Content uploaded by Zhenpeng Chen
All content in this area was uploaded by Zhenpeng Chen on Aug 04, 2017
Content may be subject to copyright.
Through a Gender Lens: An Empirical Study of Emoji Usage
over Large-Scale Android Users∗
Zhenpeng Chen, Xuan Lu, Sheng Shen, Wei Ai, Xuanzhe Liu, Qiaozhu Mei
Emojis have gained incredible popularity in recent years and become a new ubiquitous language
for Computer-Mediated Communication (CMC) by worldwide users. Various research eﬀorts have
been made to understand the behaviors of using emojis. Gender-speciﬁc study is always meaningful
for HCI community, however, so far we know very little about whether and how much males and
females vary in emoji usage. To bridge such a knowledge gap, this paper makes the ﬁrst eﬀort
to explore the emoji usage through a gender lens. Our analysis is based on the largest data set
to date, which covers 134,419 users from 183 countries, along with their over 401 million messages
collected in three months. We conduct a multi-dimensional statistical analysis from various aspects
of emoji usage, including the frequency, preferences, input patterns, public/private CMC-scenario
patterns, temporal patterns, and sentiment patterns. The results demonstrate that emoji usage can
signiﬁcantly vary between males and females. Accordingly, we propose some implications that can
raise useful insights to HCI community.
On April 11, 2015, Andy Murray, a world-wide known tennis player, announced his wedding on Twitter1.
Unlike any other formal announcement, such an inspiring tweet consists of no words, but 51 emojis
Undoubtedly, emojis have gained incredible popularity in recent years. Compared to traditional
information representations such as text messages, pictures, or even emoticons, emojis are considered
to be more lively, more expressive, and more semantically rich, and thus appreciated by Internet users,
particularly on smartphones. The prevalence of emojis has been an amazing phenomenon of social
innovation and appreciation. Interwoven into our daily communications, they have been a new ubiquitous
Prior research in the ﬁelds of Computer-Mediated Communication (CMC) and Human-Computer
Interaction (HCI) has taken an active interest in studying the similar instant messaging elements to
emoji, such as emoticons (e.g., ;-)). HCI research on emoji use, by contrast, is still in an early stage.
Given that users generate a large volume of emojis, some recent research interests have been made to
understand the behaviors of using emojis across apps , across platforms , and even across cultures .
Among these eﬀorts, one missing issue is the emoji-usage behaviors across genders.
Is the gender issue signiﬁcant and worthy? Probably yes. It is worth mentioning that Google recently
announced the support of some gender-oriented emojis2, implying the gender is a non-trivial issue in
emoji. Furthermore, identifying the gender diﬀerences is always an important topic in the HCI research
community. Existing studies have demonstrated that there could exist some diﬀerences in how males
and females use non-verbal cues in face-to-face oﬄine speech [4, 5, 6]. As a result, one may be curious if
there are diﬀerences in how males and females use in online CMC. In particular, with a conversational
arXiv:1705.05546v1 [cs.HC] 16 May 2017
user interface, an estimate of gender can be quite useful to infer the possible user proﬁles. For example,
web content/ad providers can make use of such gender diﬀerences in user’s behavior and decide to deliver
proper contents or recommend advertisements accordingly.
Indeed, due to the lack of labelled gender information in most state-of-the-art research, we currently
know quite little about the diﬀerence of emojis usage between male and female users. In this paper,
we made the ﬁrst descriptive analysis to bridge such a knowledge gap. Our work takes uniqueness at
two folds. First, we introduce the largest data set to date, from a leading input method app, namely
Kika. Such a data set covers 134,419 users from 183 countries, and their 401 million messages collected in
three months. Second, we conduct a multi-dimensional statistical analysis from various aspects of emoji
usage, including the frequency, preferences, input patterns, CMC-scenario sensitive patterns, temporal
patterns, and sentiment patterns. The analysis over the large-scale data set can at best make our results
comprehensive, and thus evidences the statistically signiﬁcant diﬀerences of emoji usage between female
and male users.
More speciﬁcally, we make the following ﬁndings:
•The frequency of emoji usage is quite diverse. Females are more likely to use emojis than males do.
•The emoji usage patterns seem to be diﬀerent . Female users are more likely to use only one emoji
or discretely use multiple emojis in a message, while male users tend to consecutively use multiple
•The preferred emojis are diﬀerent. For example, females are more likely to use face-related emojis,
while male users are more likely to use heart-related emojis.
•The emoji usage can be signiﬁcantly aﬀected by diﬀerent CMC scenarios for both males and females,
i.e., females are more likely to use emojis than males in public communication such as Twitter;
instead, in private communication, the situation is reversed.
•The sentiment implied by emojis can vary between males and females for a speciﬁc time, such as
weekdays, weekends, and festivals.
The rest of this paper is organized as follows. Section 2 summarizes status quo research eﬀorts. Sec-
tion 3 describes our data set and how the ethical issues are preserved. Section 4 presents our measurement
approach and the studied research questions. Section 5 makes the descriptive analysis at a macro level,
covering the frequency, choice, and input preferences of emoji, and evidences the gender diﬀerence exactly
exists in emoji usage. Section 6 and 7 extends the basic analysis from two contexts, i.e., CMC scenarios
and temporal-sentiment. Besides the preceding analysis, Section 8 discusses about the implications that
could help further potential applications, e.g., how our ﬁndings along with some knowledge their derived
can help improve the UI layout design and play as possible indicators of inferring the gender information
by using only emojis without any texts, and analyzes the limitations that could narrow our study and
results. Section 9 ends up the paper with concluding remarks.
2 Related Work
We start with summarizing the relevant background and literature. As the marked popularity of emoji,
many researchers have taken an active interest in investigating it. However, there is very limited research
on emoji usage from gender perspective. Inspired by the prior literature about emoticons, sentiment
analysis, and gender diﬀerences in non-verbal expressivity, we consider emoji as a new kind of non-verbal
cue and investigate gender impacts on emoji usage in mobile communication.
2.1 Emoticons and Emojis
In everyday verbal communications, we often use body language and facial expressions to better express
our complex emotions like humor, doubt, and sarcasm. These cues take up about 93% of everyday
communication . However, they cannot be used in text communication. With the growth of computer-
mediated communication, to supply the absence of these cues, people gradually turn to a kind of symbolic
representations where emotion or aﬀect is referenced pictorially using alphanumerics, punctuations, or
other characters, commonly called “emoticons” . Researchers have been trying to understand the senti-
ment and non-verbal cues provided by emoticons. Emoticons can be used to strengthen the expression 
or express emotions , humor , intimacy , and irony . There has also been some research
on the emoticon usage across countries , across cultures , and across statuses . What’s more,
some prior work focused on the gender gap in emoticon usage. Tossell et al. found that females send
more messages with emoticons while males use a more diverse range of emoticons . Hwang reported
female students are more likely to use emoticons to express emotion or intimacy and manage meaning
than male ones based on a Korean sample . Wolf found that females are expressive in emoticon usage
but the diﬀerence in frequency of emoticon usage is not statistically signiﬁcant on the mixed gender
However, the limited morphological variation of ASCII symbols limits the expressive power of emoti-
cons. The nonstandard creation and use of them introduce some challenge to data analysis. On the
contrast, emojis are preloaded and deﬁned standardly. They can not only be used to express emotions,
but also be able to represent various objects such as foods and sports. Since the debut on Twitter and
Instagram, emojis quickly won a lot of fans. Some ﬁndings even showed that the Twitter users who adopt
emojis tend to reduce their usage of emoticons . Besides linguistic functions such as replacing words
and describing contents, emojis also have non-verbal eﬀects such as decorating text, adjusting tones,
providing additional emotional or situational information, and engaging the recipient [20, 21, 22, 23, 24].
