Conference PaperPDF Available

Construct of Sarcasm on Social Media Platform

Authors:

Abstract and Figures

The basic idea behind machine learning-based systems , or artificial intelligence in general, is mimicking how humans operate. This idea is particularly true for our problem, sarcasm detection on social networking sites (SNSs). Therefore, before proceeding to build a system that can detect sarcasm on SNSs, we attempt to understand how humans do the same. Many studies propose approaches based on personal experience and word-level definition of "sarcasm" [1], [2]. However, in this paper, we aim to find more general themes that are typical with users while detecting and expressing sarcasm on SNSs through a qualitative study to build a more effective sarcasm detection model.
Content may be subject to copyright.
Construct of Sarcasm on Social Media Platform
Dipto Das and Anthony J. Clark
Department of Computer Science
Missouri State University
Springfield, Missouri, USA
Email: dipto175@live.missouristate.edu
Abstract—The basic idea behind machine learning-based sys-
tems, or artificial intelligence in general, is mimicking how
humans operate. This idea is particularly true for our problem,
sarcasm detection on social networking sites (SNSs). Therefore,
before proceeding to build a system that can detect sarcasm
on SNSs, we attempt to understand how humans do the same.
Many studies propose approaches based on personal experience
and word-level definition of “sarcasm” [1], [2]. However, in this
paper, we aim to find more general themes that are typical with
users while detecting and expressing sarcasm on SNSs through
a qualitative study to build a more effective sarcasm detection
model.
I. INTRODUCTION
Sarcasm is an interesting aspect of human communication.
It is different and usually more complex to understand than
positive and negative statements. There are two opposing
meanings of a sarcastic statement: one literal meaning and one
intended meaning [3], whereas for non-sarcastic statements,
literal and intended meanings of a statement are the same.
Sarcasm detection has long been ignored from sentiment anal-
ysis perspective. Tepperman et al. [4] introduced the first work
in sarcasm detection in 2006 from a computer science view
point. After that, there has been some research in this field.
Most of these works are machine learning-based detection
approaches [3], [5], [6]. However, all of those works depend
on a specific definition of sarcasm or the authors’ hypotheses
based on personal experience on social media instead of
being grounded into psycholinguistic theory about sarcasm.
We argue that to make machine learning based models more
reliable and robust, we need to learn from cognitive theory of
sarcasm.
Sarcastic statements are usually associated with non-verbal
cues in in-person communication. According to Gibbs et
al. [7], to understand a sarcastic remark, one has to understand
both verbal and non-verbal cues at the same time. However,
in this age of social networking sites (SNS), a large portion
of conversations take place online and non-verbal cues of in-
person sarcasm need to be expressed in a different way due to
the limited communication modes available on SNS platforms.
Though sarcasm has been well-researched by psycholinguists,
most studies were conducted before the rise of social media,
and they do not consider how sarcasm on SNS platforms is
different from that of in-person communication. For exam-
ple, non-verbal cues like amplitude change of voice or air-
quotations are not possible on SNS platforms. In this paper,
we are interested to study how the limited number of available
modes of data on social media platforms imposes changes on
sarcasm.
Since the objective of our study is to understand the
construct of sarcasm on social media, we interviewed active
social media users who have regular exposure to sarcasm in
our targeted setting. We asked participants how they detect
sarcasm on SNS platforms; when and how they express
their sarcastic remark. To understand the role of language
in this regard, we carefully recruited participants from two
different language speaking people. This helped us understand
how people from all over the world can detect and express
sarcasm in a multilingual setting as well. Besides knowing
an individual’s method for detecting and expressing sarcasm,
we were also interested to know about general response to
sarcastic contents on these platforms.
Since there has been no prior theory about the construct of
sarcasm on SNSs, we used a grounded theory based approach
to analyze our data. We compared our data analysis with that
of the study by Gibbs et al. [7] to see how non-verbal cues
of in-person sarcasm change according to the modes available
on SNS platforms.
Our qualitative study results in two different models for
sarcasm on SNS platforms. The first one is on detection
and expression of sarcastic posts and the second one is on
use and non-use of sarcasm on SNSs. Our study critiques
traditional unimodal machine learning approaches [3], [5]. It
re-emphasizes the importance of multimodal approaches like
some research works [1], [8], [9] and provides pointers about
potential directions for multimodal sarcasm detection models.
The rest of the paper is organized as follows: section II gives
an overview of prior related works in literature; section III
discusses the methodology of our study; section IV and V
describes our two models respectively; section VI points out
the design implications of the study; and finally, we conclude
with a brief discussion.
II. LITERATURE REVIEW
Most works that study sarcasm are from linguistics, psy-
chology, and cognitive science. Gibbs et al. [7] conducted
experiments with 256 undergraduate students, where they
showed how non-literal interpretations of sarcastic statements
are processed by humans before the literal meaning. They
said that when a sarcastic statement is made in an in-person
conversation, and the audience have access to non-verbal
cues besides the verbal statements, the audience translate the
statements into the corresponding intended meaning, i.e., non-
literal meaning before translating the statements into their
surface/literal meaning. They also discussed how sarcasm
impacts how long the participants of a conversation remember
a particular statement. They highlight the ease of processing
and memory for sarcastic utterances. In a collection of several
empirical and theoretical works, Gibbs et al. [10] discuss the
theory of irony, especially comprehension of sarcasm in verbal
form, social contexts, and functions of irony.
Sarcasm detection as a field of computer science can be
placed under the field of sentiment analysis, which first
drew the attention of computer science researchers in 2006.
