Social Network Analysis and Mining (2021) 11:81

https://doi.org/10.1007/s13278-021-00776-6

REVIEW PAPER

A review on sentiment analysis and emotion detection from text

Pansy Nandwani1 · Rupali Verma1

Received: 6 April 2021 / Revised: 25 June 2021 / Accepted: 10 July 2021

© The Author(s), under exclusive licence to Springer-Verlag GmbH Austria, part of Springer Nature 2021

Abstract
Social networking platforms have become an essential means for communicating feelings to the entire world due to rapid
expansion in the Internet era. Several people use textual content, pictures, audio, and video to express their feelings or
viewpoints. Text communication via Web-based networking media, on the other hand, is somewhat overwhelming. Every
second, a massive amount of unstructured data is generated on the Internet due to social media platforms. The data must be
processed as rapidly as generated to comprehend human psychology, and it can be accomplished using sentiment analysis,
which recognizes polarity in texts. It assesses whether the author has a negative, positive, or neutral attitude toward an item,
administration, individual, or location. In some applications, sentiment analysis is insufficient and hence requires emotion
detection, which determines an individual’s emotional/mental state precisely. This review paper provides understanding into
levels of sentiment analysis, various emotion models, and the process of sentiment analysis and emotion detection from text.
Finally, this paper discusses the challenges faced during sentiment and emotion analysis.

Keywords Affective computing · Natural language processing · Opinion mining · Pre-processing · Word embedding

1 Introduction depressed. “Emotion detection,” “affective computing,”

“emotion analysis,” and “emotion identification” are all
Human language understanding and human language gen- phrases that are sometimes used interchangeably (Munezero
eration are the two aspects of natural language processing et al. 2014). People are using social media to communicate
(NLP). The former, however, is more difficult due to ambi- their feelings since Internet services have improved. On
guities in natural language. However, the former is more social media, people freely express their feelings, arguments,
challenging due to ambiguities present in natural language. opinions on wide range of topics. In addition, many users
Speech recognition, document summarization, question give feedbacks and reviews various products and services
answering, speech synthesis, machine translation, and other on various e-commerce sites. User's ratings and reviews on
applications all employ NLP (Itani et al. 2017). The two multiple platforms encourage vendors and service provid-
critical areas of natural language processing are sentiment ers to enhance their current systems, goods, or services.
analysis and emotion recognition. Even though these two Today almost every industry or company is undergoing
names are sometimes used interchangeably, they differ in a some digital transition, resulting in vast amounts of struc-
few respects. Sentiment analysis is a means of assessing if tured and unstructured increase data. The enormous task for
data is positive, negative, or neutral. companies is to transform unstructured data into meaning-
In contrast, Emotion detection is a means of identifying ful insights that can help them in decision-making (Ahmad
distinct human emotion types such as furious, cheerful, or et al. 2020)
For instance, in the business world, vendors use social
media platforms such as Instagram, YouTube, Twitter, and
* Pansy Nandwani Facebook to broadcast information about their product and
[email protected]; efficiently collect client feedback (Agbehadji and Ijabad-
[email protected]
eniyi 2021). People’s active feedback is valuable not only
* Rupali Verma for business marketers to measure customer satisfaction and
[email protected]
keep track of the competition but also for consumers who
1
Computer Science and Engineering Department, Punjab want to learn more about a product or service before buying
Engineering College, Chandigarh, India