Similar to emoticons, there exists prior research on emoji usage across countries and cultures [25, 1, 26].
However, there are very limited studies on emoji usage from gender perspective. Some related work
that did not focus on gender perspective reported several ﬁndings about gender impacts on emojis. For
example, Nishimura examined data from a blog site in Japan and found that women tend to use more
emojis than men . Pohl et al. investigated the gender distribution of users tweeting with emoji and
found females are more than males . However, there is no existing work to study gender and emoji
through user input data at scale. To bridge this research gap, we investigate the gender eﬀects on emoji
usage from large-scale Android users comprehensively.
2.2 Sentiment analysis
Traditionally, sentiments and emotions are collected through survey-based , audio/video based [29, 30],
biometric-based [31, 32], and behavior-based approaches [33, 34, 35]. As the growth of text communica-
tion, researchers turned to identify users’ sentiment from the text they typed. Although the researchers
have proposed various advanced sentiment analysis techniques, it is still challenging to identify the senti-
ment and emotions from free text. Some researchers started to integrate non-verbal cues into sentiment
analysis of text. For example, Filho et al. proposed an approach which detected users’ facial expres-
sion reactions to conversations when text chatting . In addition, some researchers attempt to use
emoticons and emojis to model the text sentiment [37, 38].
In this paper, we need to know the sentiment behind emojis and take emojis as indictors to sense
users’ sentiment. Therefore, we investigate the existing approaches to measure the sentiment or semantic
representations of emojis. In the previous literature, there are three main approaches: (1) based on
the sentiment annotated by participants [39, 3, 40], (2) based on the their textual descriptions and
annotation tags deﬁned by the Unicode Standard [1, 23, 41], and (3) based on the context in which the
emojis occur [23, 42, 43, 44, 45]. We adopt the third approach in this paper. We apply a state-of-art
embedding model to get the semantic representations of emojis and then infer their sentiment.
2.3 Gender and Non-verbal Expressivity
Similar to emoticons, emojis are also proved to oﬀer non-verbal cues that poorly conveyed by text .
Therefore, to some extent, this work is about gender and non-verbal expressivity. Conventional wisdom
(a) Kika proﬁle page (b) Emoji supports (c) Multi languages support
Figure 1: The Snapshots of Kika Keyboard. Fig. 1(a) describes the proﬁle page on Google Play; Fig. 1(b)
illustrates the number of emojis supported in Kika; Fig. 1(c) illustrates the multiple languages supported
leads us to believe that females are more emotionally expressive than males . The intense expressivity
of emotion is also reﬂected in non-verbal communication. There are considerable prior studies reported
that females are more non-verbally expressivity than males [4, 5, 6]. Females are evidenced to show a
greater number of facial activity than males [47, 48] and observers can identify emotional states from
female more accurately than from male faces . Further, as the growth of the “facial expression”
(emoticons) in text, researchers started to investigate the relationship between gender and emoticon
usage and found female superiority in emoticon usage [18, 17, 16]. Now, in the times of emojis, there are
very limited studies on gender and the new kind of non-verbal cues as is mentioned above. In this paper,
we devote to enriching the ﬁndings of gender impacts on the rising non-verbal cues – emoji.
3 The Data Set
In this section, we brieﬂy describe our data set and how we process the data with strict ethical consid-
3.1 Data Collection and Description
Following our previous study in emoji usage , this paper still uses the data set that is collected via the
Kika Keyboard3(abbreviated as Kika in the rest of this paper), a leading Android input method app
in Google play (Fig. 1). As a free third-party keyboard supporting 82 languages and more than 3,000
emojis and emoticons, it has gained millions of downloads and installs across the world, and was ranked
as the top 25 most downloaded apps of Google play in 2015. During the period of data collection, Emoji
v4.04(containing 2,389 emojis) was the latest released version by Unicode Standard. We use it as the
list of candidate emojis to search our data set. Finally, we captured a total of 1,356 diﬀerent emojis from
To improve the user experience, Kika explicitly notiﬁes its users that some information will be col-
lected, such as the user proﬁle information (optional at user registration) and the user-input messages
4http://unicode.org/emoji/charts/full-emoji-list.html. Note that the emoji list has recently been updated. The latest
release version is Emoji v5.0 now.
(a) Top 10 countries with the most active users (b) Gender distribution
Figure 2: Demographic distribution of users. Fig. 2(a) illustrates the top 10 countries with the most
active users and their gender distributions in the data set; Fig. 2(b) shows the gender distribution in the
(deﬁned as the content user typed before “Send” action) for analysis. However, such information can
be collected from only the users who agree with the user-term statements. As is declared in Kika’s Pri-
vacy Policy5, no passwords or sensitive data are recorded and all data have been anonymized before the
As an input method which can run at a system-wide level, Kika can collect data that are not limited to
particular applications such as Twitter (compared to studies using Twitter data [25, 20, 43]). Moreover,
Kika can capture the contexts in other apps where it is enabled and launched. Hence, it enables us to
make more comprehensive analysis of emoji usage.
The data covers 134,419 active users who oﬀer their proﬁle information. These users come from 183
countries and regions and their 401 million messages from December 4, 2016 to February 28, 2017. All
active users posted at least one message through Kika keyboard during this period. We use an anonymized
user identiﬁer (device ID replaced with random string) to represent every single active user. For each
user, we use the demographic information (including gender and the country where he/she comes from)
and the messages typed by him/her in this period (with timestamps and the apps where the messaged
Then, we report the demographic distribution of users covered in our data set. As is illustrated in
Fig. 2(a), the top 10 countries ranked by their active users are Brazil, Indonesia, Mexico, US, Philippines,
Egypt, India, Thailand, Colombia, and Argentina. The country distribution is heavily long-tailed, so that
a small fraction of countries have a disproportionately large share of users. Users from these 10 countries
constitute about 70.9% of all users. However, in the aggregate, we have comparable male and female
users. As shown in Fig. 2(b), 47% of the users are male and the other 53% are female.
3.2 User Privacy and Ethical Consideration
Indeed, our study is based on the sensitive gender information of users, hence we take careful steps to
preserve the ethics in our research. First, our work is approved by the Research Ethical Committee of
the institutes (a.k.a, IRB) of the authors. In the entire life cycle of this study, we consider user privacy
as a critical concern. Kika replaced the device ID of each user with a randomized string before storage,
so one can not identify any individual user with information in the data set. All the data is stored on a
private, HIPPA-compliant cloud server, with strict access authorized by Kika. Although we calculated
the message length in Section 6 and labelled emojis with sentiment by means of analyzing the contexts
in Section 7.1, our analysis pipeline was entirely governed by Kika employees to ensure the compliance
extracted only the metadata for research.
Indeed, every empirical study could have some limitations in the used data set, which can potentially
aﬀect the analysis along with the derived results. Note that our data set only contains the users who
voluntarily share the gender information, and the user distribution is not uniform among all countries.
Hence, there could exist some selection bias. However, to the best of knowledge, our data set is the largest
to date and contains a representative number of users whose gender proﬁles are known. In particular,
compared to previous HCI research that studied gender diﬀerence based on limited or controlled user
groups [50, 51], e.g., by questionnaire from volunteers, the scale of our data set can make our analysis
more comprehensive and the results statistically meaningful.
Our empirical study aims to explore whether there exist some diﬀerences in emoji usage between male
and female users. As an emerging popular CMC fashion, emoji is considered to be quite semantically
rich, and can imply some sentiment even without any texts. In our study, we mainly make descriptive
analysis by employing various statistic measures such as OLS regression, two-tailed z-test, and Pairwise
We focus on the following analysis.
