J. Inst. Eng. India Ser.

B (February 2022) 103(1):259–272

https://doi.org/10.1007/s40031-021-00620-7

REVIEW PAPER

Exploiting Emojis in Sentiment Analysis: A Survey

Vandita Grover1

Received: 18 December 2020 / Accepted: 15 May 2021 / Published online: 12 June 2021
 The Institution of Engineers (India) 2021

Abstract Sentiment analysis is now a prominent field of Introduction

interest owing to a growing trend of users expressing their
opinions on social media, review pages, feedback forms, Sentiment analysis identifies text sentiment and classifies
and other online channels. The machine learning approach the underlying sentiment’s polarity into positive, negative,
to sentiment analysis focuses on feature extraction methods or neutral. Text can be analyzed for sentiments at three
like constructing lexicons to learn sentiment polarity or levels – Document level, Sentence level, and Aspect Level
learning word embeddings and applying them for their use [1].
in machine learning algorithms for sentiment classification.
1. Document Level At the document level sentiment
But most popular machine learning approaches still cannot
analysis investigates whether the entire document has a
capture nuanced emotions like sarcasm, irony, etc. Emojis
positive, negative, or neutral sentiment. Documents
are now being used along with text by the users to express
like blogs, forum discussions, reviews, author insights,
emotions and hence can help researchers improve senti-
feedback questions in surveys, etc., generally have
ment classification tasks. Sentiment analysis powered by
detailed, yet multiple opinions at different positions in
emojis is still in the nascent phase and has gained some
the document. An overall sentiment score of a
pace in the last five years. The primary goal of this paper is
document can help infer the overall sentiment of the
to discuss the use of emojis that supplement the text to
document and extract useful information without the
express different emotions. This paper compares some
need to go through the entire document.
traditional text-based word embeddings and lexicons. Then
2. Sentence Level analysis classifies the sentiment polar-
the paper discusses the evolution of emoji-based lexicons
ity at the sentence level. Sentences are considered as
and emoji embeddings. Further, some deep learning
short documents and there are two classes of classifi-
approaches using emojis to improve existing sentiment
cation [2]. One is to determine if a sentence expresses a
classification tasks are studied. The main contribution of
sentiment or not and the other to classify the sentence
this paper is to survey various approaches to use emojis in
as positive, negative, or neutral. Sentence level clas-
sentiment analysis which to the best of our knowledge has
sification has become more popular with the rise in
not been done till now.
social media usage. YouTube comments and Twitter
tweets can help understand user sentiment better on
Keywords Sentiment analysis  Emojis 
trending topics like elections, pandemic, and even to
Sentiment classification  Emotion detection
glean insights on user view of the products.
3. Aspect Level Many times, it becomes important to
understand user opinion at a nuanced level. For
& Vandita Grover instance, a user may like the story and cinematography
vanditagrover@andc.du.ac.in of a movie but may not be too happy about the acting.
Thus, it is important to understand the different aspects
1
Department of Computer Science, Acharya Narendra Dev
College, University of Delhi, New Delhi, India

123
260 J. Inst. Eng. India Ser. B (February 2022) 103(1):259–272

at a granular level and analyze users’ sentiments for • Multi-attribute Split may use the similarity of docu-
different aspects. ments to words or phrases or discriminants like Fisher
discriminant for the split.
A summary of the three levels of sentiment analysis is
given in Fig. 1.
Sentiment classification techniques can be classified as Unsupervised Learning
follows.
Does not have pre-trained or pre-labeled data. Since we do
Supervised Machine Learning Approach not invest time in training our classifier model, the learning
process is simplified and also becomes faster.
In the supervised approach training data is available for
• In the lexicon-based approach, a sentiment lexicon is
each class of the sentiment. The classification model is
created. Sentiment words like adverbs or adjectives and
trained to learn document characteristics. Classification
sometimes nouns and words are associated with posi-
algorithms include.
tive or negative orientation. Then overall polarity of the
• Probabilistic models like naı̈ve Bayes, Bayesian Net- document is computed by combining all the positive
work, Maximum Entropy [2] use mixture models that and negative polarities of the words present in the
assume that each class is a component of the mixture. document and according to the collective score the
Each mixture component provides the probability of document is classified as positive for a positive overall
sampling a term of the component [1]. score, negative for a negative overall score, and neutral
• Linear classifiers like Support Vector Machines (SVM) if the overall score is zero.
and Neural Networks interpret the document as an n-
The following strategies are typically used for automatic
dimensional feature space X = (9 1, 9 2, …, xn),
lexicon creation
combine it with a vector A of the same dimension,
where A = (a1, a2, …, an) and a scalar B. The output Y 1. A dictionary-based lexicon takes a pre-determined set
of the linear classifier determines the sentiment class of of words with known orientations and then the set is
our target [1], where Y is given as augmented with similar or opposite words iteratively,
0 till no new words are found [1].
Y ¼ AX þ B
2. Corpus-based lexicon takes context into account for
Decision-tree classifiers decompose the training determining the sentiment of words. Syntactic patterns
documents’ feature space to split on condition(s) on and words that frequently occur together with an initial
attribute(s), recursively, till the leaf nodes contain at least set of words, together determine new sentiment words
a minimum threshold of records for classification [1]. with a context-specific orientation to build the corpus.
There are a variety of splits:
• Single Attribute Split uses the presence or absence of
certain words at a node to determine the split.