13
Vol.:(0123456789)
 81 Page 2 of 19 Social Network Analysis and Mining (2021) 11:81

it. Sentiment analysis assists marketers in understanding techniques, researchers face significant challenges, including
their customer's perspectives better so that they may make dealing with context, ridicule, statements conveying several
necessary changes to their products or services (Jang et al. emotions, spreading Web slang, and lexical and syntactical
2013; Al Ajrawi et al. 2021). In both advanced and emerging ambiguity. Furthermore, because there are no standard rules
nations, the impact of business and client sentiment on stock for communicating feelings across multiple platforms, some
market performance may be witnessed. In addition, the rise express them with incredible effect, some stifle their feel-
of social media has made it easier and faster for investors to ings, and some structure their message logically. Therefore,
interact in the stock market. As a result, investor's sentiments it is a great challenge for researchers to develop a technique
impact their investment decisions which can swiftly spread that can efficiently work in all domains.
and magnify over the network, and the stock market can be In this review paper, Sect. 2, introduces sentiment analy-
altered to some extent (Ahmed 2020). As a result, sentiment sis and its various levels, emotion detection, and psycho-
and emotion analysis has changed the way we conduct busi- logical models. Section 3 discusses multiple steps involved
ness (Bhardwaj et al. 2015). in sentiment and emotion analysis, including datasets, pre-
In the healthcare sector, online social media like Twitter processing of text, feature extraction techniques, and vari-
have become essential sources of health-related informa- ous sentiment and emotion analysis approaches. Section 4
tion provided by healthcare professionals and citizens. For addresses multiple challenges faced by researchers during
example, people have been sharing their thoughts, opin- sentiment and emotion analysis. Finally, Sect. 5 concludes
ions, and feelings on the Covid-19 pandemic (Garcia and the work.
Berton 2021). Patients were directed to stay isolated from
their loved ones, which harmed their mental health. To save
patients from mental health issues like depression, health 2 Background
practitioners must use automated sentiment and emotion
analysis (Singh et al. 2021). People commonly share their 2.1 Sentiment analysis
feelings or beliefs on sites through their posts, and if some-
one seemed to be depressed, people could reach out to them Many people worldwide are now using blogs, forums, and
to help, thus averting deteriorated mental health conditions. social media sites such as Twitter and Facebook to share
Sentiment and emotion analysis plays a critical role in their opinions with the rest of the globe. Social media has
the education sector, both for teachers and students. The become one of the most effective communication media
efficacy of a teacher is decided not only by his academic cre- available. As a result, an ample amount of data is generated,
dentials but also by his enthusiasm, talent, and dedication. called big data, and sentiment analysis was introduced to
Taking timely feedback from students is the most effective analyze this big data effectively and efficiently (Nagamanjula
technique for a teacher to improve teaching approaches (San- and Pethalakshmi 2020). It has become exceptionally crucial
geetha and Prabha 2020). Open-ended textual feedback is for industry or organization to comprehend the sentiments of
difficult to observe, and it is also challenging to derive con- the user. Sentiment analysis, often known as opinion mining,
clusions manually. The findings of a sentiment analysis and is a method for detecting whether an author’s or user’s view-
emotion analysis assist teachers and organizations in taking point on a subject is positive or negative. Sentiment analysis
corrective action. Since social site's inception, educational is defined as the process of obtaining meaningful informa-
institutes are increasingly relying on social media like Face- tion and semantics from text using natural processing tech-
book and Twitter for marketing and advertising purposes. niques and determining the writer’s attitude, which might be
Students and guardians conduct considerable online research positive, negative, or neutral (Onyenwe et al. 2020). Since
and learn more about the potential institution, courses and the purpose of sentiment analysis is to determine polarity
professors. They use blogs and other discussion forums to and categorize opinionated texts as positive or negative,
interact with students who share similar interests and to dataset’s class range involved in sentiment analysis is not
assess the quality of possible colleges and universities. Thus, restricted to just positive or negative; it can be agreed or
applying sentiment and emotion analysis can help the stu- disagreed, good or bad. It can also be quantified on a 5-point
dent to select the best institute or teacher in his registration scale: strongly disagree, disagree, neutral, agree, or strongly
process (Archana Rao and Baglodi 2017). agree (Prabowo and Thelwall 2009). For instance, Ye et al.
Sentiment and emotion analysis has a wide range of (2009) applied sentiment analysis on reviews on European
applications and can be done using various methodologies. and US destinations labeled on the scale of 1 to 5. They
There are three types of sentiment and emotion analysis associated 1-star or 2-star reviews with the negative polarity
techniques: lexicon based, machine learning based, and deep and more than 2-star reviews with positive polarity. Gräbner
learning based. Each has its own set of benefits and draw- et al. (2012) built a domain-specific lexicon that consists
backs. Despite different sentiment and emotion recognition of tokens with their sentiment value. These tokens were