•Usage Pattern Analysis. We ﬁrst make a macro-level descriptive analysis of emoji-usage patterns
between male and female users from various aspects, including frequency pattern, consecutive/dis-
crete pattern, co-used pattern, and so on.
•CMC Scenario Analysis. Beyond the macro-level analysis of emoji usage, we further focus
on the CMC scenario analysis, i.e., whether the emoji usage can vary in diﬀerent communication
situations. Such an analysis is motivated by the fact that emoji can be used in both public apps
(e.g., Twitter) and private communication apps (e.g., WhatsApp), so it is interesting whether the
emoji usage can reﬂect gender diﬀerence under diﬀerent contexts.
•Temporal-Sentiment Analysis. Emoji is widely used to express sentiment beyond plain text
only. We ﬁnally explore the emoji usage along with the potential sentiment of male and female
users. More speciﬁcally, we focus on the diﬀerence from a temporal perspective, as time is an
important incentive factor that can lead to the changes of sentiment.
5 Usage Pattern Analysis
From the data set, we observe 30.5 million messages containing at least one emoji, accounting for 7.6%
of all messages. Beyond the overall popularity of emojis, however, we are curious if males and females
use emojis diﬀerently. How often do male and female users tend to use emojis in text messages? Which
emojis do male and female users tend to use, respectively? Are there any input patterns of using emojis?
In this section, we begin with some macro-level descriptive analysis of gender diﬀerence in emoji usage.
5.1 Frequency of Emoji in Messages
Previous studies have suggested that females are more non-verbally expressive than males [4, 5, 6]. More
speciﬁcally, gender diﬀerence has been observed in the frequency of emoticon usage [11, 16]. Since emojis
are another type of non-verbal cue, could we observe a similar diﬀerence in their usage frequency? We
begin with examining whether males or females use emojis more frequently in their messages. That is,
Table 1: Male and female frequency of emoji usage. (“**”: p-value <0.01)
Gender %emoji-msg(in total) %emoji-msg(2016.12) %emoji-msg(2017.01) %emoji-msg(2017.02)
Male 7.02 6.88 6.84 7.45
Female 7.96 7.78 7.77 8.41
z-score -347.58** -197.79** -215.37** -182.97**
Table 2: Emoji usage patterns. (“**”: p-value <0.01 after Bonferroni correction)
Male (%) Female (%) z-score
Pat1 39.50 40.61 -59.87**
Pat2 57.55 56.09 77.96**
Pat2.1 37.85 36.82 55.87**
Pat3 10.16 10.56 -35.08**
we compare the percentage of messages containing emojis (%emoji-msg) between male and female users,
both for the entire time range and in each month.
As shown in Table 1, in general, the 7.02% of messages sent by male users contain at least one emoji,
while the percentage for female user is 7.96%. The diﬀerence is signiﬁcant under two-tailed z-test and
holds in each month during the period. Therefore, we can conclude that females have higher proportion
of messages containing emojis in their input.
Finding (F1): In general, female users are more likely to use emojis in text messages than males.
This superiority is reﬂected from female higher proportion of messages containing emoji.
5.2 Consecutive/Discrete Usage Patterns
Above, we looked at the message-level diﬀerence, yet those emoji-messages aren’t quite the same. Some
may have only one emoji, while some other may contain multiple emojis. We would like to explore if
males and females have diﬀerent input patterns of using emojis. Below are the three typical patterns
how users input emojis:
•Pat1: Only one emoji in one message, e.g., I you.
•Pat2: Consecutive use of multiple emojis in one message, e.g., I love you .
•Pat3: Discrete use of multiple emojis in one message, e.g., I love you .
Note that, the Pat2 and Pat3 are not mutually exclusive. A user may use emojis in one message
both consecutively and discretely (e.g., I love you ). Furthermore, a consecutive use of the same
emoji is considered to be an emphasis and thus can imply much stronger sentiment . Similarly, people
sometimes lengthen words to emphasize their sentiment, such as “coooooooooooooolllll”  and females
are reported to have a higher frequency of use this skill than males in Twitter . Additionally, users
often lengthen the “mouth” of emoticon to indicate a strong aﬀect, such as :)) . Inspired by these
previous lessons, we also focus on male and female consecutive use of the same emoji. We deﬁne the
•Pat2.1: Consecutive use of the same emoji in one message.
For both male and female users, we compute the proportion of the four patterns in the emoji-messages.
Then we conduct the z-test to measure the gender diﬀerences in usage of these patterns. We apply
Bonferroni correction  to adjust p-values for multi-hypothesis testing. The results are summarized in
Figure 3: The top 10 most used emojis in male and female users
From Table 2, we see interesting diﬀerences between female and male users. Female users are more
likely to use only one emoji or discretely use multiple emojis in a message. In the contrast, male users
are more likely to consecutively use multiple emojis. Such a preference for consecutive emojis holds true
even if we focus solely on the consecutive use of the same emoji.
Finding (F2): The emoji usage patterns are diﬀerent between genders. Female users are more likely to
use only one emoji or discretely use multiple emojis in a message, while male users tend to consecutively
use multiple emojis. Furthermore, male users are more likely to consecutively use the same emoji to
reinforce their sentiment.
5.3 The Choice of Emojis
We have found gender diﬀerences in how emojis are used in messages. Yet little do we know about which
emojis are used. Do men and women have the same “goto” emojis? Do they co-use emojis in the same
way? In this part, we will examine if male and female users diﬀer in their choice of emojis. We not
only investigate frequently used emojis or emoji categories, but also explore which emojis are likely to be
co-used by male and female users.
5.3.1 Frequently Used Emojis
We start by examining the favorite emojis of male and female users. For each emoji, we calculate the
percentage of its occurrence in all emoji occurrence used by male or female users, and summarize the ten
most frequently used emojis in Fig. 3. Both emoji usage distribution of males and females are long-tailed.
A small set of emojis have a disproportionately large share of all occurrence. There are not signiﬁcant
diﬀerences in the top 10 frequently used emojis. The top 5 emojis are even the same for male and female
users, i.e., (face with tears of joy), (red heart), (smiling face with heart-eyes), (face blowing a
kiss), and (loudly crying face). However, we do notice diﬀerences for particular emojis. For example,
although is the most popular emoji in both males and females, it accounts for 18.9% in male emoji
usage, but 22.1% in females, with a diﬀerence of 3.2%. This diﬀerence is non-negligible given the heavily
Finding (F3): Male and female users share the same top 5 most used emojis. However, there is
obvious diﬀerence in usage proportion of some emojis.
Table 3: Face-related and heart-related emojis. (“**”: p-value <0.01)
Examples Male usage (%) Female usage (%) z-score
Face-related emojis 56.11 58.17 -201.60**
Heart-related emojis 19.41 17.62 222.01**
5.3.2 Frequently Used Categories
Above, we can ﬁnd that the most popular emojis are about expression, being either faces or hearts. In
our data set, we ﬁnd 69 face-related emojis and 15 heart-related emojis in total. The face-related emojis
emphasize the facial expressions through the eye, eyebrow, or mouth shape. Diﬀerent shapes are used
to express diﬀerent aﬀects (e.g., positive and negative) and meanings such as happy ( ), blue ( ), and
angry ( ). Instead of reﬂecting emotions in the facial expression, the heart-related emojis emphasize the
color and shape of heart (such as , , and ) to convey the love and aﬀects of users directly. These
84 emojis comprise 75.5% of total emoji usage for males and 75.8% for females. Indeed, these are the
two most popular categories of emojis.