Fig. 1 Levels of Sentiment Analysis

123
J. Inst. Eng. India Ser. B (February 2022) 103(1):259–272 261

Hybrid Approach an underlying model of word representation will not

result in improving sentiment classification as demon-
Can combine the supervised and unsupervised approaches strated by Bansal and Srivastava in [8]. Also, the size
to determine the sentiment of a document. Some approa- of the training set can drastically impact the quality of
ches like in [3] by Elshakankery et al. propose a hybrid the word vectors [4]. Being statistical in nature it
method that builds a lexicon to be further used to train a cannot handle out of vocabulary words and may
machine learning model. require creating new representations for different
In most of the work done so far, the only text has been languages.
used to extract features to create lexicons and word 2. Global Vectors for Word Representation (GloVe)
embeddings. These have been fed to machine learning proposed by Pennington et al. in [9] is a global log-
classifiers to learn the underlying sentiment polarity. bilinear regression model for learning word represen-
tations in an unsupervised manner. It captures word
analogy, word similarity, and identifies named entities.
Popular Text Sentiment Analysis Approaches If i and j are two words say ice and steam whose
with Word Embeddings and Lexicons relationship is to be analyzed with word k, say solid.
Ice and solid are more likely to co-occur together as
Most machine learning sentiment analysis follows the text- compared to steam and solid. To capture such co-
based approach to extract features to learn word embed- occurrences, more weights are assigned to frequently
dings or create word lexicons. In a supervised approach co-occurring words as compared to rare co-occur-
word embeddings or vectors are learned to be fed to tra- rences.
ditional machine learning classifiers or deep layered net- GloVe captures word analogies and takes into account
works. Sentiment lexicon creation is an unsupervised global occurrences of words in a given text corpus, but,
approach where a sentiment lexicon is created from the He et al. in [10] demonstrate that in the GloVe
word corpus to assign each word a sentiment polarity. In embeddings sentiment information is not adequately
subsequent subsections, popular word embeddings and captured by GloVe.
lexicons widely used for sentiment analysis are discussed. 3. Sentiment Specific Word Embedding to Classify Twitter
Sentiments (SSWE) was proposed by Tang et al. [5].
Text Feature Extraction for Word Embeddings They extend the C&W model [11] proposed by
Collobert et al. and use three neural networks to
The text needs to be represented or modeled in the form of incorporate sentiment information to learn SSWE. The
the real-valued vectors in a vector space. Each word rep- basic model SSEWh predicts the sentiment polarity of
resents some point in the vector space whose dimensions an n-gram while sliding the n-gram window across the
can be determined through experimentation. These word sentence. A hard constraint of i.e. [1,0] for positive and
vectors are learned through neural networks with the [0,1] for negative sentiment classification is applied on
backpropagation algorithm. The word vectors eventually resulting multinomial distribution from a SoftMax
learned can then be passed to the machine learning or deep layer on top of the C&W layer. This hard constraint is
learning classifiers. Some popular approaches to learn word relaxed in the SSEWr model, such that if the predicted
embeddings are discussed as follows. positive score is greater than the predicted negative
score the tweet sentiment polarity is positive, and,
1. Google’s Word2Vec [4] is based on continuous bag-of-
negative in the reverse situation. The SoftMax layer is
words (CBoW) [5] and skip-gram [6] architectures for
removed from SSEWr as probabilistic interpretation is
word representations. CBoW captures local context
not required. SSEWh and SSEWr are unified with the
while the Skip-gram model takes into consideration the
C&W model in the SSEWu model to capture sentiment
context utilizing the partial sequence structure of a
as well as syntactic information. This SSEWu model
sentence as a window for the target word. Skip-gram
predicts the language model score and sentiment score
captures the syntactic properties too [7].
respectively of an n-gram. The results show that the
Word2Vec can be used to cluster words that have
sentiment-specific word embeddings effectively distin-
semantically similar features.
guish between words of opposite polarity.
An important advantage for Word2Vec embeddings is
that they take into account semantic similarity and The model captures positive and negative sentiment
work well even for noisy data. Since context coverage information along with the syntactic information. But the
is largely dependent on the type or domain of the model assumes each word’s sentiment polarity in the local
document, sentiment classification with skip-gram as window context and not how it contributes in the context of

123
262 J. Inst. Eng. India Ser. B (February 2022) 103(1):259–272

the entire tweet or sentence [12]. Thus, a word may have a semantic relationship or their meanings i.e. if they are
positive polarity but it might have been used in the nega- polysemic or monosemic [16]. Words of similar
tive context in the Tweet to express sarcasm. Consider the meanings are grouped in synsets. Each sense of the
tweet, ‘‘I love being ignored #sarcasm’’ [13], where the word is a node and all word sense nodes are linked by
polarity of the word love is positive but the global senti- various semantic relationships like synonyms or
ment of the tweet is negative. antonyms, meronyms or holonyms (consists of),
A comparative analysis of the above-discussed word hypernyms or hyponyms (is-a) [16]. SentiWordNet
embedding approaches is tabulated in Table 1. ranks synsets or terms based on subjectivity i.e., the
meaning they represent in the part of speech. Each
Text Sentiment Lexicons term is extracted and matched with the SentiWordNet
database to assign a positive or negative orientation.
The lexicon approach to sentiment analysis computes the All the terms are used to compute the overall score of
polarity of each word and combines the sentiment scores of the document to determine its expression of sentiment.
each word to compute the overall polarity of the document. The major drawback of this approach is that since
Few sentiment-specific lexicons that have gained popular- individual terms are extracted, documents that do not
ity in the past few years are discussed below. express any opinion (neutral) or the ones that express
sarcasm or irony can often be missed.
1. SentiWordNet [15] is a readily available resource that
2. VADER was developed by Hutto and Gilbert as
contains term sentiment information of words in
Valence Aware Dictionary for sEntiment Reasoning
WordNet [16]—an electronic lexical database. Words
[17] by using quantitative and qualitative features to
in the WordNet database are organized based on
produce a gold-standard lexicon [17]. They used

Table 1 Comparison of some Word Embedding Approaches for Text Sentiment Classification
Model Approach Type Advantages Disadvantages Complexity

Word2Vec Vector representations using Semi- Cosine Dependent on training O (N log V)

CBoW or skip-gram supervised distance of set size N: training corpus size
architecture vectors
Limited to training V: unique words in the
helps
corpus vocabulary vocabulary
capture
linguistic Only localcontext can be
similarities handled
Word sense
disambiguation not
possible
GloVe Works on Global counts and takes Unsupervised Takes global Sentiment information Worst Case: O(|V|2)
into account frequently co- co- not adequately |V|: words in the vocabulary X
occurring words occurrences captured matrix denotes word-word co-
of word occurrence which is sparse
Captures word Limited to training |C| corpus size
analogies corpus vocabulary
|V| combination of word-word co-
occurrence and Xi,j = k/ (ri,j)a is
modeled as power-law function
of frequency rank ri,j
(
Tighter Bound:
 OðjC jÞ  a\1
|jj X j
O jC j1=a a [ 1
Sentiment Learn sentiment specific word Supervised Takes into Focuses on sentiment SSWEh gets computationally
Specific embedding by training three account polarity of a word in expensive with the SoftMax
Word different neural networks and syntactic as the local context and layer taking O(|V|) time to
Embedding integrating sentiment well as not the global compute a probability
information in their loss semantic sentiment of the distribution over V words
functions information sentence SSWEr although removes the
softmax layer it still takes
7 weeks to complete with
|V|= 130,000 [14]