13
Social Network Analysis and Mining (2021) 11:81 Page 3 of 19 81

gathered from customer reviews in the tourism domain to As stated earlier, sentiment analysis and emotion analy-
classify sentiment into 5-star ratings from terrible to excel- sis are often used interchangeably by researchers. However,
lent in the tourism domain. Moreover, Sentiment analysis they differ in a few ways. In sentiment analysis, polarity
from the text can be performed at three levels discussed in is the primary concern, whereas, in emotion detection, the
the following section. Salinca (2015) applied machine learn- emotional or psychological state or mood is detected. Sen-
ing algorithms on the Yelp dataset, which contains reviews timent analysis is exceptionally subjective, whereas emo-
on service providers scaled from 1 to 5. Sentiment analysis tion detection is more objective and precise. Section 2.2
can be categorized at three levels, mentioned in the follow- describes all about emotion detection in detail.
ing section.
2.2 Emotion detection
2.1.1 Levels of sentiment analysis
Emotions are an inseparable component of human life.
Sentiment analysis is possible at three levels: sentence level,
These emotions influence human decision-making and
document level, and aspect level. At the sentence-level or
help us communicate to the world in a better way. Emo-
phrase-level sentiment analysis, documents or paragraphs
tion detection, also known as emotion recognition, is the
are broken down into sentences, and each sentence’s polar-
process of identifying a person’s various feelings or emo-
ity is identified (Meena and Prabhakar 2007; Arulmurugan
tions (for example, joy, sadness, or fury). Researchers have
et al. 2019; Shirsat et al. 2019). At the document level, the
been working hard to automate emotion recognition for the
sentiment is detected from the entire document or record
past few years. However, some physical activities such as
(Pu et al. 2019). The necessity of document-level senti-
heart rate, shivering of hands, sweating, and voice pitch also
ment analysis is to extract global sentiment from long texts
convey a person’s emotional state (Kratzwald et al. 2018),
that contain redundant local patterns and lots of noise. The
but emotion detection from text is quite hard. In addition,
most challenging aspect of document-level sentiment clas-
various ambiguities and new slang or terminologies being
sification is taking into account the link between words
introduced with each passing day make emotion detection
and phrases and the full context of semantic information
from text more challenging. Furthermore, emotion detection
to reflect document composition (Rao et al. 2018; Liu et al.
is not just restricted to identifying the primary psychological
2020a). It necessitates a deeper understanding of the intri-
conditions (happy, sad, anger); instead, it tends to reach up
cate internal structure of sentiments and dependent words
to 6-scale or 8-scale depending on the emotion model.
(Liu et al. 2020b). At the aspect level, sentiment analysis,
opinion about a specific aspect or feature is determined. For
instance, the speed of the processor is high, but this prod- 2.2.1 Emotion models/emotion theories
uct is overpriced. Here, speed and cost are two aspects or
viewpoints. Speed is mentioned in the sentence hence called In English, the word 'emotion' came into existence in the
explicit aspect, whereas cost is an implicit aspect. Aspect- seventeenth century, derived from the French word 'emo-
level sentiment analysis is a bit harder than the other two as tion, meaning a physical disturbance. Before the nineteenth
implicit features are hard to identify. Devi Sri Nandhini and century, passion, appetite, and affections were categorized
Pradeep (2020) proposed an algorithm to extract implicit as mental states. In the nineteenth century, the word 'emo-
aspects from documents based on the frequency of co-occur- tion' was considered a psychological term (Dixon 2012). In
rence of aspect with feature indicator and by exploiting the psychology, complex states of feeling lead to a change in
relation between opinionated words and explicit aspects. thoughts, actions, behavior, and personality referred to as
Ma et al. (2019) took care of the two issues concerning emotions. Broadly, psychological or emotion models are
aspect-level analysis: various aspects in a single sentence classified into two categories: dimensional and categorical.
having different polarities and explicit position of context Dimensional Emotion model This model represents emo-
in an opinionated sentence. The authors built up a two-stage tions based on three parameters: valence, arousal, and power
model based on LSTM with an attention mechanism to solve (Bakker et al. 2014. Valence means polarity, and arousal
these issues. They proposed this model based on the assump- means how exciting a feeling is. For example, delighted
tion that context words near to aspect are more relevant and is more exciting than happy. Power or dominance signi-
need greater attention than farther context words. At stage fies restriction over emotion. These parameters decide the
one, the model exploits multiple aspects in a sentence one position of psychological states in 2-dimensional space, as
by one with a position attention mechanism. Then, at the illustrated in Fig. 1.
second state, it identifies (aspect, sentence) pairs according Categorical Emotion model
to the position of aspect and context around it and calculates In the categorical model, emotions are defined discretely,
the polarity of each team simultaneously. such as anger, happiness, sadness, and fear. Depending upon

13
 81 Page 4 of 19 Social Network Analysis and Mining (2021) 11:81

Fig. 1  Dimensional model of

emotions

the particular categorical model, emotions are categorized most commonly used emotion states in different models
into four, six, or eight categories. include anger, fear, joy, surprise, and disgust, as depicted in
Table 1 demonstrates numerous emotion models that the figure above. It can be seen from the figure that emotions
are dimensional and categorical. In the realm of emotion on two sides of the axis will not always be opposite of each
detection, most researchers adopted Ekman and Plutchik’s other. For example, sadness and joy are opposites, but anger
emotion model. The emotional states defined by the models is not the opposite of fear.
make up the set of labels used to annotate the sentences or
documents. Batbaatar et al. (2019), Becker et al. (2017),
Jain et al. (2017) adopted Ekman’s six basic emotions. Sai- 3 Process of sentiment analysis
lunaz and Alhajj (2019) used Ekman models for annotating and emotion detection
tweets. Some researchers used customized emotion models
by extending the model with one or two additional states. Process of sentiment analysis and emotion detection comes
Roberts et al. (2012) used the Ekman model to annotate the across various stages like collecting dataset, pre-processing,
tweets with the 'love' state. Ahmad et al. (2020) adopted the feature extraction, model development, and evaluation, as
wheel of emotion modeled by Plutchik for labeling Hindi shown in Fig. 3.
sentences with nine different Plutchik model states, decreas-
ing semantic confusion, among other words. Plutchik and 3.1 Datasets for sentiment analysis and emotion
Ekman’s model's states are also utilized in various hand- detection
crafted lexicons like WordNet-Affect (Strapparava et al.
2004) and NRC (Mohammad and Turney 2013) word–emo- Table 2 lists numerous sentiment and emotion analysis
tion lexicons. Laubert and Parlamis (2019) referred to the datasets that researchers have used to assess the effec-
Shaver model because of its three-level hierarchy structure tiveness of their models. The most common datasets are
of emotions. Valence or polarity is presented at the first SemEval, Stanford sentiment treebank (SST), international
level, followed by the second level consisting of five emo- survey of emotional antecedents and reactions (ISEAR)
tions, and the third level shows discrete 24 emotion states. in the field of sentiment and emotion analysis. SemE-
Some researchers did not refer to any model and classified val and SST datasets have various variants which differ
the dataset into three basic feelings: happy, sad, or angry. in terms of domain, size, etc. ISEAR was collected from
Figure 2 depicts the numerous emotional states that can multiple respondents who felt one of the seven emotions
be found in various models. These states are plotted on a (mentioned in the table) in some situations. The table
four-axis by taking the Plutchik model as a base model. The shows that datasets include mainly the tweets, reviews,