In the traditional verbal communication studies for the real life [5, 47, 48, 56], females are reported to
show more facial-related activities than males. Previous studies also suggest that females are more likely
to express love in real life [57, 58, 59]. Since emojis act as non-verbal cues in text communication, could
we observe the similar behavior characteristics in their usage? In other words, are females more likely to
use face- and heart-related emojis in text communication, just like their habits in real life? To answer
these questions, we aggregate the usage of face and heart emojis, and see if males and females use them
We calculate the proportion of face- and heart-related emojis in total emoji usage for both male and
female users, and conduct a z-test to compare gender preferences for these emojis. Table 3 summarizes
the results. Female users are signiﬁcantly more likely to use face-related emojis. Such an observation
could be interpreted by previous studies on verbal communication. However, we are surprised to ﬁnd that
male users are more likely to use heart-related emojis than females in online communication. Such an
observation is contrary to psychological literature where males are reported to be less willing to express
love in real life. This ﬁnding implies that males reserve to express their love in the forms of verbal and
text in real life, but they turn to the ubiquitous language, emoji, to supplement their love expression for
Finding (F4): Females are more likely to use face-related emojis than males, while male users are
more likely to use heart-related emojis than females.
5.3.3 Co-used Emojis
We have already demonstrated gender diﬀerence in using the popular emojis. Yet we are also interested
in the diﬀerence at the long tail. Since the distributions of the long-tail emojis are small and trivial, it is
not straightforward to directly characterize such diﬀerence. Therefore, we group similar emojis together
by their co-occurrences and quantitatively characterize the diﬀerence between genders.
We leverage male and female data to cluster emojis that are frequently used together, respectively.
We use Point Mutual Information (PMI)  to measure the co-occurrence of every two emojis. The
PMI of two emojis e1and e2can be computed as
PMI(e1,e2) = log p(e1,e2)
where p(e1) represents the usage frequency of e1,p(e2) represents the usage frequency of e2, and
p(e1,e2) represents the frequency of the co-occurrences. A larger PMI indicates the two emojis are more
likely to occur together. We use the PMI of every two emojis to build a network for males and females,
respectively. We connect each emoji to ﬁve emojis that have the largest positive PMI with it. In such a
network, each node represents an emoji and the weight of each edge is PMI of the two nodes.
We then perform community detection using the classic Fast Unfolding algorithm  and split the
network into several communities. The nodes within one community have more connections (larger PMI)
with each other, while the nodes from diﬀerent communities have fewer connections (lower PMI). We
detect 56 communities for males and 55 communities for females. Then we analyze these communities
and ﬁnd some interesting phenomena. For example, males like to use sport-related emojis (such as
and ) with and , while female prefer to use the sport-related emojis with , and . What’s
more, females like to co-use the clothes-, shoes-, and bag-related emojis with (shopping bags), while
male users do not like this usage.
Finding (F5): Male and female users are diﬀerent in emoji co-occurrences. For example, males like
to use sport-related emojis with and , while female prefer to use the sport-related emojis with ,
and . Females like to co-use the clothes-, shoes-, and bag-related emojis with (shopping bags),
while male users do not like this usage.
5.4 Summary of Findings
From the ﬁndings F1-F5 derived from our descriptive analysis, we can have a basic understanding of
how emojis are used by females and males. Generally, female and male users can statistically have
quite signiﬁcantly diﬀerent behaviors in using emojis. Such ﬁndings evidence our initial hypotheses. We
further explore whether there could exist some complex context-aware usage that could be reﬂected by
gender diﬀerences. In the following sections, we choose typical contexts from two major aspects, i.e.,
communication scenario and temporal-sentiment.
6 CMC Scenario analysis
Based on the macro-level descriptive analysis, we have demonstrated how males and females use emoji
diﬀerently. However, we argue that knowing the general gender diﬀerence is deﬁnitely not enough, but
can be explored more deeply. Since users widely interweave emojis in daily life and work, it is more
interesting to investigate how users use emojis under diﬀerent contexts. In this section, we raise a
hypothesis that the usage of emojis can vary between genders under diﬀerent CMC scenarios. Such a
hypothesis is motivated from two folds. First, female and male users could have diﬀerent representation
preferences in CMC. Second, even though there have been some non-emoji-oriented studies exploring
gender diﬀerences in CMC, most of them are conducted at the public channel such as Twitter. It is
unclear whether the results can be still consistent for private channels such as WhatsApp. Since Kika
keyboard is a system-wide app, it makes us capable of conducting such a study.
For simplicity, we choose two popular representative apps from these two categories, i.e., Twitter
and Whatsapp, respectively. Accordingly, we only take US users into account in this section. For public
communication scenarios, we focus on the messages collected from Twitter, since Twitter is a popular
CMC channel for propagating information . For private communication scenarios, we select messages
collected from WhatsApp. Although Twitter can also be used to make private conversations now, What-
sApp is relatively more frequently used for private communication purpose. We investigate the emoji
usage of male and female users in these two scenarios, respectively.
We measure the frequency of emoji usage using the metric %emoji-msg deﬁned in Section 5. Table
4 displays the %emoji-msg and the z-test result of male and female users in public and private commu-
nication. In Twitter, it is observed that females are more likely to use emojis compared to males. Such
a ﬁnding is consistent with the previous work that has discovered that females have a higher frequency
of emoticon use than men in Twitter .
In contrast, it is interesting to ﬁnd that males tend to use emojis more frequently than females in
private communication scenario. In order to better understand this phenomenon, we further analyze the
Table 4: Frequency (%emoji-msg) in diﬀerent scenarios (“**”: p-value <0.01)
Male Female z-score
Twitter 11.18 12.21 -4.67**
WhatsApp 7.92 7.05 29.44**
Table 5: OLS Regression: Gender and number of emoji (“**”: p-value<0.01)
# of emoji in a message
(1) WhatsApp (2) Twitter
Gender -0.1287** 0.0573
Message length (excluding emojis) 0.1438** 0.0263**
Constant 2.3298** 1.8226**
Observations 255,490 11,455
number of emojis in each message typed in the two scenarios. We perform regressions to analyze the
gender diﬀerence and set Gender as a dummy variable (Male=0, Female=1) . Since one may expect
that more emojis are used in longer messages, we control the message length excluding emojis in the
regressions. Then we conduct the multiple-variable Ordinary Least Square regression  for Twitter
and WhatsApp, respectively. Regression results are summarized in Table 5. As expected, the longer the
message is, the more emojis it contains. The coeﬃcients of the dummy variable Gender are negative and
signiﬁcant in the regression on the WhatsApp data. Such a result indicates that males prefer using more
emojis than females, given the same length of a message in WhatsApp. Combining the ﬁnding shown
above, we can conclude that males not only use emojis more frequently, but also tend to use more emojis
in one message than females in private communication. However, in Twitter, gender does not have such
an impact on the number of emojis used in one message.
In addition, at the ﬁrst sight of Table 4, we may get the hasty conclusion that both males and
females are more likely to use emojis in public communication. In other words, emojis are more popular
in information propagation (i.e., Twitter) than in conversations (i.e., WhatsApp). However, previous
work has reported that emoticons are more popularly used in conversations than in information propa-
gation . Do emojis and emoticons diﬀer in this dimension? With this doubt, we consider the potential
interpretation behind this phenomenon.
As is shown above, message length could be an inﬂuential factor to emoji usage, so it is uncon-
scionable to compare the emoji usage in public and private communication directly. We can not rule out
the possibility that users tend to send longer messages in Twitter than in WhatsApp (median length of
messages in Twitter / WhatsApp is 5 words / 3 words, respectively), resulting in the higher likelihood
to use emojis in Twitter. Therefore, we measure the frequency by the number of emoji in one message.