123
J. Inst. Eng. India Ser. B (February 2022) 103(1):259–272 263

popular word banks like LIWC [18], ANEW [19], and • The constraint of Sentiment Prior is a general-purpose
GI [20]. For lexical features, they included western sentiment lexicon that gives good prior information.
emoticons of the form:-), sentiment-related acronyms Xn
minf j¼1 I G
j jSj  Gj jg
like LOL, ROFL, etc., and slangs like meh, Nah, etc.
that have some sentimental value. They could generate
9000 candidate lexical features with this process. The IjG = 1 and Gj = L(wj) if wj exists on lexicon L, 0 otherwise
intensity rating [– 4–4] of candidate features was This favors sentiment score assignment S that is closest to
collected from 10 independent human raters to estab- the general-purpose lexicon G.
lish the ground truth, where – 4 was extremely • The constraint for Overall Sentiment Rating that is
negative and 4 was extremely positive. 7500 lexical overall sentiment rating is positively correlated with the
features with a non-zero mean rating and standard sentiment score of the words in the document.
deviation less than five were kept with their valence Xm
scores that indicated their sentiment polarity. The minf i¼1 I O i jf ðd i ; SÞ  Oi jg
lexical features were combined with five generalizable
rules to include commonly used grammatical and
Oi is the overall sentiment of the text di normalized to [–
syntactical conventions that express the intensity of the
1,1]
sentiments, viz. punctuation, capitalization, degree
f(di, S) predicts the sentiment score of the text di and can be
modifiers, contrastive conjunction, and negation.
defined by averaging the sentiment scores of (aj,wj)
VADER exploits commonly used lexical features to
compute sentiment polarity but ignores features with X
n

mean zero which might hold a clue to ironical f ð d i ; SÞ ¼ X ij Sj

y¼1
statements. Also, statements that may appear neutral
or even positive may have underlying sarcasm, which
VADER doesn’t capture. Xi,j is the number of times a pair j occurs in text di. In the
3. Context-Aware Sentiment Lexicon Lu et. al in [21] case of negation clauses, while estimating Xi,j the count of
construct an automated context-aware sentiment lexi- the occurrence is replaced with -1.
con to supplement a general-purpose lexicon, based on IiO indicates if the overall sentiment is defined.
an assumption that a set of domain-specific aspects • The constraint for Similar Sentiment that is words wj,
A = {A1, A2,…, Ak} are available beforehand. wk appear as synonyms in thesaurus or pairs j and k are
concatenated with ‘and’ then their sentiments will tend
where Lc is the sentiment lexicon dictionary of opinion
to become similar. The objective function is thus given
words w conditioned on various domain aspects. Instead of
as
relying on human labeling several signals like those from ( )
the General-purpose sentiment lexicon, overall sentiment X n X n

rating like TripAdvisor, Epinion, Amazon reviews, The- min Ajk jSj  Sk j
j¼1 k¼1
saurus like WordNet, and Linguistic Heuristics like and,
but, and negation rules are used. Signals from multiple where Ajk e [0,1] is the confidence that pairs j and k have a
sources are combined so that context-dependent informa- similar sentiment.
tion is not missed if a signal is weak. • The constraint for Opposite Sentiment is if words wj, wj
To generate candidate lexicon entries clauses are chosen appear as antonyms in thesaurus or pairs j and k are
as a unit to associate opinion words more accurately with concatenated with ‘but’ then their sentiments will tend
the aspects. The opinion text is segmented by tagging the to become opposite. Here the sign of the sentiment
clause with the aspects whose word cluster overlaps with score is involved because sentiments are at two
the words in the clause. The non-aspect and non-stop words extremes, but their absolute scores are assumed to be
are considered to be opinion words of the tagged aspects. close. Two non-negative variables Sj? and Sj- are
For more than one tag for a clause with different aspects, introduced where at most one is active and the other is
the potential opinion words are paired with each aspect. zero. The active variable represents the sign of Sj and
Sj e [-1,1] is the sentiment score of aspect-opinion pair the value represents the absolute value of Sj.
(aj,wj)
This can be represented as
An objective function is then designed to assign a sen-
timent score to lexicon entries, under the following
constraints:

123
264 J. Inst. Eng. India Ser. B (February 2022) 103(1):259–272

( )
n 
X  sentiment score close to zero. This makes detection of
min Sþ
j þ S
j ; nuanced emotions of sarcasm and irony difficult to spot.
j¼1 Depend on domain-specific characteristics of the text
subject to Sj ¼ Sþ  þ 
j þ Sj and Sj ; Sj  corpus that makes it difficult to generalize the sentiment
for j = 1 to n analysis approach.
Thus, novel approaches are needed to improve senti-
which forces at least one of Sj? or Sj - to be zero. The
ment classification. Emojis can become powerful tools in
final objective function thus becomes
(   !) sentiment analysis. For instance, they can supplement the
 þ 
Pn Pn Sj  Sj þ text and help determine complex emotions. This new
min j¼1 k¼1 B jk where Bjk e [0,1]
jS þ
j  Sk j
approach in the form of Emoji Powered Sentiment Anal-
ysis is discussed in the next section. Research work that
is the confidence that pairs j and k have opposite sentiment.
focuses on constructing emoji lexicons and emoji-based
The optimization problem is then transformed into an embeddings over the past few years are also discussed. In
equivalent linear programming problem with appropriate subsequent sections, some of the proposed deep learning
weighting parameters for each constraint set to a degree of approaches using emojis have been discussed followed by
trust on each source. tracking the temporal evolution of these approaches,
The proposed method performed best for F1 measure on challenges, and lastly, the future scope in the field of
two data sets used and gave mixed results for precision and emojis in sentiment classification.
recall against baselines, viz. Random for randomly guess-
ing sentiment, MPQA to predict the sentiment of the
opinion word with MPQA lexicon [22], General Inquirer Emojis in Sentiment Analysis
[23], and the Global Prediction Method [24].
This work handles opposing sentiments by taking into Emojis are fast becoming a new, simpler, yet expressive
account positive and negative signs. But the framework mode of communication across digital platforms and thus
depends heavily on signals like prior information, ratings hold tremendous potential to improve sentiment classifi-
which may differ across datasets, thus making it difficult to cation and emotion detection tasks. Emojis are pictographs
generalize the lexicon. having textual descriptions that have been converted to
A comparative analysis of the above lexicon approaches Unicode codes by Google [25].
is given in Table 2. According to a study by Konrad, et al. in [26] emojis are
used for—expressing emotions; emphasizing messages,
Issues and Challenges clarifying, acknowledging and reciprocal exchange, etc.

As seen above, the lexicon and word embedding approa- Here is an example of a conversation involving
ches tend to. emojis.

