship supported this work. We
would also like to thank our anonymous reviewers for their
comments and suggestions for follow-up research.
References
Clark, H.H. 1996. Using Language. Cambridge University Press.
Davidov, D., Tsur, O., and Rappoport, A. 2010. “Enhanced Sen-
timent Learning Using Twitter Hashtags and Smileys.” In ACL
'10, 24149. Association for Computational Linguistics.
Davis, M., and Edberg, P. 2015. “Unicode Emoji.” Technical
Report 51.
Derks, D., Fischer, A.H., and Bos, A. 2008. “The Role of Emo-
tion in Computer-Mediated Communication: A Review.” Com-
puters in Human Behavior 24 (3): 76685.
Dimson, T. 2015. “Emojineering Part 1: Machine Learning for
Emoji Trends.” Instagram Engineering Blog.
Huffaker, D.A., and Calvert, S.L. 2006. “Gender, Identity, and
Language Use in Teenage Blogs.” JCMC 10 (2): 0000.
Kelly, R., Watts, L. 2015. “Characterising the Inventive Appro-
priation of Emoji as Relationally Meaningful in Mediated Close
Personal Relationships.” Experiences of Technology Appropria-
tion: Unanticipated Users, Usage, Circumstances, and Design.
Liebman, N., and Gergle, D. 2016. “It’s (Not) Simply a Matter of
Time: The Relationship Between CMC Cues and Interpersonal
Affinity.” In CSCW '16.
Liu, K.L., Li, W.J., and Guo, M. 2012. “Emoticon Smoothed
Language Models for Twitter Sentiment Analysis.” In AAAI '12.
Lo, S.K. 2008. “The Nonverbal Communication Functions of
Emoticons in Computer-Mediated Communication.” CyberPsy-
chology & Behavior 11 (5): 59597.
Novak, P.K., Smailović, J., Sluban, B., and Mozetič, I. 2015.
“Sentiment of Emojis.” PloS One 10 (12).
Park, J., Barash, V., Fink, C., and Cha, M. 2013. “Emoticon
Style: Interpreting Differences in Emoticons Across Cultures.” In
ICWSM '13.
Pavalanathan, U., and Eisenstein, J. 2016. “Emoticons vs. Emojis
on Twitter: A Causal Inference Approach.” AAAI 2016 Spring
Symposium on Observational Studies through Social Media and
Other Human-Generated Content.
SwiftKey. 2015. “Most-Used Emoji Revealed: Americans Love
Skulls, Brazilians Love Cats, the French Love Hearts.” SwiftKey
Blog. April 21.
Taboada, M., Brooke, J., Tofiloski, M., Voll, K., and Stede, M.
2011. Lexicon-Based Methods for Sentiment Analysis. Computa-
tional Linguistics 37, 2: 267307.
Tossell, C.C., Kortum, P., Shepard, C., Barg-Walkow, L.H.,
Rahmati, A., and Zhong, L. 2012. “A Longitudinal Study of
Emoticon Use in Text Messaging from Smartphones.” Computers
in Human Behavior 28 (2): 65963.
Walther, J.B., and D’Addario, K.P. 2001. “The Impacts of Emoti-
cons on Message Interpretation in Computer-Mediated Commu-
nication.” SSCR 19 (3): 32447.
... A common limitation of these approaches, however, is that they do not rely on users' perceptions . This is not a trivial issue, since a user's interpretation does not always match the intended meaning of an emoji (i.e., the meaning intended by emoji developers, which can be found on the Unicode website) and there are likely to be differences between perceivers' interpretations (Jaeger et al., 2019;Miller et al., 2016;Rodrigues et al., 2018). Moreover, a given emoji can be used to represent several meanings . ...
... Moreover, a given emoji can be used to represent several meanings . Also, emojis with the same meaning can have different visual representations across different operating systems and social platforms, while affective properties may be different across distinct visual representations (e.g., Franco & Fugate, 2020;Miller et al., 2016;Tigwell & Flatla, 2016). ...