We deﬁne a dummy variable Type (Private communication=0, Public communication=1) to distinguish
the two scenarios. For male and females users, we conduct OLS regressions, respectively. We report the
results of these regressions in Table 6. When typing messages of the same length, both male and female
users are more likely to use more emojis in private communication than in public one. From this per-
spective, indeed, emojis are more popularly used in private communication than in public communication.
Finding (F6): Females and males can perform quite diﬀerently in emoji usage under diﬀerent CMC
scenarios. Females are more likely to use emojis than males in public communication. However, in
private communication, males are more active in emoji usage, both in the frequency of using emoji
and the number of emojis contained in one message. Emojis are more popularly used in private
communication than in public communication.
Table 6: OLS Regression: Communication scenarios and number of emoji in one message (“**”: p-
# of emoji in a message
(1) Male (3) Female
Type -1.1749** -1.1964**
Message length (excluding emojis) 0.1832** 0.1187**
Constant 2.0370** 2.4369**
Observations 104,606 165,786
7 Temporal-Sentiment analysis
Finally, we aim to examine the emoji usage with temporal-sentiment context. Since emojis can imply
richer semantics and sentiment , analyzing the sentiment of emojis provides a more detailed summa-
rization of emoji usage. In addition, usually, it is argued that time and social events can be incentives
to aﬀect sentiment changes. As a result, we explore whether emojis can play as an new indicator of
sentiment to sense male and female response to some speciﬁc time or events, e.g., weekends and festivals.
7.1 Inferring Sentiment behind Emojis
To sense users’ response from the emoji usage, at ﬁrst, we need to infer the sentiment behind emojis.
As is summarized in Section 2, there are three main approaches to measure the sentiment or semantic
representations of emojis: (1) based on the sentiment annotated by participants, (2) based on the their
textual descriptions and annotation tags deﬁned by the Unicode Standard, and (3) based on the context
in which the emojis occur. As previous work has evidenced that everyone can interpret emojis in his/her
own way , we do not adopt the ﬁrst approach. What’s more, the oﬃcial descriptions are so simple that
sometimes they can not provide the sentiment of emojis. For example, the Unicode Consortium deﬁnes
the emoji as “red heart”. We employ a text analysis tool, named LIWC (Linguistic Inquiry and Word
Count)6, to validate the sentiment of the text “red heart” and ﬁnd that there is no sentiment conveyed
in it. The short description only describes the ideogram but covered up its emotional contents. However,
we often use the emoji to convey certain sentiment such as “I you”. Therefore, it seems to be
unreasonable to take the description as users’ interpretation. We ﬁnally adopt the third approach and
analyze the emoji sentiment from their contexts. Novak et al.  engaged human annotators to label
the tweets that contained emojis and further inferred the sentiment of emojis from the sentiment of the
tweets that they are used in. However, it is so challenging to annotate in the same way due to the large
scale of our data set. Instead, we choose to apply a embedding model to project words and emojis onto
the same high-dimensional vector space as in [23, 44, 45] and then use the vector space to generate the
semantic representations of emojis.
We choose LINE , a network embedding model, to compute the similar words of each emoji from
the context. It should be noted that we do not have the expertise in processing multiple languages other
than English. Prior work has reported that some emojis are interpreted in a diﬀerent way from language
to language . In order to rule out the impacts of culture and language on emoji interpretation, we
only focus on English messages collected from American users to analyze the emoji sentiment in this
section. We follow the approach in , setting LINE with the second-order proximity (LINE-2nd) to
get semantic representation of each emoji. First, we construct a co-occurrence graph from the corpus to
represent the semantic structure. In the co-occurrence graph, each node represents a token (it could be a
Table 7: Categorization of emojis with sentiment
Aﬀect Condition # of emojis
POS Sposemo>Snegemo 747
NEU Sposemo=Snegemo >0 150
NEG Sposemo<Snegemo 352
word, an emoticon, or an emoji), and the edge between the nodes represents the co-occurrence of the pair
of tokens. The second-order proximity between two nodes is the similarity between their neighborhood
network, describing how likely the two tokens occur in contexts. For example, we often use “I you” to
replace “I love you”. In this case, “love” is a similar word of because they have similar “neighborhood”
(contexts). We then compute the euclidean distance of two tokens in the embedding space to represent
their semantic similarity. For each emoji, we use a k-near-neighbors algorithm (kNN) to get the most
semantically similar words in embedding space, where k is set to 150, as its semantic representation. We
further employ LIWC to calculate the sentiment for the semantic representation of each emoji. LIWC has
language limitations and can not analyze text from all languages. It is another reason why we only focus
on English messages in this subsection. We only consider two measures of aﬀect derived from LIWC:
posemo (positive aﬀect) and negemo (negative aﬀect). After applying LIWC, semantic representations of
1,249 emojis are classiﬁed with sentiment. In other words, we obtain sentiment scores of the 1,249 emojis.
We further label each of them with one of the three sentiment polarity: positive (POS), neutral (NEU),
and negative (NEG). The result of categorization is summarized in Table 7. Most of these emojis are
positive, which is consistent with previous work [1, 43].
7.2 Emoji Response to Some Speciﬁc Time
Based on the sentiment of emojis, we then move to the temporal analysis of emoji usage along with the
potential sentiment of male and female users, respectively. As is mentioned above, we focus on US in
this section. We use the timestamps of messages as the time indicator (when the message containing
the emojis are posted). We have introduced in Section 3 that the timestamps of messages are collected
as the server-side time (GMT-8). We can not determine the real system time when the messages are
posted based on the country information because US spans four diﬀerent time zones (including GMT-8)
and we don’t have more detailed location information of each user. However, if we conduct daily-grained
analysis, this time bias can be negligible. For each day in the period that our data set spans (December
4, 2016 to February 28, 2017), we calculate the daily sentiment of male and female users reﬂected from
emoji usage, respectively. We derive day-to-day sentiment by counting positive and negative emojis. We
deﬁne the positive score sd,g on day dof gender gas the ratio of positive versus negative emojis, counting
from that day’s messages typed by g:
sd,g =countd(pos. emoji ∧g)
countd(neg. emoji ∧g)=p(pos. emoji |d,g)
p(neg. emoji |d,g)
We then leverage the results of the daily emoji usage to investigate the male and female users’ temporal
“emoji response” to some speciﬁc time or events, e.g., weekends and festivals.
7.2.1 At Weekdays and Weekends
We ﬁrst aggregate the daily results and look at the diﬀerences of emoji usage in weekdays and weekends.
As shown in Table 8, both male and females are more likely to use relatively more positive emojis in
weekends than in weekdays. It implies that people generally tend to use more positive emojis at non-work
time. What’s more, male users seem to be more sensitive to weekends in emoji usage as the growth rates
of positive scores are higher than females. This may be correlated with the diﬀerent kinds of stress that
males and females suﬀer from. In previous work, males are reported to mainly feel the stress from work
while females have more daily stress such as family and health-related events . At weekends, males
Table 8: Positive scores in weekdays and weekends
Weekdays Weekends Growth rate (%)
Male 15.85 18.38 15.96
Female 10.89 12.61 15.79
Table 9: Rank of Christmas-related emojis in male and female usage
Male Female Male Female
December 24, 2016 12 7 >20 11
December 25, 2016 9 5 >20 10
may be temporarily away from work pressure (the main pressure in their mind), but females still suﬀer
from their annoying family and health stress although they don’t need to go to work at weekends, either.