Separate opposing sentiments. Also, sometimes words or A: How was your day?
phrases are dropped because of a lower mean score or B: Fine.

Table 2 Comparison of above Sentiment Lexicon Approaches

Lexicon Approach Level Advantages Disadvantages

SentiWordNet Computes sentiment polarity of the words in Document- Takes into account the Individual words are used to compute
the WordNet database. All the words are level semantic meaning of sentiment polarity so parts in the
used to compute the overall document the words in a document that express irony or
sentiment score document sarcasm can often be missed
VADER Combine lexical features with grammatical Sentence Can compute polarity and Ignores features with mean zero that may
and syntactical conventions to compute Level sentiment intensity miss sarcasm or irony
sentiment polarity using commonly used
lexical features
Context- Uses linear programming optimization Aspect Takes into account Heavily dependent on signals, making it
Aware framework under four constraints of Prior Level opposing sentiments in difficult to generalize across different
Sentiment knowledge, Overall Rating, and Opposing a context datasets
Lexicon and Similar Sentiments

123
J. Inst. Eng. India Ser. B (February 2022) 103(1):259–272 265

EmoTwi50 dataset with 50 pairs of emojis for their func-

tional similariÐty and relatedness. Here functional simi-
larity means if the two emojis are similar and relatedness
means if they can be used together. The variants of skip-
gram modeled varied on.
1. Dimensions of vectors: where various dimensions
between 50 -700 were tested and model with 300
dimensions gave better results.
2. Different window sizes between 3–12 (for skip-gram
model) were experimented with and models with
Fig. 2 ML approaches to Emoji Powered Sentiment Classification smaller window sizes (between 3–6) gave better results
as tweets are generally short in length.
The use of has made it evident that B’s response is 3. The models were also trained on different datasets as
neither neutral nor positive. Here ‘‘Fine’’ is being used follows. The raw filter dataset had the entire context of
sarcastically and the use of emoji has made it clear that B is tweets, punctuations, and emojis. The clean filter
not actually Fine. removed punctuation and stop words from the tweet.
Singh et al. in [27] replace emojis with their text The only filter retained only the emojis in the dataset.
description, as it is common to find word embeddings of A quantitative and qualitative evaluation for the emoji
the text. They experiment with a two-stack BiLSTM neural embeddings from different variations of the skip-gram
network and show improvement in sentiment classification model was conducted.
and irony detection, concluding that emojis are more For quantitative evaluation, a human gold standard was
effective in sentiment analysis. created where 8 participants rated the 50 pairs of emojis
Barbieri et al. have researched various aspects of emojis, that were then compiled in the EmoTwi50 dataset. Simi-
viz exploring emoji usage and meaning over different larity, relatedness, and average Pearson correlation were
languages [28], emoji-based skip-gram models [29], computed between human gold standard and different
Twitter emoji usage in Barcelona and Madrid [30], vector space skip-gram models generated. The clean
studying emojis predictability [31], studying emoji pre- dataset showed more balanced results for similarity and
diction [32] and predicting emojis [33] and [34]. relatedness, which shows that the best way to filter the
Various approaches to construct emoji-based lexicons dataset is by removing punctuations and stopwords.
and learning emoji embeddings for improving sentiment The models could correlate emojis in the EmoTwi50
classification have been experimented with. Apart from dataset, better for relatedness than the similarity task. Thus,
traditional machine learning algorithms, nowadays while creating embeddings from Word2Vec [4] algorithm
researchers have started to work with deep learning-based both emojis in the pair were considered in the context of
classifiers given the tremendous text data on social media each other which influenced the creation of both their
and the web and the unprecedented processing power of embeddings.
modern-day computer systems. For qualitative evaluation, the emoji vectors of the 100
Figure 2 shows various machine learning approaches for emojis in the EmoTwi50 dataset were plotted using
Emoji-powered sentiment analysis. t-Distributed Stochastic Neighbor Embedding (t-SNE). The
t-SNE technique helped reduce the 300-dimensional vec-
Emoji Embeddings tors into two dimensions. Similar vectors like happy faces
were grouped together but well separated from the sad face
Some attempts to learn emoji embeddings to complement vectors.
text word embeddings for sentiment classification tasks are Also, K-means clustering was used to build 11 clusters
been discussed below. of 300 most frequent emojis. Despite the noise in a few
clusters most of the clusters were reported to have a clear
A Vector Space Skip-Gram Model for Twitter Emojis identity and consistency in groupings.
This work presents emoji embeddings created based on
Barbieri et al. in [29] mapped words and emojis of tweets emoji context. But how these emojis impact the sentiment
in the same vectorial space to build various skip-gram polarity when combined with the words in a sentence is yet
models. Several variants of the skip-gram model with to be studied.
different vector dimensions and different lengths of win-
dow size were experimented with to compile the

123
266 J. Inst. Eng. India Ser. B (February 2022) 103(1):259–272

Emoji2Vec represent the emoji on different platforms; Ri is set of

related emojis; Hi is the set of categories ei belongs to; Si is
In [35] Eisner et al. trained Emoji2Vec on Unicode the set of senses in which ei can be used.
descriptions of emojis to augment sentiment classification Using EmojiNet meaning of the emojis in three different
performance of classifiers. The emojis along with their forms were derived, viz.
name and description were collected from the Unicode
1. Emoji Descriptions a description of the emoji’s
emoji list [36]. For every emoji and its word description,
intended use
the corresponding vectors of the words in the description
2. Emoji Sense Labels emoji’s sense and part of speech
were combined using Word2Vec.
tag pairs
The emojis along with their name and description were
3. Emoji Sense Definitions text descriptions that explain
collected from the Unicode emoji list and were trained
the sense label for which an emoji can be used in a
using Word2Vec embeddings. Thus, for every emoji and its
sentence.
word description, the corresponding vectors of the words in
the description were combined using Word2Vec. These three machine-readable emoji meanings are con-
X
N verted to vector representations using publicly available
vd j ¼ wd k word embeddings like Word2Vec. Twitter and Google
k¼1 News corpora are used to train emoji embedding models.
where vdj is a vector representation of the emoji The learned emoji vectors are used to replace the emojis.
description. The emoji embeddings are then obtained by averaging the
wdk is the word embedding vector for the kth word in the word vectors of all the words in emoji definitions to form a
description of emoji j. 300-dimension vector. Sense_all set combines all the
A vector xi for every emoji in the training set was learned emoji embeddings. Thus, by using sense definitions
trained and probability for a match between representation from EmojiNet different emoji embeddings were evaluated
vector xi and description vector vdi was modeled using to measure the similarity of emojis. 10 human evaluators