Article
Full-text available
This article presents subjective norms for 1031 emojis in six dimensions: visual complexity, familiarity, frequency of use, clarity, emotional valence, and emotional arousal. This is the largest normative study conducted so far that relies on subjective ratings. Unlike the few existing normative studies, which mainly comprise face emojis, here we present a wide range of emoji categories. We also examine the correlations between the dimensions assessed. Our results show that, in terms of their affective properties, emojis are analogous to other stimuli, such as words, showing the expected U-shaped relationship between valence and arousal. The relationship between affective properties and other dimensions (e.g., between valence and familiarity) is also similar to the relationship observed in words, in the sense that positively valenced emojis are more familiar than negative ones. These findings suggest that emojis are suitable stimuli for studying affective processing. Emoji-SP will be highly valuable for researchers of various fields interested in emojis, including computer science, communication, linguistics, and psychology. The full set of norms is available at: https://osf.io/dtfjv/.
... It is possible, then, that allowing for the use of emoji would increase the extent to which emotions are successfully communicated in digital contexts. On the other hand, however, emoji can be ambiguous (Miller et al., 2016) and could possibly undermine emotion communication. Hence, the role of emoji in implicit emotional communication is an avenue that is worth investigating further. ...
Article
Full-text available
In this research I explored the communication of emotions in digital contexts. Specifically, how well are people able to implicitly communicate discrete emotional states with words alone, and what are some of the correlates of this ability? In two experiments, senders created text messages designed to communicate 22 specific emotions (e.g., disgust), without naming the emotion, and receivers were asked to identify the emotion being conveyed. Senders and receivers indicated their degree of confidence that they successfully conveyed/recognized each emotion, and all participants completed measures of empathy and perspective taking. Emotion recognition (50%) far exceeded chance (5%) when a multiple-choice procedure was used (Experiment 2) and was substantial (20%) when participants were required to generate their own emotion labels (Experiment 1). When receivers failed to recognize the specific emotion, their errors were almost always of the same valence as the conveyed emotion (85% in Experiment 1 and 91% in Experiment 2), a rate that far exceeded chance (50%). Even though implicit emotional communication was relatively successful, the confidence rating of senders (but not receivers) was unrelated to communicative success. Emotion communication was more successful when the receiver was female and higher in empathy and perspective taking. In contrast, the gender and empathy level of senders was unrelated to communicative success. Overall, these results demonstrate that people can, to varying degrees, communicate emotions in digital contexts with words only.
... Crucially, WhatsApp has their own renderings of emoji so that they look identical on the phone of the sender and recipient, even if they are using different operating systems (cf. Miller et al. 2016). If an emoji is sent on its own without accompanying text, an enlarged version of the emoji is sent to the recipient. ...
Chapter
The aim of this chapter is to introduce and describe how digital technologies, in particular smartphones, can be used in research in two areas, namely (i) to conduct personality assessment and (ii) to assess and promote physical activity. This area of research is very timely, because it demonstrates how the ubiquitously available smartphone technology—next to its known advantages in day-to-day life—can provide insights into many variables, relevant for psycho-social research, beyond what is possible within the classic spectrum of self-report inventories and laboratory experiments. The present chapter gives a brief overview on first empirical studies and discusses both opportunities and challenges in this rapidly developing research area. Please note that the personality part of this chapter in the second edition has been slightly updated.
... In addition, the word cloud of emojis fails to capture the diversity of emoji appearances customized by emoji vendors. As different emoji renderings across viewing platforms may cause communication errors and diverse interpretations (Miller Hillberg et al., 2018;Miller et al., 2016), it is very important to support vendor-wise emoji visualization. Although several online platforms, such as Talkwalker 1 , Slido 2 , and Poll Everywhere 3 , offer emoji cloud services for various needs, they are not open-source and fail to provide APIs or functions for flexible usages in text mining. ...