Finding (F7): Both males and females use emoji more positively in weekends compared to weekdays,
7.2.2 Near Festival
To investigate the “emoji response” of males and females to festivals, we select the Christmas. We
present the variation tendency of positive scores in December, 2016 in Fig. 4 to help us understand
the Christmas’ inﬂuence that can be reﬂected by emojis. In Fig. 4, we can ﬁnd the obvious sentiment
ﬂuctuation near Christmas for both males and females. From December 23 to 25, the sentiment of males
and females became more and more positive compared to other periods. Both the male and female users
were immersed in the festive joy. However, females seem to have more serious so-called “pre-holiday
blue”. “Pre-holiday blue” means that people get stressed and depressed about the holidays before they
arrive. We can see that males underwent a decrease of positive scores from December 21 to 23, while
the scores of females started to decrease much earlier, i.e., from December 18, a week before Christmas.
In addition, both males and females underwent quite obvious “post-holiday blue” reﬂected from their
decrease in positive scores after Christmas.
Another interesting ﬁnding is that people’s enthusiasm towards the Christmas is likely to be reﬂected
by those festival-related emojis. We calculate the most used 20 emojis of male and female users per
day in the three months. On December 24 and 25, we ﬁnd the obvious increase in the usage of and
, especially for female users. We list the rank of the two emojis (ranked by the usage frequency) in
December 24 and 25, for male and female users, respectively. As is illustrated in Table 9, both males
and females are sensitive to Christmas as they use the two emojis more frequently on December 24 and
25 (the two emojis don’t occur in the top 20 emoji list except the two days). Additionally, female users
seem to be more sensitive, reﬂected from the rank.
Finding (F8): Both males and females are relatively more positive and use Christmas-related emojis
( and ) more frequently on December 24 and 25, especially female users. Additionally, male and
females users have obvious pre- and post-holiday blue, especially females.
So far we have demonstrated the emoji-usage behaviors through a gender lens and indeed found some
gender-speciﬁc diﬀerences. We then discuss some implications following our previous ﬁndings, and try
to explore some possible opportunities for app developers. Meanwhile, we analyze the potential threats
that could aﬀect the results of this study.
Figure 4: Positive scores in December, 2016. Note: For simplicity, we use day to represent 2016-12-day
in the ﬁgures
8.1.1 Improving Emoji Keyboard Layout
As we have found some gender diﬀerences in emoji usage, the most intuitive and straightforward implica-
tion is to improve user experiences on current smartphone keyboard. In current OS-native and third-party
input methods, emojis are always shown in paging control and each page contains some emojis that are
displayed in a rather ﬁxed layout. When we want to type in an emoji, we need to swipe left or right to
search for it. This approach could be problematic as the number of emoji grows, which attracts many
researchers to optimize emoji entry speed such as . Indeed, our analysis can provide some hints for
smartphone input-method developers, not limited to Kika but also including other keyboard developers
or even OS vendors, to optimize their keyboard layout. For example, besides those top emojis such as
, we ﬁnd that females prefer face-related emojis while males prefer heart-related emojis. Therefore, the
ranking list of emojis shown on the keyboards’ layout should be more gender-aware to users. Additionally,
current keyboards can recommend the possible words or emojis that users may type in the next input.
Based on our observations, keyboard developers can improve their algorithms from gender perspective
instead of only listing the Most-Recently-Used emojis. In addition, we also ﬁnd some scenario-related
and temporal-related emoji usage patterns, which can further help keyboard developers improve their
emoji recommendations under some speciﬁc contexts.
8.1.2 Non Privacy-Invasive User Proﬁling
User proﬁling is crucial for Internet service or content providers. Knowing the possible user proﬁles, such
as gender, age, and other preferences, has been proved to be an eﬃcient way that not only improves
the user experiences, but also increases the accuracy of online ads recommendation to increase potential
ad clicks and revenues. In the times of “app economy”, app developers rely on in-app ads that require
accurate user proﬁling . Today’s in-app ads are mostly irrelevant, justly derided as taking a “spray
and pray” , which greatly aﬀects the revenue of app developers. In the social applications such as
facebook, a user is asked to ﬁll in his/her proﬁles, e.g., typically at registration phase, so that others
can better know his/her possible background. However, in most mobile applications, developers can
not collect such information. Indeed, collecting users’ typed texts in an input-method and using NLP
techniques to analyze the user proﬁle is a feasible approach. However, it may result in collecting some
sensitive information and thus hurt user privacy. Note that the analysis in this paper does not touch
any texts, but relies on the usage of emoji only. At least, it has been indicated that the inferred emoji
usage can help distinguish a users’ gender, and implies that the emoji usage could be a possible signal
for inferring user proﬁles in a non privacy-invasive fashion. In our future work, we plan to synthesize
the derived emoji-usage patterns and more information such as input traces and smartphone brands,
and devise machine learning models for better user proﬁling. We believe that such eﬀorts are exactly
actionable, as we can use the derived patterns as features to train the model and use the existing “gender-
labelled” users as ground truth for validation. In addition, since smartphone keyboard is a system-wide
application, it could expose some APIs to other applications that can predict their users proﬁles, which
may result in a new business model.
8.2 Threats to Validity
Every empirical study can have its own limitations and thus aﬀect the generalization of results. One major
potential threat of this study could come from the coverage of our data set. We focus on the active users
in the data set collected by the Kika keyboard. Indeed, most popular smartphone manufacturers support
emojis in their built-in input methods, and there are also some other popular third-party input methods
supporting emojis in the market. In our opinion, such a threat could be not so signiﬁcant to the current
macro-level analysis, as our data set covers a large number of real-world users from various countries
and thus can promise the gender-speciﬁc study statistically comprehensive. However, the temporal and
communication scenario analysis of emoji usage is conducted over only US users, hence the derived
patterns can not be fully generalized to users from other countries.
Indeed, besides gender, there are still some confounding user-proﬁle factors that may inﬂuence the
emoji usage such as country and language. It would be interesting to synthesize such information with
gender in our future work, and thus make the gender diﬀerence at a ﬁner granularity.
9 Conclusion and future Work
In this paper, we have presented an empirical study of emoji usage through a gender lens. Our study
was based on a unique and large data set collected by Kika – a popular input method app. The data
set covers 401 million messages typed by 134,419 active users from 183 countries and regions over three
months. 47% of these users are male and the other 53% are female. We conduct a multi-dimensional
statistical analysis from various aspects of emoji usage, including the frequency, preferences, input pat-
terns, CMC-scenario sensitive patterns, temporal patterns, and sentiment patterns. We applied rigorous
statistical tests to ensure the credibility of our ﬁndings. Indeed, we demonstrate that there are exactly
some gender diﬀerences in emoji usage. We drew on our observations and ﬁndings to present some impli-
cations such as distinguishing users’ gender through their emoji usage and then providing more eﬀective
recommendations. To the best of our knowledge, our study is the ﬁrst to qualitatively analyze emoji
usage through a gender lens based on large-scale users and their realistic usage data.
In the future, we would like to further bridge the gap between gender and emoji usage. We plan to
synthesize the derived emoji-usage patterns and more information such as input traces and smartphone
brands, and devise machine learning models for better user proﬁling.
 X. Lu, W. Ai, X. Liu, Q. Li, N. Wang, G. Huang, and Q. Mei, “Learning from the ubiquitous
language: An empirical analysis of emoji usage of smartphone users,” in Proceedings of the 2016
ACM International Joint Conference on Pervasive and Ubiquitous Computing, UbiComp 2016, 2016,
 C. Tauch and E. Kanjo, “The roles of emojis in mobile phone notiﬁcations,” in Proceedings of
the 2016 ACM International Joint Conference on Pervasive and Ubiquitous Computing, UbiComp
Adjunct 2016, 2016, pp. 1560–1565.