r xTi  vd j_ were roped in to create the EmoSim508 data set. Emo-
The approach presents an interesting take on creating Sim508 consists of 508 Emoji pairs to analyze how well
vector representation of emojis trained by their description, the emoji embeddings performed in the similarity calcu-
but it does not capture the local context-dependent usage of lation task and results showed that the proposed emoji
emojis which could be important in capturing nuanced embedding models could be used in real-world NLP and
emotions of sarcasm, frustration, irony, etc. sentiment analysis tasks.
Although the embedding approach takes into account
Semantic-Based Measure for Emoji Similarity emoji meaning, description, and intended use to learn the
embeddings. It is yet to be seen how these sense-based
Wijeratne et al. [37] built a model to learn emoji embed- emoji embeddings perform on classification tasks like
dings from emoji meanings found in EmojiNet [38]. These sarcasm and irony detection.
emoji meanings in EmojiNet [38] have been constructed
from the Unicode Consortium, Emojipedia, the Emoji Emoji Lexicons
Dictionary, and BabelNet to create a nonuple. This nonuple
is a mapping of emojis which includes names, their In this subsection different emoji-based lexicons to
meanings, descriptions, keywords, and images with their understand underlying sentiments of text supplemented
possible senses. The Emoji Dictionary lists sense labels as with emojis are discussed.
different uses of the emoji. For instance, if an emoji has a
face keyword and a face can be used as a noun or a verb or Automatic Construction of Emoji Lexicon
an adjective then the emoji with the keyword ’face’ in its
description will have three sense labels. Then the BabelNet Kimura and Katsurai in [39] constructed an emoji senti-
is used which lists different sense definitions for a sense ment lexicon automatically by extracting sentiment words
label. The most appropriate sense definition is selected from WordNet-Affect [40]. WordNet Affect works like an
using a Word Sense Disambiguation task. Finally, each English dictionary to assign affective sentiment labels to
emoji is represented as ei = (ui, ni, ci, di, Ki, Ii, Ri, Hi, Si). synsets of WordNet [16]. For a word in the dataset, the
For every ei: ui is its Unicode representation; ni is its name; corresponding synset in WordNet was derived and if the
ci is the shortcode; di is the description; Ki is keywords set synset also existed in the WordNet-Affect it was labeled by
to describe its intended meaning; Ii is images used to one of the sentiment categories in a set S; eg. S could be
{joy, love, sadness}. A word was labeled with the most

123
J. Inst. Eng. India Ser. B (February 2022) 103(1):259–272 267

frequent sentiment and added to the sentiment word list. tweet or a sentence. For instance, consider a tweet I am
The word was discarded if the number of occurrences of fine. This tweet may have a neutral sentiment, but if it is
different affective sentiments was equal to avoid ambigu- clubbed with an emoji-like (neutral face emoji) [42], the
ity. An emoji sentiment score vector ES(ei, s) was com- tweet’s polarity may change to negative because even
puted for a target emoji by evaluating the ratio of co- though the emoji has a neutral face it is used to convey
occurrence frequency of the emoji with a sentiment word. mild irritation and concern or dry humor [42].
Large ES(ei, s) indicates a close relationship of emoji i with Thus, it needs to be experimented on how the sentiment
the sentiment s. The corresponding emoji vector with s polarity of the text along with the polarity of an emoji
dimensions is given as {ES (ei, s)}s e S. together help determine the overall polarity of the tweet.
The method does not require manual sentiment labeling Studying this may help uncover complex emotions like
and builds an automatic sentiment lexicon by computing a sarcasm, irony, frustration, etc.
five-dimensional sentiment score for an emoji. But this
approach has a limited set of words as it refers only to the
words in WordNet-Affect. The method is unable to dis- Sentiment classification with emojis – the deep
tinguish between nuanced emotions of anger, fear, and learning approach
disgust. A reason for this could be the discarding of equal
co-occurrence of different affective sentiments which Some researchers have made attempts to improve the
needs to be further evaluated. sentiment classification tasks for given text corpus. Deep
learning models like LSTM and Recurrent Neural Net-
Sentiment of Emojis works are fast becoming popular for sentiment classifica-
tion with emojis. Here is a discussion on some work that
Novak et al. [41] computed the sentiment of the emojis incorporates emojis while learning the sentiment labels
from the sentiment of the tweets in 13 European languages, using deep learning techniques.
labeled for sentiment by 83 native speakers. The sentiment
of a tweet was represented by a 3 valued discrete variable Bi-sense Emoji Embedding and Attention Based
for the sentiment class, c 2 f1; 0; 1g, where the sentiment LSTM for Twitter Sentiment Analysis
for each tweet is negative, neutral, and negative in that
order. A discrete probability distribution for each category Chen et al. in [43] propose an emoji attention scheme for
(p-, p0, p?) was formed, Twitter Sentiment Analysis. They consider each emoji as a
  special word and initialize emoji embeddings using fas-
p ; p0 ; pþ ;
X tText [44]. The emoji is represented by two vectors that
pc ¼ 1; mean two senses; one each for positive and negative sen-
c timent context. They propose two schemes, viz. Word-
where the components p- denotes negative, p0 denotes guide Attention-based LSTM and Multi-Level LSTM. A
neutral, and p? component denotes the positivity of the senti-emoji embedding is obtained by computing the
emoji. weighted average of bi-sense emoji embeddings using a
s the sentiment score is the mean of the probability self-select attention mechanism.
distribution of emojis in tweets with label ’c’ estimated In the Word-guide attention method, each context word
from the number of occurrences N, evaluated as. guides the attention weights so that the model can self-
X select emoji embedding sense it should attend on.
s¼ pc c; In the Multi-level Attention-based LSTM the modified
senti-emoji embedding vector is at the sentence (tweet)
The Emoji Sentiment Ranking lexicon of 751 frequent
level. So instead of the word, the attention function utilizes
emojis from 1.6 million annotated tweets is created. These
the final state vector of the last LSTM unit.
emojis are ranked according to their sentiment score s and
VADER is used for the data annotation process that
can be used by various machine classification algorithms
generates weak sentiment labels. Tweets with strong sen-
for sentiment classification.
timent signals are kept while the tweets with scores in the
This approach presents a publicly available lexicon for
range (-0.6, 0.6) are discarded. Tweets with absolute sen-
annotating the text that occurs with emojis.
timent scores in (0.60, 0.70) are manually labeled by two
Although the emoji polarity is determined based on the
graduate students. If the two annotations disagree the
frequency occurrence of emoji in tweets across different
Tweets are discarded and so are neutral tweets. They also
sentiment polarities, the sentiment score (s) of an emoji
validate the information that emojis occur in tweets that
may not be sufficient to determine the overall polarity of a
contain contradictory information to the emoji sentiment.