Tepperman et al. [4] developed the first work that recognized
the problem of sarcasm detection from the perspective of
computer science. They experimented with sarcasm recogni-
tion using cues like contextual (e.g., acknowledgement, agree-
ment/disagreement), prosodic, and spectral features (e.g. pitch,
energy, duration of each word). Given the limited capability
of natural language processing at that time, they proposed
a na¨
ıve approach of detecting sarcasm from text data. They
emphasized on the nature of sarcasm of being associated
with several commonly used phrases. In their work, they only
searched for the phrase “yeah right” as an indicator of sarcasm.
Several studies have invested effort to define what it means
to be “sarcasm”. Gonzalez-Ibanez et al. [3] identified the
opposite nature of literal and intended meaning of micro-blog
posts as sarcasm. According to them, sarcasm is different
from positive or negative statements made on social media.
It conveys negative sentiment while the literal meaning (also
termed as surface sentiment) of the statement is positive and
likewise, conveys positive intended sentiment with apparently
negative surface meaning. That means, the study by [3] argues
that sarcasm has one intended and one surface sentiment
that have opposite polarity, i.e., positive surface meaning
with negative intended meaning, and vice-versa. For example,
in a statement like: “Thank you for ruining my day., the
phrase “thank you” is used with criticizing intention (i.e.,
negative intended meaning), whereas the phrase itself literally
expresses gratitude (i.e., positive surface meaning). However,
several other studies do not agree with [3] in this regard.
Filatove et al. [11] argue that sarcasm always has positive
literal meaning with a negative intended meaning. They also
present observations of sarcasm having clear victims in micro-
blogging platforms, including social media, blogging sites, etc.
They discussed sarcasm and irony inter-changeably in their
work. Kreuz et al. [12] from a linguistic perspective agree with
the argument of Filatova et al. [11] on sarcasm having always
positive literal meaning with negative intended meaning.
Clift et al. [13] explained sarcasm as a phenomenon of diver-
gence between the spoken words and their intended meaning
with the Traditional Oppositional Model (TOM). However, this
model was criticized for ignoring the requirement of these
two aspects of meaning happening at the same time. Sperber
et al. [14] suggested that audiences just process the intended
meaning of sarcasm in a model named “Echoic/Interpretation
Model”. Later building on this model, the “Echoic Reminder
Model” was proposed and reemphasized by Kreuz et al. [12]
and Colston et al. [15] discussed the role of generally expected
situation or social norms. Instead Kumon-Nakamura et al. [16]
suggested sarcasm is achieved by mentioning part of an
expected situation that has occured while some other part was
violated. Later Colston et al. [17] in their book, discussed
how verbal sarcasm can be viewed as violation of expecta-
tion, and the pragmatically insincere or contrary relationship
between literal and intended meaning of statements. This is
echoed in the studies by [3], [11], [12] where we can see
sarcasm as violations of Grice’s maxims [18]. According to
Grices Maxims [18], there are two of the major principles for
cooperative dialogue: the maxim of quality and the maxim
of manner. The maxim of quality states that one tries to be
truthful and does not give information that is false or that is not
supported by evidence. The maxim of manner says that one
tries to be clear as one can in what one says avoiding obscurity
and ambiguity. According to Tepperman et. al. [4], sarcastic
speech always violates at least of one of Grice’s maxims for
cooperative dialogue.
Bamman et al. [6] gave importance to context information
for the task of sarcasm detection. They tried to capture extra-
linguistic information from the context of an utterance of
sarcasm on Twitter. According to them, inclusion of properties
of author, audience, and the immediate communicative envi-
ronment can contribute to the sarcasm detection task. Their
argument also situates itself in a line with linguistic study by
Utsumi et al. [19] who discuss the comprehension of verbal
irony for in-person conversational settings. The role of context
can also be explained with the expectation of certain social
norms as in [12], [15], [17], and thus reestablishes the incident
of violating Grice’s maxims [18].
III. METHODOLOGY
Our qualitative study started with goals to (1) understand
how users recognize sarcastic contents on social media, with
or without context, (2) study what factors impact the ways
of how they express sarcasm, and (3) study how users on
social media respond to sarcasm. To achieve these goals, we
conducted an interview-based qualitative study with social
media users situated in Missouri, United States and Dhaka,
Bangladesh. Our data collection consisted of semi-structured
interviews with 20 participants from these two countries.
A. Semi-Structured Interviews
We conducted semi-structured interviews with participants
between November and December 2018. The interviews tar-
geted understanding participants’ social media using practices
and their ways of recognizing as well as conveying sarcasm.
The first author (23 years old, Male) in this work was born and
brought up in Bangladesh, and has been living in the United
States for more than one year. He speaks both local languages,
Bengali and English. Since use of sarcasm is very common on
social media, we began by recruiting participants who were
active on social media. We adopted a blend of convenience
sampling, purposive sampling, and snowball sampling [20].
First, two participants were recruited from the authors’ social
network by convenience sampling. Second, since the focus
area of this research is the social media platform, authors
posted the recruitment flyer of this research on social media. In
the flyer, we described the inclusion criteria for our study and
gave a high level overview of the objective of the study. We
distributed the flyer through departmental email. Second, we
used social media itself as a channel for recruiting participants
since most of the users on this platform will inherently satisfy
one of the inclusion criteria. We shared the recruitment flyer
on the social media. As a result, the subjects of interests in this
research could be easily reached through purposive sampling.
Third, as previous literature suggest, by keeping the comment
section public for tagging improves the response rate [21],
we welcomed tagging other potential participants. Again, our
participants recruited through convenience sampling in the first
phase helped us recruit additional participants. Thus, snowball
sampling in both online and in-person social network helped
us to recruit potential subjects. We also utilized in-person
communication and recruited participants through word-of-
mouth. In total, we recruited 20 participants speaking two
different languages from two different countries.