 H. Miller, J. Thebault-Spieker, S. Chang, I. L. Johnson, L. G. Terveen, and B. Hecht, ““blissfully
happy” or “ready to ﬁght”: Varying interpretations of emoji,” in Proceedings of the 10th International
Conference on Web and Social Media, ICWSM 2016, 2016, pp. 259–268.
 S. L. Ablon, D. P. Brown, E. J. Khantzian, and J. E. Mack, Explorations in aﬀect development and
meaning. Routledge, 2013.
 L. Marianne and B. Mahzarin, “Toward a reconsideration of the gender-emotion relationship,” Emo-
tion and Social Behavior, vol. 14, pp. 178–201, 1992.
 B. N. J. and J. A. Hall., “Beliefs about female and male nonverbal communication,” Sex Roles,
vol. 32, no. 1, pp. 79–90, 1995.
 M. A, Silent Messages (1st ed.). Belmont, CA:Wadsworth, 1971.
 J. Park, V. Barash, C. Fink, and M. Cha, “Emoticon style: Interpreting diﬀerences in emoticons
across cultures,” in Proceedings of the Seventh International Conference on Weblogs and Social
Media, ICWSM 2013, 2013.
 W. J. B and D. K. P, “The impacts of emoticons on message interpretation in computer-mediated
communication,” Social Science Computer Review, vol. 19, no. 3, pp. 324–347, 2001.
 L. Shao-Kang, “The nonverbal communication functions of emoticons in computer-mediated com-
munication,” CyberPsychology & Behavior, vol. 11, no. 5, pp. 595–597, 2008.
 E. Dresner and S. C. Herring, “Functions of the nonverbal in CMC: Emoticons and illocutionary
force,” Communication theory, vol. 20, no. 3, pp. 249–268, 2010.
 D. Daantje, B. A. ER, and V. G. Jasper, “Emoticons and online message interpretation,” Social
Science Computer Review, 2007.
 F. Ruth, urcan Alexandra, T. Dominic, H. Nicole, D. Harriet, and T. Amelia, “Sarcasm and emoti-
cons: Comprehension and emotional impact,” The Quarterly Journal of Experimental Psychology,
vol. 69, no. 11, pp. 2130–2146, 2016.
 M. K. M and O. Sae, “Pragmatic play? some possible functions of english emoticons and japanese
kaomoji in computer-mediated discourse,” in Association of Internet Researchers Annual Conference,
vol. 8, 2007.
 S. E. Tchokni, D. ´
O. S´eaghdha, and D. Quercia, “Emoticons and phrases: Status symbols in social
media,” in Proceedings of the Eighth International Conference on Weblogs and Social Media, ICWSM
 C. Tossell, P. T. Kortum, C. Shepard, L. H. Barg-Walkow, A. Rahmati, and L. Zhong, “A longitu-
dinal study of emoticon use in text messaging from smartphones,” Computers in Human Behavior,
vol. 28, no. 2, pp. 659–663, 2012.
 H. H. Sung, “Gender diﬀerences in emoticon use on mobile text messaging: evidence from a korean
sample,” International Journal of Journalism & Mass Communication, vol. 2014, 2014.
 A. Wolf, “Emotional expression online: Gender diﬀerences in emoticon use,” Cyberpsy., Behavior,
and Soc. Networking, vol. 3, no. 5, pp. 827–833, 2000.
 P. Umashanthi and E. Jacob, “Emoticons vs. emojis on twitter: A causal inference approach,” arXiv
preprint arXiv:1510.08480, 2015.
 K. Ryan and L. Watts, “Characterising the inventive appropriation of emoji as relationally mean-
ingful in mediated close personal relationships,” Experiences of Technology Appropriation: Unantic-
ipated Users, Usage, Circumstances, and Design, 2015.
 H. Cramer, P. de Juan, and J. R. Tetreault, “Sender-intended functions of emojis in US messaging,”
in Proceedings of the 18th International Conference on Human-Computer Interaction with Mobile
Devices and Services, MobileHCI 2016, 2016, pp. 504–509.
 H. Tianran, G. Han, S. Hao, N. T. vy Thi, and L. Jiebo, “Spice up your chat: The intentions and
sentiment eﬀects of using emoji,” in Proceedings of the 11th International Conference on Weblogs
and Social Media, ICWSM 2017, 2017, p. to appear.
 H. Pohl, C. Domin, and M. Rohs, “Beyond just text: semantic emoji similarity modeling to support
expressive communication ,” ACM Transactions on Computer-Human Interaction (TOCHI),
vol. 24, no. 1, pp. 6:1–6:42, 2017.
 R. Zhou, J. Hentschel, and N. Kumar, “Goodbye text, hello emoji: Mobile communication on wechat
in china,” in Proceedings of the 2017 CHI Conference on Human Factors in Computing Systems,
CHI 2017, 2017.
 F. Barbieri, G. Kruszewski, F. Ronzano, and H. Saggion, “How cosmopolitan are emojis?: Exploring
emojis usage and meaning over diﬀerent languages with distributional semantics,” in Proceedings of
the 2016 ACM Conference on Multimedia Conference, MM 2016, 2016, pp. 531–535.
 L. Nikola and F. Darja, “A global analysis of emoji usage,” ACL 2016, pp. 82–89, 2016.
 Y. Nishimura, “A sociolinguistic analysis of emoticon usage in japanese blogs: Variation by age,
gender, and topic,” Selected Papers of Internet Research, vol. 16, 2015.
 R. Wang, G. M. Harari, P. Hao, X. Zhou, and A. T. Campbell, “Smartgpa: How smartphones
can assess and predict academic performance of college students,” in Proceedings of the 2015 ACM
International Joint Conference on Pervasive and Ubiquitous Computing, UbiComp 2015, pp. 295–
 C. Orellana-Rodriguez, E. Diaz-Aviles, and W. Nejdl, “Mining emotions in short ﬁlms: User com-
ments or crowdsourcing?” in Proceedings of the 22nd International World Wide Web Conference,
WWW 2013, 2013, pp. 69–70.
 D. Sanchez-Cortes, J. Biel, S. Kumano, J. Yamato, K. Otsuka, and D. Gatica-Perez, “Inferring mood
in ubiquitous conversational video,” in Proceedings of the 12th International Conference on Mobile
and Ubiquitous Multimedia, MUM 2013, 2013, pp. 22:1–22:9.
 E. A. Carroll, M. Czerwinski, A. Roseway, A. Kapoor, P. Johns, K. Rowan, and M. M. C. Schrae-
fel, “Food and mood: Just-in-time support for emotional eating,” in 2013 Humaine Association
Conference on Aﬀective Computing and Intelligent Interaction, ACII 2013, 2013, pp. 252–257.
 J. M. dos Reis Costa, A. T. Adams, M. F. Jung, F. Guimbeti`ere, and T. Choudhury, “Emotioncheck:
Leveraging bodily signals and false feedback to regulate our emotions,” in Proceedings of the 2016
ACM International Joint Conference on Pervasive and Ubiquitous Computing, UbiComp 2016, 2016,
 R. LiKamWa, Y. Liu, N. D. Lane, and L. Zhong, “Moodscope: building a mood sensor from smart-
phone usage patterns,” in The 11th Annual International Conference on Mobile Systems, Applica-
tions, and Services, MobiSys 2013, 2013, pp. 389–402.
 J. Hernandez, M. E. Hoque, W. Drevo, and R. W. Picard, “Mood meter: Counting smiles in the
wild,” in The 2012 ACM Conference on Ubiquitous Computing, Ubicomp 2012, 2012, pp. 301–310.