13
 Table 1  Emotion models defined by various psychologists
Emotion model Type of model No. of states Psychological states Representations Discussion
Social Network Analysis and Mining

Ekman model (Ekman 1992) Categorical 6 Anger, disgust, fear, joy, sadness, surprise – Ekman’s model consisted of six emotions,
which act as a base for other emotion models
like Plutchik model
Plutchik Wheel of Emotions (Plutchik 1982) Dimensional – Joy, pensiveness, ecstasy, acceptance, sadness, Wheel Plutchik considered two types of emotions:
fear, interest, rage, admiration, amazement, basic (Ekman model + Trust +Anticipation)
anger, vigilance boredom, annoyance, sub- and mixed emotions (made from the combina-
(2021) 11:81

mission, serenity, apprehension, contempt, tion of basic emotions). Plutchik represented

surprise, disapproval, distraction, grief, emotions on a colored wheel
loathing, love, optimism, aggressiveness,
remorse, anticipation, awe, terror, trust,
disgust
Izard model (Izard 1992) – 10 Anger, contempt, disgust, anxiety, fear, guilt, – –
interest, joy, shame, surprise
Shaver model (Shaver et al. 1987) Categorical 6 Sadness, joy, anger, fear, love, surprise Tree Shaver represented the primary, secondary
and tertiary emotions in a hierarchal manner.
The top-level of the tree presents these six
emotions
Russell’s circumplex model (Russell 1980) Dimensional – Sad, satisfied, Afraid, alarmed, frustrated, – Emotions are presented over the circumplex
angry, happy, gloomy, annoyed, tired, model
relaxed, glad, aroused, astonished, at ease,
tense, miserable, content, bored, calm,
delighted, excited, depressed, distressed,
serene, droopy, pleased, sleepy
Tomkins model (Tomkins and McCarter Categorical 9 Disgust, surprise-Startle, anger-rage, anxiety, – Tomkins identified nine different emotions out
1964) fear-terror, contempt, joy, shame, interest- of which six emotions are negative. Most of
Excitement the emotions are defined as a pair
Lövheim Model (Lövheim 2012) Dimensional – Anger, contempt, distress, enjoyment, terror, Cube Lövheim arranged the emotions according to
excitement, humiliation, startle the amount of three substances (Noradrena-
line, dopamine and Serotonin) on a 3-D cube
Page 5 of 19
81

13
 81 Page 6 of 19 Social Network Analysis and Mining (2021) 11:81

Fig. 2  Illustration of various emotional models with some psychological states

Fig. 3  Basic steps to perform

sentiment analysis and emotion
detection

feedbacks, stories, etc. A dimensional model named 3.2 Pre‑processing of text

valence, arousal dominance model (VAD) is used in the
EmoBank dataset collected from news, blogs, letters, etc. On social media, people usually communicate their feel-
Many studies have acquired data from social media sites ings and emotions in effortless ways. As a result, the data
such as Twitter, YouTube, and Facebook and had it labeled obtained from these social media platform's posts, audits,
by language and psychology experts in the literature. Data comments, remarks, and criticisms are highly unstruc-
crawled from various social media platform's posts, blogs, tured, making sentiment and emotion analysis difficult
e-commerce sites are usually unstructured and thus need for machines. As a result, pre-processing is a critical
to be processed to make it structured to reduce some addi- stage in data cleaning since the data quality significantly
tional computations outlined in the following section. impacts many approaches that follow pre-processing.

13
 Table 2  Datasets for sentiment analysis and emotion detection
Dataset Data size Sentiment/emotion analysis Sentiments/emotions Range Domain
Social Network Analysis and Mining

Stanford Sentiment Treebank 118,55 reviews in SST-1 Sentiment analysis Very positive, positive, negative, 5 Movie reviews
(Chen et al. 2017) very negative and neutral.
9613 reviews in SST-2 Sentiment analysis Positive and negative 2 Movie reviews
SemEval Tasks (Ma et al. 2019; SemEval- 2014 (Task 4): 5936 Sentiment analysis Positive, negative and neutral 3 Laptop and Restaurant reviews
Ahmad et al. 2020) reviews for training and 1758
reviews for testing
(2021) 11:81