Participation in the study was voluntary. The average com-
pletion time of the interviews was around 25 minutes. The in-
terviews were conducted one-on-one. We gave the participants
a high level overview of the study objective at the beginning
of the interview. We encouraged them to ask any question
they might have, and we obtained written consents from
participants before the interviews with the informed consent
form. The consent form was devised keeping it at a high
school standard reading level. However, we also summarized
the consent form in their native language. Interviews were
conducted at a place preferred by each participant, or over
Skype, and in his/her native language. Interviews were audio-
recorded with permission from the participants.
Interviews were semi-structured and guided by a list of
topics. We collected participant demographic information like
their age, gender, most recent occupation, highest attended
educational level, etc. We asked about their experience about
using social media, e.g., with whom they mostly interact with,
what kind of contents they usually see in their newsfeed. We
then asked questions that sought an understanding of how they
recognize and express sarcasm, including their views about
overall user response to sarcastic contents on social media.
B. Participants Characteristics
Our 20 participants (16 males and 4 females) came from
two different language speaking communities originated from
two different countries and ranged in age from 19 to 34 years
(average = 25.1 years, standard deviation = 4.48) . With respect
to their social media usage, all of our participants satisfy these
following criteria:
Must have an account with at least one SNS for more
than a year.
Must be an active user on SNS with spending 5-7 hours
per week.
Participants possessed a range of socio-economic back-
grounds. Five of them are undergraduate students, six are
graduate students, six are employed having undergraduate or
graduate degrees, and three are currently unemployed. More
TABLE I
DEM OGR AP HIC S OF PART ICI PANT S (N=20) IN T HE IN TE RVIE W ON
SA RCA SM U SE ON SNS
ID Gender Age Language
P1 Male 33 English
P2 Male 29 Bengali
P3 Male 21 English
P4 Male 28 English
P5 Female 22 English
P6 Male 22 English
P7 Female 29 English
P8 Male 20 Bengali
P9 Male 31 Bengali
P10 Male 34 Bengali
P11 Male 30 English
P12 Female 22 Bengali
P13 Male 20 English
P14 Male 25 Bengali
P15 Female 24 Bengali
P16 Male 25 Bengali
P17 Male 21 Bengali
P18 Male 25 Bengali
P19 Male 19 English
P20 Male 22 English
detailed information about our recruited participants are shown
in Table I.
The participants we studied represent two different sets of
social media users. The participants recruited from the United
States were mostly users of both Twitter and on Facebook. On
the other hand, participants collected from Bangladesh were
mostly active on Facebook, some of them having accounts
on Twitters that they do not use often. Participants from the
United States use English in all their social media activities
whereas participants from Bangladesh varied in their language
use on social media. They used both Bengali and English on
social media, as well as a version of Bengali called “Banglish”,
Bengali words using English alphabet.
C. Data Collection and Analysis
The data we collected resulted in a total of 283 minutes
(4 hours 43 minutes) of audio-recorded interview data and a
collection of field notes. The first author of this paper tran-
scribed the interviews and translated them to English. These
qualitative data were analyzed using an inductive approach.
We utilized grounded theory [22] as the inductive method on
the interview scripts. Since to the best of our knowledge, there
has been no research on theory about users’ sarcasm behavior
on online platforms, we in the early phase of our study,
aimed to have insights/theories about users’ sarcasm behavior
on social media. Therefore, grounded theory data analysis
meets our need. As core phenomenon, we are interested to
study how users detect sarcastic remarks on social media. We
studied what factors initiate the circumstances of a sarcastic
conversation to occur or a sarcastic remark to appear as a part
of a conversation as the causal condition. This leads to our
studies of strategies, i.e., how users express sarcasm on social
media. Then we study what consequences or impacts sarcasm
has on users’ interaction on social media.
After we conducted interviews, we prepared transcriptions
of the sessions. We identified parts of the participants’ quotes
where they discussed their methods of expressing sarcasm. A
participant mentioned his use of interjections inappropriately
to convey sarcasm. We open-coded this response descriptively
as “wrong use of interjection”. Repeated patterns in users’
interaction give rise to axial codes. For example, “wrong use
of interjection” and “association of wrong adjectives” are two
open codes categorized under “opposing sentiments as parts
of a single sentence”. The final codes were agreed upon when
themes came to a saturation. In selective coding phase, we
integrated the emerged axial codes into theoretical models. Our
qualitative study resulted in two separate models: (1) sarcasm
detection and expression model for SNSs and (2) sarcasm use
and non-use model for SNSs.
IV. SARCASM DET EC TI ON A ND EXPRESSION PRACTICES
Before discussing how sarcasm shapes users’ responses to a
content on social media, it is important to understand how our
participants recognize and express sarcasm on social media.
Broadly, the subjects we interviewed recognized sarcasm in
two ways: (1) unusual emotion/sentiment expression style and
(2) usual patterns of sarcastic posts.
A. Unusual Style of Sentiment Expression
The topics that are usually discussed on social media are
often subjective human interaction. That means, users discuss
their views, give opinions, and express their feelings about a
matter. As discussed earlier, a substantial amount of research
has been done to analyze the sentiment and emotion of these
user generated contents on social media. Usually, a particular
content/post generated by a user contains his/her views, and
thus the sentiment towards the corresponding topic. However,
in case of sarcasm, our participants report that this sentiment
in a particular post might seem unusual.
1) Exaggeration of Sentiments: Many of our participants
agree that exaggeration of sentiments in text is a sign of
a post being sarcastic. They think that in well-constructed
sarcasm, there are two objectives (1) to point out a flaw of
a targeted person (this was previously identified by previous
works) and (2) to entertain others if an audience is available,
which is common in usual social media settings. According to
participant P8,
“It does not matter what emotion you are showing, exag-
geration of it will automatically make your targeted person
confused whether it is sarcasm or not, since it is so common.