Conference Paper
Full-text available
This paper proposes EmojiCloud, an open-source Python-based emoji cloud visualization tool, to generate a quick and straightforward understanding of emojis from the perspective of frequency and importance. EmojiCloud is flexible enough to support diverse drawing shapes, such as rectangles, ellipses, and image masked canvases. We also follow inclusive and personalized design principles to cover the unique emoji designs from seven emoji vendors (e.g., Twitter, Apple, and Windows) and allow users to customize plotted emojis and background colors. We hope EmojiCloud can benefit the whole emoji community due to its flexibility, inclusiveness, and customizability.
... The followers can see the story and interact with the person who has posted it. Users can add different colors to the videos they post, and the communication can be text-based too [19]. The feature right after 2016 became so popular that whatever happened, people would say, "Put it in your story". ...
Article
Full-text available
Storytelling content is where the facts are conveyed by emotion and that make people more engaged and want to take action or change their surroundings. Stories fascinate people and can easily be remembered compared to the facts alone. The much hyped feature "stories" of Instagram, a trendy social media platform, has become a game-changer for influencer marketing. The present study extends reactance theory in the context of Instagram's millennial users. Previous researchers have tested the effectiveness of the stories feature of this particular social media platform. Therefore, in line with the earlier studies, we propose a sequential mediation model that investigates the effect of storytelling content (made by Instagram Influencers) on audience engagement using two sequential mediation mechanisms of relatability and trust. Data were obtained using a cross-sectional study design from 273 millennial users of Instagram. Our results justify the direct and indirect hypothesized relationship through Process Macros. We found that relatability and trust play a significant role in building a strong relationship between storytelling content and audience engagement. Ultimately, the research findings suggest that professionals should be more creative while making the content on Instagram to engage the millennial market. Moreover, this research has tried to fill the gap in the literature on Instagram "stories" as an advertising platform.
... The accepted emoji creation date is sometime in the late 1990's and Japan has been the most often cited emoji birthplace (Miller, Thebault-Spieker, Chang, Johnson, Terveen, & Hecht, 2016;Magnus, 2016). Pictorial characters, very similar to present-day emoji, called smileys among other names, were already in use in web chartrooms, websites and Instant Messaging programs like ICQ going back to the late 1990's and the early 2000's (Voiskounsky, 1999;Weverka, & Taylor, 2000;Oakley & O'Modhrain, 2002;Olivine, 2006). ...
Thesis
Full-text available
Sequences of emoji characters are often used in online communication to convey meaning in conjunction with words, but also in their place. This thesis explores the use of emoji-only utterances in instant messaging, looking at them from a grammatical point of view, building on Jackendoff and Wittenberg’s (2014) hierarchy of grammatical complexity for the structure of sentence-level utterances. In this context, emoji, if used in isolation might posit word-like properties and show grammatical patterns and orders, similar to words. The present study investigates whether emoji showcase grammar on their own and identifies the grammatical patterns appearing in utterances where they substitute for words. According to its results, emojionly communication shows grammatical patterns on the lower levels of Jackendoff and Wittenberg’s hierarchy.
Article
The online digital sphere has recently witnessed a rapid growing of amateur audiovisual translation, which not only reinvigorates the current body of translation scholarship but also calls for a paradigm shift from the entrenched monolingualism. Though a plethora of studies have been conducted on fansubbing, little is known about blogger subtitling. Drawing on the netnographic data from a subtitling video blogger, @暴躁鸭仔 (baozao yazai, Grumpy Duckling for English) on a microblogging site (Weibo) in the mainland of China, we explored trans-translating by examining the translanguaging involved in the blogger’s subtitling translation. It is revealed that the blogger makes a hybrid, integrated use of multiple semiotic, orthographic, discursive and modal resources, without being confined by the boundaries amongst nameable languages, or linguistic/non-linguistic codes. The orchestration of the unified heterolingual repertoire builds up a social-media translanguaging space, wherein the tensions between translation and translanguaging due to distinct conceptual schemas are manipulated to yield synergy insofar as they dovetail into ingenious language making. We concluded that such creative translation can be retheorized as a process of trans-translating to encapsulate the subtitling experience in light of its dynamic and fluid nature, and transformative power of translanguaging. The results of this research might add to the picture of amateur subtitling and translation and that of translanguaging as well as their relation to one another.