123
268 J. Inst. Eng. India Ser. B (February 2022) 103(1):259–272

Both Word Attention and Multi-Level Attention schemes category improved the bi-polarity of sentiment prediction.
achieved superior overall performance outperforming But the model detected sarcasm and irony with precision at
baseline T-LSTM and E-LSTM models. 65% and recall at 72.22%.
Now it remains to be studied whether the discarded
neutral tweets were discarded because they were actually DEEPMOJI
neutral or contained contradictory information. If the
tweets were discarded because the annotations disagreed or In [46] Felbo et al. train the DEEPMOJI model to learn rich
because they contained contradictory sentiments, those representations with emotional labels over a dataset of
tweets need to be studied for sarcasm or irony. 1246 million tweets with 64 common emojis. The target
task is to improve sentiment classification by treating
HEMOS emojis as a proxy of emotional content. The emojis were
utilized as noisy labels for pretraining. For each emoji
In the literature [45] Li et al. proposed, HEMOS, humor irrespective of the number of occurrences in a tweet; a
detecting deep learning model. The attention-based Bi- separate tweet is saved for pretraining with the emoji as a
LSTM model was applied for four categories of classifi- label. This captures diverse emotional content and makes
cation, viz, positive, negative, optimistic humor for jokes, pre-training a single-label classification. A variant of
self-mockery, jocosity, etc., and pessimistic humor for LSTM with an embedding layer of 256 dimensions is used
sarcasm and irony on Chinese Weibo posts. The model was for word embedding. Each embedding dimension is con-
fed with on Weibo’s humorous posts with internet slangs strained within [-1,1] by using a hyperbolic tanh function.
and emojis projected into a vector representation. An Two bi-directional LSTM layers with 512 hidden units in
attention mechanism was used to extract important words each direction along with an attention mechanism to decide
that contributed to the meaning of the posts. the importance of each word for prediction are utilized. A
For an n-word sentence, an embedding matrix MnXd is single parameter per input channel is utilized:
generated by the word encoder, where d is the embedding et ¼ ht wa
size of the word embedding. M is fed as an input to the bi-
directional LSTM. The sentence-level vector is built pro- ht: word representation at time t.
gressively from word vectors using the attention structure. wa: weight matrix of attention layer
The attention mechanism extracts words with important expðet Þ
meaning to the post. at ¼ PT
i¼1 expðei Þ
The word annotation with hidden state ht of the bi-di-
rectional LSTM is fed to one-layer MLP to learn ut the at: attention importance scores of each time step
hidden representation. The word’s importance for word- X
T
level context vector uw is measured as the similarity of ut. v¼ ai hi
i¼1
ut ¼ tanhðW w ht þ bw Þ
v: representation vector of the text.
The softmax function is used to obtain normalized
v is passed to the softmax layer for the final classifica-
importance weight at.
  tion task.
exp u> t uw A chain-thaw transfer learning approach is proposed,
at ¼  
Rt exp u> t uw
that sequentially un-freezes and fine-tunes a layer at a time
is used for fine-tuning the model for the target tasks like
The sentence vector c is computed as the weighted sum emotion, sarcasm, and sentiment detection.
of the word annotations. They show that the model performs better in general
X
c¼ at ht when emojis are used in pre-training and the diversity of
t the emoji set impacts the performance of the model with
the performance dipping with the reduced emoji set.
uw, the context vector is randomly initialized. The
classification probabilities y’ are given as
0
Emoji Based Sentiment Classification of Arabic
y ¼ softmaxðW c c þ bc Þ Tweets with Deep RNNs
where the weight Wc and the bias bc are computed during
Al-Azani and El-Alfy in [47] explored recurrent structures
the training process.
LSTM and GRU (gated recurrent unit) networks to detect
The results showed that slang and emoji lexicons rec-
sentiment polarity of Arabic microblogs. They take into
ognized humor and the optimistic and pessimistic humor

123
J. Inst. Eng. India Ser. B (February 2022) 103(1):259–272 269

account the Emoji Sentiment Ranking [41] by Novak et al. also understand fine-grained emotions of humor and
for feature extraction. Of the 751 emojis, 429 emojis sarcasm.
appeared in the Arabic dataset and the ones that don’t
appear in the dataset are ignored. The Principal Component
analysis is then applied to reduce the dimensionality to 100 Challenges
features. Different deep learning architectures LSTM and
GRU in both unidirectional and bi-directional forms are The research work discussed in this paper shows that using
applied to detect sentiment polarity of the Arabic micro- emojis in the sentiment classification task can help improve
blog. The bi-directional GRU followed by the by-direc- the performance of the classifiers but this approach is not
tional LSTM was the most effective combination for without its challenges. Here are a few challenges that need
sentiment classification. to be addressed.
1. Emoji representations are different across platforms
and new emojis get introduced to the Unicode standard
Temporal Evolution of Emoji-Based Sentiment
from time to time [48], which makes it difficult to
Analysis
understand their meaning.
2. The use of emojis can vary across cultures, social
The study of emojis in sentiment analysis has gained
context, and author preference, thus the original intent
traction over the past five years. The evolution of emoji-
may be lost in interpretation [48].
based sentiment analysis has been summarized in Table 3.
3. Emojis can enhance or reverse the polarity of a tweet
From the year 2015 to 2017 we see a focus on creating
or sentence sentiment. Thus, the tweet sentiment and
emoji sentiment lexicons and learning emoji embeddings to
the sentiment of the emoji need to be evaluated
improve sentiment classification tasks by incorporating the
separately as well as how the overall sentiment
emojis. Years 2017 to 2020 the interest has grown to
changes when the text in question and the emoji are
incorporate emojis in deep learning networks like LSTM
used together.
and RNNs to not only improve sentiment classification but