Your audience will often find it funny, so you get some people
on your side at least, even if the person who was your target
does not get the sarcasm.
While discussing this context further, an interesting reason-
ing was posed by our participants. According to them, when
one tries to make a general post, the objective is usually
to inform, to share opinion that will eventually lead the
audience to some direction. However, in posts with sarcasm,
the composer has no such motivation rather the sole goal here
is to make people laugh and that can be done by making the
post subjective. We found this reasoning plausible during our
quantitative analysis presented by [9].
2) Opposing Sentiments: In a subjective writing, a person
shares his/her positive or negative sentiment. As previous
studies have suggested, a sarcastic remark often has a negative
intended meaning. Our participants share the same view as the
study by Cliche et al. [23]. They say that in a sarcastic post
we can expect to observe opposing sentiments as part of the
text. This might be evident by their sentence construct: “Wow!
This is ugly” (example given by P6); here, the sentiment in the
first sentence is positive whereas it is negative for the second
sentence. As P7 gave us an example, “Terribly terrific”, such
phenomena can be observed at word level as well.
3) Wrong Use of Punctuation: All of our participants agree
that wrong use of punctuation is a usual clue for identifying a
sarcastic post. They say that this clue often occurs in sarcastic
remarks as a part of a conversation. Our participant P19 gives
his opinion with an example.
“Suppose, you are surprised and want to say “wow”, what
mark will you use? You will use exclamation mark with that.
But “wow” with a period after that just says that you are not
much impressed, rather you might be annoyed and are trying
to show your annoyance or callousness with a cold wow.
However, they also agree that though it is a usual clue, it
is not a very reliable clue. They think users generally want
to use social media with minimum effort. If they mistakenly
use wrong punctuation with a sentence, they often do not
care too much to edit the post to correct a single punctuation
mark. They might rather explain that it was a mistake and
correct later only if someone else pointed out at the wrong
punctuation.
B. Usual Structures/Patterns of Sarcastic Posts
Participants said that they look for clues in different parts
of a post. Some participants reported that the users who have
been on social media for a certain amount of time notice the
following things as notions of sarcastic posts: (1) exaggeration
of usually necessary emotions in writing, (2) popularly used
patterns of sarcastic posts that users learn with time, and (3)
opposing emotions/sentiments in different parts of a single
post.
1) Reference to Recent Objects: Our participants agree on
a very interesting aspect of sarcastic contents on social media.
They think there is a temporal factor to the pattern of sarcastic
posts on SNSs. As our participant P1 said,
“You know when Star Wars is a very popular movie. But
when a new Star Wars movie comes you can expect to see a
lot of sarcastic comments referencing to famous quotes from
the movie. Like, people might try to use “May the force be
with you.
We were curious to know whether it is the repetition of what
we explored as “reference to iconic object” earlier. Therefore,
we asked the participants about this. However, they think these
two are related but different factors. P1 clears up this in this
way:
“... No, you see, there are obviously some fans who can tell
you the movie’s name and what happened in a particular scene
when they hear a quote. But most people are not like that. They
watch, enjoyed, and may re-watch before a new movie in that
franchise comes. That’s when the craze is revived, and it will
make sense to use these reference only at that time. But sure,
if I am talking with my friends who, I know, lives in Star
Wars like me, hahaha! Then sure! I can use those reference
anytime.
P17 shares a different perspective about the temporal factor
of sarcastic posts’ pattern. He thinks recent events that get
popularity online may impact what users refer to for being
sarcastic. He thinks the frequency of these references are
maximum a little after when the original event got popularity.
With time, users are posed more new events that might be
referenced for sarcasm, and the earlier ones are not used as
many times as when they were first seen; however, regular
users might recognize and use those at times. When we asked
for example, P17 said,
“Few years ago, there was a live telecast of an interview
with general people in Rajshahi or Rangpur, I don’t remember
exactly, somewhere in northern Bengal during winter. The
reporter asked how the people felt about the winter. So, one
of them told that he did not like it and could not work for
winter in local dialect, and a particular word in that dialect
means something bad in proper Bengali. People in Central
Bangladesh made fun about that part of the interview a lot.
It became a popular sarcastic clue at that time. Every year
when winter comes, you will see some people to refer to that;
not as popular as before, but still it’s used.
This shows a periodical pattern in temporal factor of sarcas-
tic posts’ structure. Several other later participants agreed with
him. For example, P18 said it is usual to use some particular
reference periodically “every four years during the world cup”.
2) Association of Popular Memes/Meme-like Contents: A
major clue that our participants reported is association of
“meme-like” contents with the posts. Meme is usually an
image or short video (sometimes GIF) that is taken directly
or with slight variation from some popular media (e.g., TV
series, movies, etc.), and spread rapidly among the internet
users. For example, as many of our participants mentioned
about the presence of photos of Matthew Perry (who played
the character of Chandler Bing in popular TV series “Friends”)
in some special postures (as shown in Figure 1(a)) in inset of
images help them to identify the sarcastic intention of the
post. Discussion with our participants also gave us an idea
about other widely used images that are perceived as clues of
sarcasm in form of images. Use of hand-drawn meme-faces,
as shown in Figure 1(b), came as another example of such
categories of visual cues. Thus, while quite different from
each other with respect to the visual representation, all of them
depict the same sentiment of “sarcasm”.
3) Capitalization: All of our participants agree that capi-
talization of words in an SNS post denotes emphasized effort
from the composer for expressing his/her emotion. As we
have discussed earlier, participants agree that extra effort for
exaggerating sentiment might be a clue to sarcastic post.