Chapter
Emojis are quickly spreading in text-based communication. Accordingly, most students and teachers in an online learning environment integrate their online text communication with emojis to explain or support their perspective. However, because of the variation in students and teachers’ age, gender and culture, it was essential to ensure that students and teachers have the same understanding of used emojis. Consequently, we ran our study to ask students and teachers about their interpretations of the most common emojis. The results show moderate agreement between students and teachers. For example, was interpreted as ‘perfect’ by students and ‘punishment’ by teachers. This variation in interpretation must be resolved to ensure acceptable online communication between students and teachers.
Article
Personalization is prevalent in modern social media marketing strategies. However, the users of social media services may perceive that personalized content is intrusive. These two components may influence consumers’ purchase intention in social media personalization campaigns. Thus, this positivism empirical research proposes a theoretical framework named PI² using the personalization-intrusiveness-intention, as well as a success measurement based upon privacy calculus theory for effective social media personalization campaigns. Further, we investigated digital mnemonics’ potential influences with emotion as social information theory and different degrees of product involvement together with the model. Based upon the analysis of a total of 800 social media users, the results indicated that personalization positively affected but intrusiveness negatively affected the purchase intention. Furthermore, personalization significantly affected purchase intention through intrusiveness and the interaction of personalization with digital mnemonics negatively influenced consumers’ purchase intention on high involvement ads. In the post-hoc analysis, the moderating effect of users’ demographic variables was found to have some influence on the PI² framework. The findings of this research provide theoretical and managerial implications for researchers and field practitioners in social media personalization.
Conference Paper
Full-text available
Nonverbal, paralinguistic cues such as punctuation and emoticons are believed to be one of the mechanisms through which interpersonal relationship development takes place in text-based interactions. We use a novel experimental apparatus to manipulate these cues in a live Instant Message conversation. Results show a positive causal relationship of conversation duration and cue use on perceived affinity, and the relationship is contingent upon whether or not partners are able to see each other’s cues. Further analysis of the dialogue reveals that reciprocity may play a central role in supporting this effect. We then demonstrate how one’s cue use is influenced by a partner’s cue use, and show that cues are often used in greeting and sign-off rituals.
Article
Full-text available
There is a new generation of emoticons, called emojis, increasingly used in mobile communications and social media. In the last two years, over ten billion of emojis were used on Twitter. Emojis are Unicode graphic symbols, used as a shorthand to express concepts and ideas. In contrast to a small number of well-known emoticons which carry clear emotional contents, there are hundreds of emojis. But what is their emotional contents? We provide the first emoji sentiment lexicon, called Emoji Sentiment Ranking, and draw a sentiment map of the 751 most frequently used emojis. The sentiment of emojis is computed from the sentiment of tweets in which they occur. We have engaged 83 human annotators to label over 1.6 million tweets in 13 European languages by the sentiment polarity (negative, neutral, or positive). About 4\% of the annotated tweets contain emojis. The sentiment analysis of emojis yields several interesting conclusions. It turns out that most of the emojis are positive, especially the most popular ones. The sentiment distribution of the tweets with and without emojis is significantly different. The inter-annotator agreement on the tweets with emojis is higher. Emojis tend to occur at the end of the tweets, and their sentiment polarity increases with the distance. We observe no significant differences in emoji rankings between the 13 languages, and propose our Emoji Sentiment Ranking as a European language-independent resource for automated sentiment analysis. Finally, the paper provides a formalization of sentiment and novel visualization in the form of a sentiment bar.
Conference Paper
Full-text available
Emoticons are a key aspect of text-based communi-cation, and are the equivalent of nonverbal cues to the medium of online chat, forums, and social media like Twitter. As emoticons become more widespread in computer mediated communication, a vocabulary of different symbols with subtle emotional distinctions emerges especially across different cultures. In this pa-per, we investigate the semantic, cultural, and social as-pects of emoticon usage on Twitter and show that emoti-cons are not limited to conveying a specific emotion or used as jokes, but rather are socio-cultural norms, whose meaning can vary depending on the identity of the speaker. We also demonstrate how these norms propagate through the Twitter @-reply network. We confirm our results on a large-scale dataset of over one billion Tweets from different time periods and coun-tries.