Table 3 Emoji-Based Sentiment Analysis – Temporal Evolution

Year Work Approach Authors

2015 Sentiment of Emojis Created emoji sentiment lexicon by computing the sentiment score Novak
based on the occurrences of an emoji with positive, neutral, or et al
negative tweets. [41]
2016 Emoji2Vec Learned emoji embeddings from description [35] Eisner
et al
Emojinet Machine-readable emoji inventory based on emoji senses [38] Wijeratne
et al
What does emoji mean Skip gram model for emoji embeddings [29] Barbeiri
et al
2017 Automatic construction of emoji sentiment lexicon Emoji lexicon creation based on the frequency of occurrence of Kimura
emoji with a sentiment word in the Wordnet-Affect [39] and
Katsurai
Measure of emoji similarity Comparison of vector representations of emojis to determine their Wijeratne
similarity [37] et al
Using millions of emoji occurrence to learn domain Used emojis as sentiment labels to learn word embeddings using bi- Felbo et al
representations for emotion, sentiment and sarcasm directional LSTMs to improve sentiment classification [46]
detection
2018 Twitter sentiment analysis with emojis Bi-sense emoji embeddings and attention-based LSTM [43] Chen et al
Emoji based sentiment classification of Arabic Deep Recurrent Neural Networks [47] Al-Azani
microblogs and El-
Alfy
2020 HEMOS Emoji representation are used for pretraining and then Bi-LSTM, Li et al
attention-based mechanism is fed word with embeddings to detect
emotion, sarcasm, and sentiment. [45]

123
270 J. Inst. Eng. India Ser. B (February 2022) 103(1):259–272

Fig. 3 5 Year Timeline of Emoji Powered Sentiment Analysis

4. To our knowledge, there is no benchmark dataset for Funding Not Applicable.

tweets and emojis for sarcasm and irony detection,
Declarations
which makes the task of identifying complex emotions
difficult. Conflict of Interest The authors declare that they have no conflict of
interest.

Conclusion and Future Work References

A five-year timeline in Fig. 3 shows that interest and 1. W. Medhat, A. Hassan, H. Korashy, Sentiment analysis algo-
research in exploiting emojis in sentiment analysis have rithms and applications: a survey. Ain Shams Eng. J. 5(4),
1093–1113 (2014)
increased tremendously in the past five years.
2. Rachid, B.A., Azza, H. and Henda, B.G., 2018. Sentiment
The above-discussed studies have shown that the emojis Analysis Approaches based on Granularity Levels.
in sentiment analysis can be used to enhance the word 3. K. Elshakankery, M.F. Ahmed, HILATSA: A hybrid Incremental
embeddings and vector space as well as create rich senti- learning approach for Arabic tweets sentiment analysis. Egyptian
Informatics Journal 20(3), 163–171 (2019)
ment lexicons. The use of emojis along with text improves
4. Mikolov, T., Chen, K., Corrado, G. and Dean, J., 2013. Efficient
the classifier performance. Also, the various deep learning estimation of word representations in vector space. arXiv preprint
models are now becoming popular tools for the sentiment http://arxiv.org/abs/1301.3781.
classification task. 5. Tang, D., Wei, F., Yang, N., Zhou, M., Liu, T. and Qin, B., 2014,
June. Learning sentiment-specific word embedding for twitter
But it is yet to be seen how emojis can help in identi-
sentiment classification. In Proceedings of the 52nd Annual
fying complex and nuanced emotions like those of sarcasm, Meeting of the Association for Computational Linguistics (Vol-
irony, dark humor, etc. and if the classification perfor- ume 1: Long Papers) (pp. 1555–1565).
mance can be improved further. 6. McCormick, C., 2016. Word2vec tutorial-the skip-gram model.
7. Komninos, A. and Manandhar, S., 2016, June. Dependency based
In the future, it can be studied to incorporate multiple
embeddings for sentence classification tasks. In Proceedings of
sense of emojis co-occurring with text and to perform the 2016 conference of the North American chapter of the asso-
multi-class classification instead of just positive and neg- ciation for computational linguistics: human language tech-
ative classes to uncover the underlying complex emotions nologies (pp. 1490–1500).
8. B. Bansal, S. Srivastava, Sentiment classification of online con-
like sarcasm, irony, frustration, etc. in the text.
sumer reviews using word vector representations. Procedia
Computer Sci. 132, 1147–1153 (2018)