Participants also agree that capitalization might also be used to
reverse the meaning or sentiment in a sentence. Our participant
P13 gave us an example of what he thinks is a popular form
of sarcasm of this pattern:
“If I say, the book is SOOOOO good that if you close it once
(a) (b)
Fig. 1. Qualitative study participants contributed/suggested samples of
images with sarcastic visual cues: (a) Matthew Perry in his popular posture
that work as indication of sarcasm for many participants. Thanks to Participant
P14 for providing us with the sample. (b) Samples of hand-drawn meme faces,
collected from: http://bit.ly/memefacesample
you wouldn’t want to open it again. It obviously has opposing
sentiments in a single sentence, but when I am using this type
of sentence in a conversation, I don’t want others to miss that
I made a sarcastic remark. So, it makes sense to emphasize to
catch their eyes.
In this step, we know how “unusual style of sentiment
expression” in a sarcastic post is achieved through a usual
pattern of posts.
4) Use of Arcane Style of Writing: We observed an inter-
esting way of conveying sarcasm among our participants from
Bangladesh. There are two forms of Bengali written language
Sadhu (more formal, used to be in practice until twentieth
century) and Cholito (less formal, currently is in practice).
Both use the same fonts, however, they vary in their preferred
use of words. Most of our participants from Bangladesh agreed
that Bengali sarcastic posts on social media are often written
in the arcane form. As one of our participants, P12 said,
“You know, no one in general, nowadays write in Sadhu
form. So, when you see a piece of text on Facebook that is in
Sadhu language, if it is not from some old books or something,
you instantly know there is something the person is trying
to do. I often find that posts written in Sadhu, are actually
sarcastic. At least the person is trying to say something funny,
if it’s not exactly sarcasm.
In this context, participants P14, P15 presented a related
insight. P14 opines that writing in this arcane form is not easy
for all as it has not been in practice for a long time. Therefore,
it is not often seen in quick sarcasm that comes as reply in a
conversation. Rather, it is seen in well-written satire posts that
took considerable effort from the writer of that post. Though
P15 agrees with P14 about the fact that this clue is not usually
seen in sarcastic comment in middle of a conversation, P15
has a different reasoning about this. P15 thinks the reason it is
not seen in “quick sarcasm” is less for the extra effort needed,
rather more for the fact that most people will not understand
the less-used words of this form of writing. According to P15,
“Who uses Facebook nowadays? Mostly young generation.
... They do not know this writing. Even many people of our
age do not know it very well. So, if you write that in middle
of conversation, they will either miss the sarcasm or ask for
explanation. It will lame if I have to explain myself after
making a sarcasm.
As we can see, though our Bengali speaking participants
agree that posts written in arcane form of Bengali writing
might be clue for the post to be sarcastic, it is often appli-
cable only for long and satirical posts for very concentrated
audience.
5) Wrong Spelling: This pattern of sarcastic posts was very
common among our participants from Bangladesh. They said
that it is a strong clue of Bengali sarcastic posts that they see
on social media. In Bengali, there are some pairs of letters
with very close sounds. In these pairs, one is softer than the
another for very similar sound. According to our participants,
using the hard sound in place of the soft one, and vice-versa
are clues of a piece of text to be sarcastic. However, they agree
that users do not do the same with text written in English.
Fig. 2. Examples of pair of soft and hard Bengali sounds for corresponding
single English sound. The list is not exhaustive.
In this context, most of our participants agree that this pat-
tern of sarcastic posts emerged recently. Though first Bengali
keyboard was published in 1988, it was fairly complicated for
users to learn. This limited the use of Bengali language on dig-
ital media. In 2014, a phonetic Bengali keyboard named Avro
was released. This made it easier for users to write Bengali
on computers, and eventually, helped increase the presence
of Bengali online. After that, it was possible to distinguish
50 letters of Bengali alphabet easily that could not be done
with 26 letters of English alphabet. Each Bengali letters pair
having similar sounds often has only one corresponding letter
in English (as shown in Figure 2). Since before 2014, most of
the Bengali users wrote Bengali using English fonts online, it
was not possible to use this hint for conveying sarcasm.
Participants P15, P16 raised another concern about this
clue to sarcasm. They said, as less educated people are not
often aware about the distinction about those sounds, they
spell words wrong unknowingly. Therefore, wrong spelling
in Bengali text can be thought as a clue to sarcasm only if the
post was composed by a person with schooling proper enough
to learn spellings of usually used words.
6) Use of Similar Sounding Words: Participants agree that
use of similar sounding words having different meanings is a
major clue for sarcastic posts on SNSs. They also think that
mashup of two words is also often deemed as sarcastic among
their audience. The reason they think it as a better clue for
sarcasm on social media is that posts are written and audience
have more time to put attention to details to understand the hint
themselves, unlike for in-person communication, it is difficult
to put such subtle hint on the go.
7) Reactions and Emojis: Our participants have commented
that reaction buttons and emojis often reverse the meaning of
a post. They described this dynamic in a bidirectional manner.
First, the post composer can associate the post with emojis that
are often used to joke on the internet. This might change the
tone of the post, in other words, make the post sarcastic by
creating a difference between surface sentiment and intended
sentiment of the post. This aligns with the theme of opposing
sentiment that we discussed earlier. As participant P2 said,
“If I see a friend to write something very serious, and put a
wink emoji at the end, I’ll know this person is being sarcastic
about his comment.
Second, all participants agree, in a sarcastic post, the
received reactions from the audience is always very mixed.
While some of the audience react to the intended meaning after
understanding the sarcasm, some might want to play along
with the sarcasm. Our participant P2 said,
“Suppose, you posted a sarcastic post about something that
annoys you, but you sarcastically said that you loved it. Many
of your peers will show annoyance as their reaction if they
understand the sarcasm. But many, specially my friends do it,
might want to keep the flow going by being positive about it
in their reactions and comments. Some might be just totally
lost.