SemEval-2018 (Affects in dataset Emotion analysis Anger, Joy, sad and fear 4 Tweets
Task): 7102 tweets in Emotion
and Intensity for ordinal clas-
sification (EI-oc)
Thai fairy tales (Pasupa and Ayut- 1964 sentences Sentiment analysis Positive, negative and neutral 3 Children tales
thaya 2019)
SS-Tweet (Symeonidis et al. 4242 Sentiment Analysis Positive strength and Negative 1 to 5 for positive Tweets
2018) strength and −1 to −5 for
negative
EmoBank (Buechel and Hahn 10,548 Emotion analysis Valence, Arousal Dominance – News, blogs, fictions, letters etc.
2017) model (VAD)
International Survey of Emotional Around 7500 sentences Emotion analysis Guilt, Joy, Shame, Fear, sadness, 7 Incident reports.
Antecedents and Reactions disgust
(ISEAR) (Seal et al. 2020)
Alm gold standard data set 1207 sentences Emotion analysis happy, fearful, sad, surprised and 5 Fairy tales
(Agrawal and An 2012) angry-disgust(combined)
EmoTex (Hasan et al. 2014) 134,100 sentences Emotion analysis Circumplex model – Twitter
Text Affect (Chaffar and Inkpen 1250 sentences Emotion analysis Ekman 6 Google news
2011)
Neviarouskaya Dataset Dataset 1: 1000 sentences and Emotion analysis Izard 10 Stories and blogs
(Alswaidan and Menai 2020) Dataset 2: 700 sentences
Aman’s dataset (Hosseini 2017) 1890 sentences Emotion analysis Ekman with neutral class 7 Blogs
Page 7 of 19
81

13
 81 Page 8 of 19 Social Network Analysis and Mining (2021) 11:81

The organization of a dataset necessitates pre-processing, dictionary, and the values present in the cells of the feature
including tokenization, stop word removal, POS tagging, map generally signify the count of the word in the sentence
etc. (Abdi et al. 2019; Bhaskar et al. 2015). Some of these or document. To carry out feature extraction, one of the most
pre-processing techniques can result in the loss of crucial straightforward methods used is 'Bag of Words' (BOW), in
information for sentiment and emotion analysis, which must which a fixed-length vector of the count is defined where
be addressed. each entry corresponds to a word in a pre-defined dictionary
Tokenization is the process of breaking down either of words. The word in a sentence is assigned a count of 0
the whole document or paragraph or just one sentence if it is not present in the pre-defined dictionary, otherwise a
into chunks of words called tokens (Nagarajan and Gan- count of greater than or equal to 1 depending on how many
dhi 2019). For instance, consider the sentence “this place times it appears in the sentence. That is why the length of
is so beautiful” and post-tokenization, it will become 'this,' the vector is always equal to the words present in the diction-
"place," is, "so," beautiful.’ It is essential to normalize the ary. The advantage of this technique is its easy implemen-
text for achieving uniformity in data by converting the text tation but has significant drawbacks as it leads to a sparse
into standard form, correcting the spelling of words, etc. matrix, loses the order of words in the sentence, and does not
(Ahuja et al. 2019). capture the meaning of a sentence (Bandhakavi et al. 2017;
Unnecessary words like articles and some prepositions Abdi et al. 2019). For example, to represent the text “are you
that do not contribute toward emotion recognition and sen- enjoying reading” from the pre-defined dictionary I, Hope,
timent analysis must be removed. For instance, stop words you, are, enjoying, reading would be (0,0,1,1,1,1). However,
like "is," "at," "an," "the" have nothing to do with senti- these representations can be improved by pre-processing of
ments, so these need to be removed to avoid unnecessary text and by utilizing n-gram, TF-IDF.
computations (Bhaskar et al. 2015; Abdi et al. 2019). POS The N-gram method is an excellent option to resolve
tagging is the way to identify different parts of speech in a the order of words in sentence vector representation. In an
sentence. This step is beneficial in finding various aspects n-gram vector representation, the text is represented as a
from a sentence that are generally described by nouns or collaboration of unique n-gram means groups of n adjacent
noun phrases while sentiments and emotions are conveyed terms or words. The value of n can be any natural number.
by adjectives (Sun et al. 2017). For example, consider the sentence “to teach is to touch a
Stemming and lemmatization are two crucial steps of pre- life forever” and n = 3 called trigram will generate 'to teach
processing. In stemming, words are converted to their root is,' 'teach is to,' 'is to touch,' 'to touch a,' 'touch a life,' 'a life
form by truncating suffixes. For example, the terms "argued" forever.' In this way, the order of the sentence can be main-
and "argue" become "argue." This process reduces the tained (Ahuja et al. 2019). N-grams features perform better
unwanted computation of sentences (Kratzwald et al. 2018; than the BOW approach as they cover syntactic patterns,
Akilandeswari and Jothi 2018). Lemmatization involves including critical information (Chaffar and Inkpen 2011).
morphological analysis to remove inflectional endings from However, though n-gram maintains the order of words, it
a token to turn it into the base word lemma (Ghanbari-Adivi has high dimensionality and data sparsity (Le and Mikolov
and Mosleh 2019). For instance, the term "caught" is con- 2014).
verted into "catch" (Ahuja et al. 2019). Symeonidis et al. Term frequency-inverse document frequency, usu-
(2018) examined the performance of four machine learning ally abbreviated as TFIDF, is another method commonly
models with a combination and ablation study of various used for feature extraction. This method represents text
pre-processing techniques on two datasets, namely SS-Tweet in matrix form, where each number quantifies how much
and SemEval. The authors concluded that removing numbers information these terms carry in a given document. It is
and lemmatization enhanced accuracy, whereas removing built on the premise that rare terms have much information
punctuation did not affect accuracy. in the text document (Liu et al. 2019). Term frequency
is the number of times a word w appears in a document
3.3 Feature extraction divided by the total number of words W in the document,
and IDF is log (total number of documents (N) divided by
The machine understands text in terms of numbers. The the total number of documents in which word w appears
process of converting or mapping the text or words to real- (n)) (Songbo and Jin 2008). Ahuja et al. (2019) imple-
valued vectors is called word vectorization or word embed- mented six pre-processing techniques and compared two
ding. It is a feature extraction technique wherein a document feature extraction techniques to identify the best approach.
is broken down into sentences that are further broken into They applied six machine learning algorithms and used
words; after that, the feature map or matrix is built. In the n-grams with n = 2 and TF-IDF for feature extraction over
resulting matrix, each row represents a sentence or docu- the SS-tweet dataset and concluded TF-IDF gives better
ment while each feature column represents a word in the performance over n-gram.