 C. Coutrix and N. Mandran, “Identifying emotions expressed by mobile users through 2d surface
and 3d motion gestures,” in The 2012 ACM Conference on Ubiquitous Computing, Ubicomp 2012,
2012, pp. 311–320.
 J. F. Filho, T. Valle, and W. Prata, “Non-verbal communications in mobile text chat: emotion-
enhanced mobile chat,” in Proceedings of the 16th international conference on Human-computer
interaction with mobile devices & services, MobileHCI 2014, 2014, pp. 443–446.
 W. Wolny, “Emotion analysis of twitter data that use emoticons and emoji ideograms,” in Informa-
tion Systems Development: Complexity in Information Systems Development - Proceedings of the
25th International Conference on Information Systems Development, ISD 2016, 2016.
 J. Zhao, L. Dong, J. Wu, and K. Xu, “Moodlens: An emoticon-based sentiment analysis system for
chinese tweets,” in The 18th ACM SIGKDD International Conference on Knowledge Discovery and
Data Mining, KDD 2012, 2012, pp. 1528–1531.
 T. G. W and F. D. R, “Oh that’s what you meant!: Reducing emoji misunderstanding,” in Proceed-
ings of the 18th International Conference on Human-Computer Interaction with Mobile Devices and
Services Adjunct, MobileHCI Adjunct 2016. ACM, 2016, pp. 859–866.
 H. Miller, D. Kluver, J. Thebault-Spieker, L. Terveen, and B. Hecht, “Understanding emoji ambi-
guity in context: The role of text in emoji-related miscommunication,” in Proceedings of the 11th
International Conference on Web and Social Media, ICWSM 2017, 2017, p. to appear.
 E. Ben, R. Tim, A. Isabelle, B. Matko, and R. Sebastian, “emoji2vec: Learning emoji representations
from their description,” arXiv preprint arXiv:1609.08359, 2016.
 D. Thomas, “Emojineering part 1: Machine learning for emoji trends,” Instagram Engineering Blog,
 P. K. Novak, J. Smailovic, B. Sluban, and I. Mozetic, “Sentiment of emojis,” PloS One, vol. 10,
no. 12, 2015.
 F. Barbieri, F. Ronzano, and H. Saggion, “What does this emoji mean? A vector space skip-gram
model for twitter emojis,” in Proceedings of the 10th International Conference on Language Resources
and Evaluation LREC 2016, 2016.
 W. Ai, X. Lu, X. Liu, N. Wang, G. Huang, and Q. Mei, “Untangling emoji popularity through
semantic embeddings,” in Proceedings of the 11th International Conference on Weblogs and Social
Media, ICWSM 2017, 2017, p. to appear.
 K. A. M and G. A. H, “Sex diﬀerences in emotion: Expression, experience, and physiology,” Journal
of Personality and Social Psychology, vol. 74, no. 3, p. 686, 1998.
 B. Ross, B. R. M, G. Nancy, and S. Beth, “Unitization of spontaneous nonverbal behavior in the
study of emotion communication,” Journal of Personality and Social Psychology, vol. 39, no. 3, pp.
 B. Ross, B. Reuben, and B. Dana, “Temporal organization of spontaneous emotional expression: A
segmentation analysis,” Journal of Personality and Social Psychology, vol. 42, no. 3, pp. 506–517,
 B. Ross, M. R. E, and C. W. F, “Sex, personality, and physiological variables in the communication
of aﬀect via facial expression.” Journal of Personality and Social Psychology, vol. 30, no. 4, p. 587,
 X. Sun, Q. Zhang, S. Wiedenbeck, and T. Chintakovid, “Gender diﬀerences in trust perception when
using IM and video,” in Extended Abstracts Proceedings of the 2006 Conference on Human Factors
in Computing Systems, CHI 2006, 2006, pp. 1373–1378.
 P. Mudliar and N. Rangaswamy, “Oﬄine strangers, online friends: Bridging classroom gender seg-
regation with whatsapp,” in Proceedings of the 33rd Annual ACM Conference on Human Factors in
Computing Systems, CHI 2015, 2015, pp. 3799–3808.
 F. R. Aylmer, Statistical methods for research workers. Genesis Publishing Pvt Ltd, 1925.
 S. Brody and N. Diakopoulos, “Cooooooooooooooollllllllllllll!!!!!!!!!!!!!! using word lengthening to
detect sentiment in microblogs,” in Proceedings of the 2011 Conference on Empirical Methods in
Natural Language Processing, EMNLP 2011, 2011, pp. 562–570.
 P. R´ois´ın, “Gender and emotional expressiveness: An analysis of prosodic features in emotional
expression,” Pragmatics and Intercultural Communication, vol. 5, no. 1, pp. 46–54, 2012.
 D. C. W, “A multiple comparison procedure for comparing several treatments with a control,”
Journal of the American Statistical Association, vol. 50, no. 272, pp. 1096–1121, 1955.
 B. R. W, S. V. J, M. R. E, and C. W. F, “Communication of aﬀect through facial expression in
humans,” Journal of Personality and Social Psychology, vol. 23, no. 3, pp. 362–371, 1972.
 F. R. A and M. C. Lynn, “Gender and age stereotypes of emotionality,” Personality and Social
Psychology Bulletin, vol. 17, no. 5, pp. 532–540, 1991.
 B. J. O and P. C. W, “The inexpressive male: A tragedy of american society,” Family Coordinator,
pp. 363–368, 1971.
 W. Richard and G. Elisabeth, “Emotion expression and the locution i love you: A cross-cultural
study,” International Journal of Intercultural Relations, vol. 30, no. 1, pp. 51–75, 2006.
 C. K. Ward and H. Patrick, “Word association norms, mutual information, and lexicography,”
Computational Linguistics, vol. 16, no. 1, pp. 22–29, 1990.
 B. V. D, G. Jean-Loup, L. Renaud, and L. Etienne, “Fast unfolding of communities in large net-
works,” Journal of statistical mechanics: Theory and experiment, vol. 2008, no. 10, p. P10008, 2008.
 S. Ye and S. F. Wu, “Measuring message propagation and social inﬂuence on twitter.com,” in Social
Informatics - Second International Conference, SocInfo 2010, 2010, pp. 216–231.
 W. J. M, Introductory econometrics: A modern approach. Nelson Education, 2015.
 J. Tang, M. Qu, M. Wang, M. Zhang, J. Yan, and Q. Mei, “LINE: large-scale information network
embedding,” in Proceedings of the 24th International Conference on World Wide Web, WWW 2015,
2015, pp. 1067–1077.
 M. M. Pilar, “Gender diﬀerences in stress and coping styles,” Personality and Individual Diﬀerences,
vol. 37, no. 7, pp. 1401–1415, 2004.
 H. Pohl, D. Stanke, and M. Rohs, “Emojizoom: emoji entry via large overview maps,” in Proceed-
ings of the 18th International Conference on Human-Computer Interaction with Mobile Devices and
Services, MobileHCI 2016, 2016, pp. 510–517.
 J. Gui, S. McIlroy, M. Nagappan, and W. G. J. Halfond, “Truth in advertising: The hidden cost of
mobile ads for software developers,” in Proceedings of the 37th IEEE/ACM International Conference
on Software Engineering, ICSE 2015, 2015, pp. 100–110.
 S. Nath, F. X. Lin, L. Ravindranath, and J. Padhye, “Smartads: Bringing contextual ads to mobile
apps,” in Proceedings of the 11th Annual International Conference on Mobile Systems, Applications,
and Services, MobiSys 2013, 2013, pp. 111–124.