Thus, a sarcastic post receives a mix of emojis and reactions
both from the composer and the audience that our participants
think as a usual pattern of sarcastic posts.
V. SA RC AS M USE A ND NO N-U SE O N SOCIAL MEDIA
We identified four kinds of SNS users with respect to their
use of sarcasm 3. This use comprises two functionalities –
detecting sarcasm and expressing sarcasm. (1) Non-users of
sarcasm means the users who cannot detect and use sarcasm on
social media. Mostly new SNS users fall into this category. (2)
Detectors are users who gain the experience needed to detect
sarcasm on SNS, but are not experienced enough to compose
sarcastic posts on their own, i.e., their sarcastic posts are often
misinterpreted by the audience. These users gain the ability to
detect sarcasm over time, though they cannot sarcasm very
effectively. (3) Consistent users are who can detect sarcastic
posts, and express sarcasm in their posts without much mis-
interpretation in most of the cases. (4) Disenchanted users
are experienced SNS users who can detect sarcasm in most
of the cases, and capable of composing such posts, however,
chose not to do so for some reasons like misunderstanding of
sarcasm from him/her among his/her peers (explained later in
this section).
A. Use of Sarcasm on Social Media
Some of our participants displayed enthusiasm for sarcasm
on social media. They think that people on social media in
general, should take social media lightly where they can make
small jokes about the happenings of their daily lives. They
believe sarcasm is a way to do that. Thus, sarcasm may work
as a driving force for making a content popular on SNSs.
According to our participant P10, this force works behind
popularity beyond online platforms as well. He describes SNS
as the place for him to get popularity, and sarcasm as the
driving force behind it. As he says,
Fig. 3. Different levels of sarcasm users on SNSs.
“I am one of the very first people in Bangladesh who were
regularly active on Facebook. There were some groups at that
time where I mostly wrote. I think my main strength is that I
write about things like politics, or day-to-day life using humor
or sarcasm. People like that. That actually made me popular.
Besides, several of our participants agree that with sarcastic
contents that refers to a recent event or that can be understood
with little or no context get a lot popularity.
B. Non-use of Sarcasm on Social Media
Unlike what we discussed earlier, some participants also
reported their reasons of non-use of sarcasm on social media.
Our participants present mainly two factors in this respect.
First, inexperience of using social media might present the
users a challenge while understating and conveying sarcasm
on social media. Our participants think older people are a large
part of this group. Our participant P1 says,
“It often happens that I am being ridiculous with my friends
on a sarcastic post, and my aunt comments in a serious tone.
Then, I have to explain that we are joking.
Second, previous bad experience of using sarcasm might
demotivate a user from using sarcasm on social media. Most
of the examples that our participants discussed had a common
pattern. They used a sarcastic remark, that was criticized
earlier. Or the flow might be opposite – where they were being
serious about something, and their audience did not take it in
the intended way staying under the hood of sarcasm. Either
way, it belittled the intention of the post, and that experience
demotivates the use of sarcasm.
VI. DESIGN IMPLICATIONS
In this section, we discuss implication from our findings
for making space for more engaging interaction among users.
SNS developers can consider these implications while design-
ing their system and customize their algorithms to organize
newsfeed.
A. Organizing SNS newsfeed
Our participants’ reasoning behind using sarcasm on SNSs
provides a design implication for SNSs. While the algorithms
social media platforms use to organize content is not known,
they might use some insights from this study. For example,
they can consider to show few sarcastic contents to new users
at first with a notification about those posts being sarcastic
and then slowly increase the amount of such contents in one’s
newsfeed if he/she likes those and as he/she becomes more
familiar with SNS sarcastic posts.
B. Checking on tempering of SNS algorithms
As we found out, sarcastic content can gain popularity
on SNSs. Thus, it is safe to assume that content having
a pattern similar to usual sarcastic contents might intrigue
users. However, clickbaits (a form of advertisement which uses
hyperlink text or a thumbnail link, and is designed to attract
attention and entice users to follow that link) might use this
insight to temper with SNS algorithms to achieve their own
objectives and trick users for that. SNSs might want to have
a check on such tempering-like activities.
C. Assisting users according to their preference
This provides another direction to the design implication
we mentioned earlier. While we proposed that SNSs can
introduce new users with sarcasm on SNSs over time, that
might seem to be an overhead to the experienced users. Thus,
instead of making that a feature of an SNS, web-browser
based extensions can be designed to help users to check
whether a particular content is sarcastic or not.
For the design implications we mentioned so far, machine
learning-based models to detect sarcasm on SNSs, as done
by [5], [8], [24], have to be made more robust and they can
benefit from the findings of this qualitative study.
VII. DISCUSSION
While using sarcasm on SNSs, users want to gain attention
from their targeted audience by making it clear instead of
doing it subtly. It becomes evident through both unusual
style of sentiment expression and usual patterns of posts.
For example, exaggeration of sentiment in statements, writing
style (e.g., capitalization, wrong spelling, arcane written form),
and association of usually sarcastic contents (e.g., memes,
emojis). This intention of making sarcasm clear to audience
can be driven by two factors–first, making it understandable
to inexperienced users, and second, targeting an engaging
interaction among users of various mix.
The temporal factor associated with sarcasm is quite in-
triguing. As we found from our study, recent events are often
referred to while making sarcasm-containing posts since it
helps many users understand that sarcastic remark. This again
relates to the strategy of users making sarcasm clear to a large
audience. Temporal factors of sarcasm on SNSs also showed a
periodic nature. As indicated by our study, while an event like
world cup, or release of new episode of a popular movie takes
place, sarcasm related to those during those times intrigue non-
followers to check out those and motivates a new larger group
of people to join the existing followers group. Thus, it achieves
an engaging interaction by motivating inexperienced users (in
this case, non-followers) with easy to follow content at first
and eventually including them into the community.