13
Social Network Analysis and Mining (2021) 11:81 Page 9 of 19 81

The availability of vast volumes of data allows a deep 3.4 Techniques for sentiment analysis and emotion
learning network to discover good vector representations. detection
Feature extraction with word embedding based on neural
networks is more informative. In neural network-based Figure 4 presents various techniques for sentiment analysis
word embedding, the words with the same semantics or and emotion detection which are broadly classified into a
those related to each other are represented by similar vec- lexicon-based approach, machine learning-based approach,
tors. This is more popular in word prediction as it retains deep learning-based approach. The hybrid approach is a
the semantics of words. Google’s research team, headed by combination of statistical and machine learning approaches
Tomas Mikolov, developed a model named Word2Vec for to overcome the drawbacks of both approaches. Transfer
word embedding. With Word2Vec, it is possible to under- learning is also a subset of machine learning which allows
stand for a machine that “queen” + “female” + “male” the use of the pre-trained model in other similar domain.
vector representation would be the same as a vector rep-
resentation of “king” (Souma et al. 2019). 3.4.1 Sentiment analysis techniques
Other examples of deep learning-based word embed-
ding models include GloVe, developed by researchers at Lexicon-based approach This method maintains a word dic-
Stanford University, and FastText, introduced by Face- tionary in which each positive and negative word is assigned
book. GloVe vectors are faster to train than Word2vec. a sentiment value. Then, the sum or mean of sentiment val-
FastText vectors have better accuracy as compared to ues is used to calculate the sentiment of the entire sentence
Word2Vec vectors by several varying measures. Yang or document. However, Jurek et al. (2015) tried a different
et al. (2018) proved that the choice of appropriate word approach called the normalization function to calculate the
embedding based on neural networks could lead to signifi- sentiment value more accurately than this basic summation
cant improvements even in the case of out of vocabulary and mean function. Dictionary-based approach and corpus-
(OOV) words. Authors compared various word embed- based approach are two types of lexicon-based approaches
dings, trained using Twitter and Wikipedia as corpora with based on sentiment lexicon. In general, a dictionary main-
TF-IDF word embedding. tains words of some language systemically, whereas a corpus
is a random sample of text in some language. The exact
meaning applies here in the dictionary-based approach and
corpus-based approach. In the dictionary-based approach,