Article
Full-text available
Emoticons are graphic representations of facial expressions that many e-mail users embed in their messages. These symbols are widely known and commonly recognized among computer-mediated communication (CMC) users, and they are described by most observers as substituting for the nonverbal cues that are missing from CMC in comparison to face-to-face communication. Their empirical impacts, however, are undocumented. An experiment sought to determine the effects of three common emoticons on message interpretations. Hypotheses drawn from literature on nonverbal communication reflect several plausible relationships between emoticons and verbal messages. The results indicate that emoticons’ contributions were outweighed by verbal content, but a negativity effect appeared such that any negative message aspect—verbal or graphic—shifts message interpretation in the direction of the negative element.
Article
Full-text available
It has been argued that the communication of emotions is more difficult in computer-mediated communication (CMC) than in face-to-face (F2F) communication. The aim of this paper is to review the empirical evidence in order to gain insight in whether emotions are communicated differently in these different modes of communication. We review two types of studies: (1) studies that explicitly examine discrete emotions and emotion expressions, and (2) studies that examine emotions more implicitly, namely as self-disclosure or emotional styles. Our conclusion is that there is no indication that CMC is a less emotional or less personally involving medium than F2F. On the contrary, emotional communication online and offline is surprisingly similar, and if differences are found they show more frequent and explicit emotion communication in CMC than in F2F.
Article
Online writing lacks the non-verbal cues present in face-to-face communication, which provide additional contextual information about the utterance, such as the speaker's intention or affective state. To fill this void, a number of orthographic features, such as emoticons, expressive lengthening, and non-standard punctuation, have become popular in social media services including Twitter and Instagram. Recently, emojis have been introduced to social media, and are increasingly popular. This raises the question of whether these predefined pictographic characters will come to replace earlier orthographic methods of paralinguistic communication. In this abstract, we attempt to shed light on this question, using a matching approach from causal inference to test whether the adoption of emojis causes individual users to employ fewer emoticons in their text on Twitter.
Article
Twitter sentiment analysis (TSA) has become a hot research topic in recent years. The goal of this task is to discover the attitude or opinion of the tweets, which is typically formulated as a machine learning based text classification problem. Some methods use manually labeled data to train fully supervised models, while others use some noisy labels, such as emoticons and hashtags, for model training. In general, we can only get a limited number of training data for the fully supervised models because it is very labor-intensive and time-consuming to manually label the tweets. As for the models with noisy labels, it is hard for them to achieve satisfactory performance due to the noise in the labels although it is easy to get a large amount of data for training. Hence, the best strategy is to utilize both manually labeled data and noisy labeled data for training. However, how to seamlessly integrate these two different kinds of data into the same learning framework is still a challenge. In this paper, we present a novel model, called emoticon smoothed language model (ESLAM), to handle this challenge. The basic idea is to train a language model based on the manually labeled data, and then use the noisy emoticon data for smoothing. Experiments on real data sets demonstrate that ESLAM can effectively integrate both kinds of data to outperform those methods using only one of them.
Conference Paper
Automated identification of diverse sentiment types can be beneficial for many NLP systems such as review summarization and public media analysis. In some of these systems there is an option of assigning a sentiment value to a single sentence or a very short text. In this paper we propose a supervised sentiment classification framework which is based on data from Twitter, a popular microblogging service. By utilizing 50 Twitter tags and 15 smileys as sentiment labels, this framework avoids the need for labor intensive manual annotation, allowing identification and classification of diverse sentiment types of short texts. We evaluate the contribution of different feature types for sentiment classification and show that our framework successfully identifies sentiment types of untagged sentences. The quality of the sentiment identification was also confirmed by human judges. We also explore dependencies and overlap between different sentiment types represented by smileys and Twitter hashtags.