With the inclusion of more modes of data, users can
make sarcastic contents dividing the context and comment in
different modes (e.g., opposing sentiments in text and image)
or establish a usual pattern of sarcastic posts over time (e.g.,
memes, reaction emoticons). We also found the impact of
availability of technology with respect to languages as well.
As our non-English (Bengali) speaking participants reported
their way of expressing sarcasm changed with the invention of
easy-to-use phonetic keyboard. It differs from both the ways
of expressing sarcasm in English and how Bengali speaking
people used to do it before the introduction of the phonetic
keyboard.
Our qualitative study ended in thematic analysis of user
sarcasm behavior on social networks. Our data analysis re-
sulted in two models. First, the sarcasm expression model
discusses how users detect and express sarcasm on social
media providing valuable insights for building sarcasm detec-
tion model/system. Second, the sarcasm use-non-use model
discusses why users choose to or not to use sarcasm on social
media platforms that help identify design implications for SNS
platforms with respect to user sarcastic content sharing. Being
the first qualitative study on construct of sarcasm on SNSs,
this paper can serve to guide future sarcasm detection system
based research.
ACKNOWLEDGMENT
We would like to thank the participants of our study and
IRB at Missouri State University for approval of our study.
We thank Mrs. Megan Clark for proof-reading the paper.
REFERENCES
[1] M. S. Razali, A. A. Halin, N. M. Norowi, and S. C. Doraisamy,
“The importance of multimodality in sarcasm detection for sentiment
analysis,” in 2017 IEEE 15th Student Conference on Research and
Development (SCOReD), pp. 56–60, IEEE, 2017.
[2] R. Schifanella, P. de Juan, J. Tetreault, and L. Cao, “Detecting sarcasm
in multimodal social platforms,” in Proceedings of the 2016 ACM on
Multimedia Conference, pp. 1136–1145, ACM, 2016.
[3] R. Gonz´
alez-Ib´
anez, S. Muresan, and N. Wacholder, “Identifying sar-
casm in twitter: a closer look,” in Proceedings of the 49th Annual
Meeting of the Association for Computational Linguistics: Human Lan-
guage Technologies: Short Papers-Volume 2, pp. 581–586, Association
for Computational Linguistics, 2011.
[4] J. Tepperman, D. Traum, and S. Narayanan, ““ yeah right”: Sarcasm
recognition for spoken dialogue systems,” in Ninth International Con-
ference on Spoken Language Processing, 2006.
[5] E. Riloff, A. Qadir, P. Surve, L. De Silva, N. Gilbert, and R. Huang,
“Sarcasm as contrast between a positive sentiment and negative situa-
tion,” in Proceedings of the 2013 Conference on Empirical Methods in
Natural Language Processing, pp. 704–714, 2013.
[6] D. Bamman and N. A. Smith, “Contextualized sarcasm detection on
twitter.,” in International AAAI Conference on Web and Social Media
(ICWSM), pp. 574–577, 2015.
[7] R. W. Gibbs, “On the psycholinguistics of sarcasm.,Journal of Exper-
imental Psychology: General, vol. 115, no. 1, p. 3, 1986.
[8] D. Das and A. J. Clark, “Sarcasm detection on facebook: A supervised
learning approach,” in 20th ACM International Conference on Multi-
modal Interaction (ICMI), (Boulder, Colorado, USA), 10 2018.
[9] D. Das, “A multimodal approach to sarcasm detection on social media,
Master’s thesis, Missouri State University, Springfield, Missouri, USA,
July 2019.
[10] R. W. Gibbs Jr and H. L. Colston, Irony in language and thought: A
cognitive science reader. Routledge, 2007.
[11] E. Filatova, “Irony and sarcasm: Corpus generation and analysis using
crowdsourcing.,” in The International Conference on Language Re-
sources and Evaluation, pp. 392–398, 2012.
[12] R. J. Kreuz and S. Glucksberg, “How to be sarcastic: The echoic
reminder theory of verbal irony.,” Journal of experimental psychology:
General, vol. 118, no. 4, p. 374, 1989.
[13] R. Clift, “Irony in conversation,Language in society, vol. 28, no. 4,
pp. 523–553, 1999.
[14] D. Sperber and D. Wilson, Relevance: Communication and cognition,
vol. 142. Harvard University Press Cambridge, MA, 1986.
[15] H. L. Colston, “On necessary conditions for verbal irony comprehen-
sion,” Pragmatics & Cognition, vol. 8, no. 2, pp. 277–324, 2000.
[16] S. Kumon-Nakamura, S. Glucksberg, and M. Brown, “How about
another piece of pie: The allusional pretense theory of discourse irony.,
Journal of Experimental Psychology: General, vol. 124, no. 1, p. 3,
1995.
[17] H. Colston and R. Gibbs, “A brief history of irony,Irony in language
and thought: A cognitive science reader, pp. 3–21, 2007.
[18] “Grice’s maxims.” https://www.sas.upenn.edu/ haroldfs/dravling/grice.html,
2018. Accessed: May 19, 2018.
[19] A. Utsumi, “Verbal irony as implicit display of ironic environment:
Distinguishing ironic utterances from nonirony,Journal of Pragmatics,
vol. 32, no. 12, pp. 1777–1806, 2000.
[20] L. A. Goodman, “Snowball sampling,” The annals of mathematical
statistics, pp. 148–170, 1961.
[21] H. S. Ferdous, D. Das, and F. M. Choudhury, “Social media question
asking (smqa): Whom do we tag and why?,” in In Proceedings of
OzCHI’18, ACM, 2018.
[22] K. Charmaz and L. L. Belgrave, “Grounded theory,” The Blackwell
encyclopedia of sociology, 2007.