Fig. 4  Techniques for sentiment analysis and emotion detection

13
 81 Page 10 of 19 Social Network Analysis and Mining (2021) 11:81

a dictionary of seed words is maintained (Schouten and used for sentiment analysis fall under supervised classifica-
Frasincar 2015). To create this dictionary, the first small set tion. Different kinds of algorithms required for sentiment
of sentiment words, possibly with very short contexts like classification may include Naïve Bayes, support vector
negations, is collected along with its polarity labels (Bern- machine (SVM), decision trees, etc. each having its pros
abé-Moreno et al. 2020). The dictionary is then updated by and cons. Gamon (2004) applied a support vector machine
looking for their synonymous (words with the same polar- over 40,884 customer feedbacks collected from surveys. The
ity) and antonymous (words with opposite polarity). The authors implemented various feature set combinations and
accuracy of sentiment analysis via this approach will depend achieved accuracy up to 85.47%. Ye et al. (2009) worked
on the algorithm. However, this technique does not contain with SVM, N-gram model, and Naïve Bayes on sentiment
domain specificity. The Corpus-based approach solves the and review on seven popular destinations of Europe and the
limitations of the dictionary-based approach by including USA, which was collected from yahoo.com. The authors
domain-specific sentiment words where the polarity label is achieved an accuracy of up to 87.17% with the n-gram
assigned to the sentiment word according to its context or model. indent Bučar et al. (2018) created the lexicon called
domain. It is a data-driven approach where sentiment words JOB 1.0 and labeled news corpora called SentiNews 1.0
along with context can be accessed. This approach can cer- for sentiment analysis in Slovene texts. JOB 1.0 consists
tainly be a rule-based approach with some NLP parsing of 25,524 headwords extended with sentiment scaling from
techniques. Thus corpus-based approach tends to have poor – 5 to 5 based on the AFINN model. For the construction of
generalization but can attain excellent performance within a corpora, data were scraped from various news Web media.
particular domain. Since the dictionary-based approach does Then, after cleaning and pre-processing of data, the annota-
not consider the context around the sentiment word, it leads tors were asked to annotate 10,427 documents on the 1–5
to less efficiency. Thus, Cho et al. (2014) explicitly handled scale where one means negative and 5 means very positive.
the contextual polarity to make dictionaries adaptable in Then these documents were labeled with positive, negative,
multiple domains with a data-driven approach. They took and neutral labels as per the specific average scale rating.
a three-step strategy: merge various dictionaries, remove The authors observed that Naïve Bayes performed better
the words that do not contribute toward classification, and as compared to the support vector machine (SVM). Naive
switch the polarity according to a particular domain. Bayes achieved an F1 score above 90% in binary classifica-
SentiWordNet (Esuli and Sebastiani 2006) and Valence tion and an F1 score above 60% for the three-class classifica-
Aware Dictionary and Sentiment Reasoner (VADER) (Hutto tion of sentiments. Tiwari et al. (2020) implemented three
and Gilbert 2014) are popular lexicons in sentiment. Jha machine learning algorithms called SVM, Naive Bayes,
et al. (2018) tried to extend the lexicon application in multi- and maximum entropy with the n-gram feature extraction
ple domains by creating a sentiment dictionary named Hindi method on the rotten tomato dataset. The training and test-
Multi-Domain Sentiment Aware Dictionary (HMDSAD) ing dataset constituted 1600 reviews in each. The authors
for document-level sentiment analysis. This dictionary can observed a decrease in accuracy with higher values of n in
be used to annotate the reviews into positive and negative. n-grams such as n = four, five, and six. Soumya and Pramod
The proposed method labeled 24% more words than the (2020) classified 3184 Malayalam tweets into positive and
traditional general lexicon Hindi Sentiwordnet (HSWN), negative opinions using different feature vectors like BOW,
a domain-specific lexicon. The semantic relationships Unigram with Sentiwordnet, etc. The authors implemented
between words in traditional lexicons have not been exam- machine learning algorithms like random forest and Naïve
ined, improving sentiment classification performance. Based Bayes and observed that the random forest with an accuracy
on this premise, Viegas et al. (2020) updated the lexicon by of 95.6% performs better with Unigram Sentiwordnet con-
including additional terms after utilizing word embeddings sidering negation words.
to discover sentiment values for these words automatically. Deep Learning-based Approach In recent years, deep
These sentiment values were derived from “nearby” word learning algorithms are dominating other traditional
embeddings of already existing words in the lexicon. approaches for sentiment analysis. These algorithms detect
Machine Learning-based approach There is another the sentiments or opinions from text without doing feature
approach for sentiment analysis called the machine learn- engineering. There are multiple deep learning algorithms,
ing approach. The entire dataset is divided into two parts namely recurrent neural network and convolutional neu-
for training and testing purposes: a training dataset and a ral networks, that can be applied to sentiment analysis and
testing dataset. The training dataset is the information used gives results that are more accurate than those provided by
to train the model by supplying the characteristics of differ- machine learning models. This approach makes humans free
ent instances of an item. The testing dataset is then used to from constructing the features from text manually as deep
see how successfully the model from the training dataset has learning models extract those features or patterns them-
been trained. Generally, the machine learning algorithms selves. Jian et al. (2010) used a model based upon neural