123
J. Inst. Eng. India Ser. B (February 2022) 103(1):259–272 271

9. Pennington, J., Socher, R. and Manning, C.D., 2014, October. 28. Barbieri, F., Kruszewski, G., Ronzano, F. and Saggion, H., 2016,
Glove: Global vectors for word representation. In Proceedings of October. How cosmopolitan are emojis? Exploring emojis usage
the 2014 conference on empirical methods in natural language and meaning over different languages with distributional
processing (EMNLP) (pp. 1532–1543). semantics. In Proceedings of the 24th ACM international con-
10. He, R., Lee, W.S., Ng, H.T. and Dahlmeier, D., 2018. Exploiting ference on Multimedia (pp. 531–535).
document knowledge for aspect-level sentiment classifica- 29. Barbieri, F., Ronzano, F. and Saggion, H., 2016. What does this
tion. arXiv preprint http://arxiv.org/abs/1806.04346. emoji mean? a vector space skip-gram model for twitter emojis.
11. Collobert, R., Weston, J., Bottou, L., Karlen, M., Kavukcuoglu, In Calzolari N, Choukri K, Declerck T, et al, editors. Proceedings
K. and Kuksa, P., 2011. Natural language processing (almost) of the Tenth International Conference on Language Resources
from scratch. Journal of machine learning research, 12(- and Evaluation (LREC 2016); 2016 May 23–28; Portorozˇ,
ARTICLE), pp.2493–2537. Slovenia. Paris: European Language Resources Association
12. S. Xiong, H. Lv, W. Zhao, D. Ji, Towards Twitter sentiment (ELRA); 2016. p. 3967–72.. ELRA (European Language
classification by multi-level sentiment-enriched word embed- Resources Association).
dings. Neurocomputing 275, 2459–2466 (2018) 30. Barbieri, F., Espinosa-Anke, L. and Saggion, H., 2016. Revealing
13. S.K. Bharti, B. Vachha, R.K. Pradhan, K.S. Babu, S.K. Jena, patterns of Twitter emoji usage in Barcelona and Madrid. Fron-
Sarcastic sentiment detection in tweets streamed in real time: a tiers in Artificial Intelligence and Applications. 2016;(Artificial
big data approach. Digital Commun. Netw. 2(3), 108–121 (2016) Intelligence Research and Development) 288: 239–44.
14. https://aylien.com/blog/overview-word-embeddings-history-word 31. Barbieri, F., Ballesteros, M. and Saggion, H., 2017. Are emojis
2vec-cbow-glove predictable?. arXiv preprint http://arxiv.org/abs/1702.07285.
15. Esuli, A. and Sebastiani, F., 2006, May. Sentiwordnet: A publicly 32. Barbieri, F., Marujo, L., Karuturi, P., Brendel, W. and Saggion,
available lexical resource for opinion mining. In LREC (Vol. 6, H., 2018. Exploring emoji usage and prediction through a tem-
pp. 417–422). poral variation lens. arXiv preprint
16. C. Fellbaum, WordNet. in Theory and Applications of Ontology: http://arxiv.org/abs/1805.00731.
Computer Applications (Springer, 2010). 33. Barbieri, F., Camacho-Collados, J., Ronzano, F., Anke, L.E.,
17. Gilbert, C.H.E. and Hutto, E., 2014, June. Vader: A parsimonious Ballesteros, M., Basile, V., Patti, V. and Saggion, H., 2018, June.
rule-based model for sentiment analysis of social media text. Semeval 2018 task 2: Multilingual emoji prediction. In Pro-
In Eighth International Conference on Weblogs and Social ceedings of The 12th International Workshop on Semantic
Media (ICWSM-14). Available at (20/04/16) . social. gatech. Evaluation (pp. 24–33).
edu/papers/icwsm14. vader. hutto. pdf (Vol. 81, p. 82). 34. Barbieri, F., Anke, L.E., Camacho-Collados, J., Schockaert, S.
18. Pennebaker, J.W., Francis, M.E. and Booth, R.J., 2001. Linguistic and Saggion, H., 2018. Interpretable emoji prediction via label-
inquiry and word count: LIWC 2001. Mahway: Lawrence Erl- wise attention LSTMs. In Proceedings of the 2018 Conference on
baum Associates, 71(2001), p.2001. Empirical Methods in Natural Language Processing (pp.
19. Bradley, M.M. and Lang, P.J., 1999. Affective norms for English 4766–4771).
words (ANEW): Instruction manual and affective ratings (Vol. 35. Eisner, B., Rocktäschel, T., Augenstein, I., Bošnjak, M. and
30, No. 1, pp. 25–36). Technical report C-1, the center for Riedel, S., 2016. emoji2vec: Learning emoji representations from
research in psychophysiology, University of Florida. their description. arXiv preprint http://arxiv.org/abs/1609.08359.
20. Stone, P.J. and Hunt, E.B., 1963, May. A computer approach to 36. http://www.unicode.org/emoji/charts/ full-emoji-list.html
content analysis: studies using the general inquirer system. 37. Wijeratne, S., Balasuriya, L., Sheth, A. and Doran, D., 2017,
In Proceedings of the May 21–23, 1963, spring joint computer August. A semantics-based measure of emoji similarity. In Pro-
conference (pp. 241–256). ceedings of the International Conference on Web Intelli-
21. Lu, Y., Castellanos, M., Dayal, U. and Zhai, C., 2011, March. gence (pp. 646–653).
Automatic construction of a context-aware sentiment lexicon: an 38. Wijeratne, S., Balasuriya, L., Sheth, A. and Doran, D., 2016,
optimization approach. In Proceedings of the 20th international November. Emojinet: Building a machine readable sense inven-
conference on World wide web (pp. 347–356). tory for emoji. In International conference on social informat-
22. Hassan, A. and Radev, D., 2010, July. Identifying text polarity ics (pp. 527–541). Springer, Cham.
using random walks. In Proceedings of the 48th Annual Meeting 39. Kimura, M. and Katsurai, M., 2017, July. Automatic construction
of the Association for Computational Linguistics (pp. 395–403). of an emoji sentiment lexicon. In Proceedings of the 2017 ieee/
23. Hatzivassiloglou, V. and McKeown, K., 1997, July. Predicting acm international conference on advances in social networks
the semantic orientation of adjectives. In 35th annual meeting of analysis and mining 2017 (pp. 1033–1036).
the association for computational linguistics and 8th conference 40. Strapparava, C. and Valitutti, A., 2004, May. Wordnet affect: an
of the european chapter of the association for computational affective extension of wordnet. In Lrec (Vol. 4, No. 1083–1086,
linguistics (pp. 174–181). p. 40).
24. Lu, Y., Zhai, C. and Sundaresan, N., 2009, April. Rated aspect 41. P.K. Novak, J. Smailović, B. Sluban, I. Mozetič, Sentiment of
summarization of short comments. In Proceedings of the 18th emojis. PloS one 10(12), 1035 (2015)
international conference on World wide web (pp. 131–140). 42. Emojipedia: https://emojipedia.org
25. https://www.unicode.org/reports/tr51/tr51-18.html 43. Chen, Y., Yuan, J., You, Q. and Luo, J., 2018, October. Twitter
26. A. Konrad, S.C. Herring, D. Choi, Sticker and emoji use in sentiment analysis via bi-sense emoji embedding and attention-
facebook messenger: implications for graphicon change. J. Com- based LSTM. In Proceedings of the 26th ACM international
put.-Mediat. Commun. 25(3), 217–235 (2020) conference on Multimedia (pp. 117–125).
27. Singh, A., Blanco, E. and Jin, W., 2019, June. Incorporating 44. https://github.com/facebookresearch/fastText
emoji descriptions improves tweet classification. In Proceedings 45. D. Li, R. Rzepka, M. Ptaszynski, K. Araki, HEMOS: A novel
of the 2019 Conference of the North American Chapter of the deep learning-based fine-grained humor detecting method for
Association for Computational Linguistics: Human Language sentiment analysis of social media. Inf. Process. Manage. 57(6),
Technologies, Volume 1 (Long and Short Papers) (pp. 102290 (2020)
2096–2101). 46. Felbo, B., Mislove, A., Søgaard, A., Rahwan, I. and Lehmann, S.,
2017. Using millions of emoji occurrences to learn any-domain

123
272 J. Inst. Eng. India Ser. B (February 2022) 103(1):259–272

representations for detecting sentiment, emotion and sar- lexica from unsupervised sentiment analysis of their descriptions.
casm. arXiv preprint http://arxiv.org/abs/1708.00524. Expert Syst. Appl. 103, 74–91 (2018)
47. Al-Azani, S. and El-Alfy, E.S., 2018, March. Emojis-based sen-
timent classification of Arabic microblogs using deep recurrent Publisher’s Note Springer Nature remains neutral with regard to
neural networks. In 2018 International Conference on Computing jurisdictional claims in published maps and institutional affiliations.
Sciences and Engineering (ICCSE) (pp. 1–6). IEEE.
48. M. Fernández-Gavilanes, J. Juncal-Martı́nez, S. Garcı́a-Méndez,
E. Costa-Montenegro, F.J. González-Castaño, Creating emoji

123