[23] M. Cliche, “The sarcasm detector.” http://www.thesarcasmdetector.com/,
2014. Accessed: May 19, 2018.
[24] D. Das and A. J. Clark, “Sarcasm detection on flickr using a cnn,” in
2018 International Conference on Computing and Big Data (ICCBD),
(Charleston, South Carolina, USA), 9 2018.
Chapter
With the advent of social media, publicly expressing opinions about businesses are hassle-free. Terms like ‘opinion mining’ and ‘microblogging’ have carved a niche. Organizations’ concern for brand reputation has accentuated the need for understanding customer’s choices and interests. Sentiment analysis scrutinizes users’ opinions through machine learning (ML) and natural language processing (NLP), a subset of which is sarcasm detection. The sarcasm detection database was fetched from Kaggle, containing nearly 30,000 tweet headlines curated for sarcasm detection, having high-quality labels with less noise. On implementing four ML algorithms, it is observed that training and testing data accuracies of random forest are 98.83% and 77.22%, linear support vector is 90.69% and 83.23%, logistic regression is 87.78% and 82.52%, Naive Bayes is 79.77% and 71.69%. This paper aims to build an RNN model based on LSTM architecture to accurately classify any given input as sarcastic or not.
Thesis
Full-text available
In recent times, a major share of human communication takes place online. The main reason being the ease of communication on social networking sites (SNSs). Due to the variety and large number of users, SNSs have drawn the attention of the computer science (CS) community, particularly the affective computing (also known as emotional AI), information retrieval, natural language processing, and data mining groups. Researchers are trying to make computers understand the nuances of human communication including sentiment and sarcasm. Emotion or sentiment detection requires more insights about the communication than it does for factual information retrieval. Sarcasm detection is particularly more difficult than categorizing sentiment. Because, in sarcasm, the intended meaning of the expression by the user is opposite to the literal meaning. Because of its complex nature, it is often difficult even for human to detect sarcasm without proper context. However, people on social media succeed in detecting sarcasm despite interacting with strangers across the world. That motivates us to investigate the human process of detecting sarcasm on social media where abundant context information is often unavailable and the group of users communicating with each other are rarely well-acquainted. We have conducted a qualitative study to examine the patterns of users conveying sarcasm on social media. Whereas most sarcasm detection systems deal in word-byword basis to accomplish their goal, we focused on the holistic sentiment conveyed by the post. We argue that utilization of word-level information will limit the systems performance to the domain of the dataset used to train the system and might not perform well for non-English language. As an endeavor to make our system less dependent on text data, we proposed a multimodal approach for sarcasm detection. We showed the applicability of images and reaction emoticons as other sources of hints about the sentiment of the post. Our research showed the superior results from a multimodal approach when compared to a unimodal approach. Multimodal sarcasm detection systems, as the one presented in this research, with the inclusion of more modes or sources of data might lead to a better sarcasm detection model.
Conference Paper
Full-text available
Social media question asking (SMQA) is an interesting application where users ask factual or subjective questions through social networks, also make invitations or seek favours, among other types of queries. Topics like what we ask, what motivates us to answer, how to integrate the traditional search engines with SMQA, etc. have been well investigated. However, the effect on tagging particular people in queries is yet to be explored. In this work, we focus on targeted queries in social networking sites, where people tag some of their friends, but also remains open to others who might want to respond. We conducted a two-phase study to investigate users tagging behaviour based on question topic and type, their rationale behind tagging those particular people, and corresponding outcomes of tagging. Our result contradicts with the existing works that tried to use automated tagging in social networks and identify design opportunities that need to be considered while developing new solutions to assist in this regard.
Conference Paper
Full-text available
Sarcasm is a common feature of user interaction on social networking sites. Sarcasm differs with typical communication in alignment of literal meaning with intended meaning. Humans can recognize sarcasm from sufficient context information including from the various contents available on SNS. Existing literature mainly uses text data to detect sarcasm; though, a few recent studies propose to use image data. To date, no study has focused on user interaction pattern as a source of context information for detecting sarcasm. In this paper, we present a supervised machine learning based approach focusing on both contents of posts (e.g., text, image) and users' interaction on those posts on Facebook.
Conference Paper
Full-text available
Sarcasm is an important aspect of human communication. However, it is often difficult to detect or understand this sentiment because the literal meaning conveyed in communication is opposite of the intended meaning. Though the field of sentiment analysis is well studied, sarcasm has often been ignored by the research community. So far, to detect sarcasm on social media, studies have largely focused upon textual features. However, visual cues are an important part of sarcasm. In this paper, we present a convolutional neural network based model for detecting sarcasm based on images shared on a popular social photo sharing site, Flickr.
Article
Full-text available
Sarcasm is a peculiar form of sentiment expression, where the surface sentiment differs from the implied sentiment. The detection of sarcasm in social media platforms has been applied in the past mainly to textual utterances where lexical indicators (such as interjections and intensifiers), linguistic markers, and contextual information (such as user profiles, or past conversations) were used to detect the sarcastic tone. However, modern social media platforms allow to create multimodal messages where audiovisual content is integrated with the text, making the analysis of a mode in isolation partial. In our work, we first study the relationship between the textual and visual aspects in multimodal posts from three major social media platforms, i.e., Instagram, Tumblr and Twitter, and we run a crowdsourcing task to quantify the extent to which images are perceived as necessary by human annotators. Moreover, we propose two different computational frameworks to detect sarcasm that integrate the textual and visual modalities. The first approach exploits visual semantics trained on an external dataset, and concatenates the semantics features with state-of-the-art textual features. The second method adapts a visual neural network initialized with parameters trained on ImageNet to multimodal sarcastic posts. Results show the positive effect of combining modalities for the detection of sarcasm across platforms and methods.