13
Social Network Analysis and Mining (2021) 11:81 Page 11 of 19 81

networks technology for categorizing sentiments which of SVM and random forest model, i.e., RFSVM, on ama-
consisted of sentimental features, feature weight vectors, zon’s product reviews. The authors concluded RFSVM,
and prior knowledge base. The authors applied the model with an accuracy level of 83.4%, performs better than SVM
to review the data of Cornell movie. The experimental with 82.4% accuracy and random forest with 81% accuracy
results of this paper revealed that the accuracy level of the individually over the dataset of 1000 reviews. Alqaryouti
I-model is extraordinary compared to HMM and SVM. et al. (2020) proposed the hybrid of the rule-based approach
Pasupa and Ayutthaya (2019) executed five-fold cross-val- and domain lexicons for aspect-level sentiment detection to
idation on the children’s tale (Thai) dataset and compared understand people’s opinions regarding government smart
three deep learning models called CNN, LSTM, and Bi- applications. The authors concluded that the proposed tech-
LSTM. These models are applied with or without features: nique outperforms other lexicon-based baseline models by
POS-tagging (pre-processing technique to identify different 5%. Ray and Chakrabarti (2020) combined the rule-based
parts of speech); Thai2Vec (word embedding trained from approach to extract aspects with a 7-layer deep learning
Thai Wikipedia); sentic (to understand the sentiment of the CNN model to tag each aspect. The hybrid model achieved
word). The authors observed the best performance in the 87% accuracy, whereas the individual models had 75%
CNN model with all the three features mentioned earlier. accuracy with rule-based and 80% accuracy with the CNN
As stated earlier, social media platforms act as a signifi- model.
cant source of data in the field of sentiment analysis. Data Table 3 describes various machine learning and deep
collected from this social sites consist lot of noise due to learning algorithms used for analyzing sentiments in multi-
its free writing syle of users. Therefore, Arora and Kansal ple domains. Many researchers implemented the proposed
(2019) proposed a model named Conv-char-Emb that can models on their dataset collected from Twitter and other
handle the problem of noisy data and use small memory social networking sites. The authors then compared their
space for embedding. For embedding, convolution neural proposed models with other existing baseline models and
network (CNN) has been used that uses less parameters in different datasets. It is observed from the table above that
feature representation. Dashtipour et al. (2020) proposed a accuracy by various models ranges from 80 to 90%.
deep learning framework to carry out sentiment analysis in
the Persian language. The researchers concluded that deep 3.4.2 Emotion detection techniques
neural networks such as LSTM and CNN outperformed the
existing machine learning algorithms on the hotel and prod- Lexicon-based Approach Lexicon-based approach is a key-
uct review dataset. word-based search approach that searches for emotion key-
Transfer Learning Approach and Hybrid Approach words assigned to some psychological states (Rabeya et al.
Transfer learning is also a part of machine learning. A 2017). The popular lexicons for emotion detection are Word-
model trained on large datasets to resolve one problem can Net-Affect (Strapparava et al. 2004 and NRC word–emotion
be applied to other related issues. Re-using a pre-trained lexicon (Mohammad and Turney 2013). WordNet-Affect is
model on related domains as a starting point can save time an extended form of WordNet which consists of affective
and produce more efficient results. Zhang et al. (2012) pro- words annotated with emotion labels. NRC lexicon consists
posed a novel instance learning method by directly mod- of 14,182 words, each assigned to one particular emotion
eling the distribution between different domains. Authors and two sentiments. These lexicons are categorical lexicons
classified the dataset: Amazon product reviews and Twitter that tag each word with an emotional state for emotion clas-
dataset into positive and negative sentiments. Tao and Fang sification. However, by ignoring the intensity of emotions,
(2020) proposed extending recent classification methods in these traditional lexicons become less informative and less
aspect-based sentiment analysis to multi-label classification. adaptable. Thus, Li et al. (2021) suggested an effective
The authors also developed transfer learning models called strategy to obtain word-level emotion distribution to assign
XLNet and Bert and evaluated the proposed approach on emotions with intensities to the sentiment words by merg-
different datasets Yelp, wine reviews rotten tomato dataset ing a dimensional dictionary named NRC-Valence arousal
from other domains. Deep learning and machine learning dominance. EmoSenticNet (Poria et al. 2014) also consists
approaches yield good results, but the hybrid approach can of a large number assigned to both qualitative and quantita-
give better results since it overcomes the limitations of each tive labels. Generally, researchers generate their lexicons and
traditional model. Mladenović et al. (2016) proposed a fea- directly apply them for emotion analysis, but lexicons can
ture reduction technique, a hybrid framework made of sen- also be used for feature extraction purposes. Abdaoui et al.
timent lexicon and Serbian wordnet. The authors expanded (2017) took the benefit of using online translation tools to
both lexicons by addition some morphological sentiment create a French lexicon called FEEL (French expanded emo-
words to avoid loss of critical information while stemming. tion lexicon) consisting of more than 14,000 words with both
Al Amrani et al. (2018) compared their hybrid model made polarity and emotion labels. This lexicon was created by

13
 81

Table 3  Work on sentiment analysis

Reference Level Technique Feature extraction Learning algorithm Domain Dataset Results

13
Songbo and Jin (2008) Sentence Machine learning – Centroid classifier, House, movie and Chn- sentiCorp Micro F1 = 90.60% with
K-nearest Classifier, education SVM and IG and macro
Winnow Classifier, F1 = 90.43%.
Page 12 of 19

Naïve Bayes, SVM

Moraes (2013) Aspect level Machine learning and Bag of words Artificial neural net- Movies, Books, GPS, – Accuracy = 86.5% on
deep learning work (ANN), Naïve Cameras movie dataset, 87.3% on
Bayes, SVM GPS dataset, 81.8% on
book dataset 90.6% on
camera dataset with ANN.
Tang et al. (2015) Document-level Deep learning Word embeddings UPNN (user product Movies Dataset collected Accuracy = 58.5% with
to dense docu- neutral network) from yelp dataset UPNN (no UP) and 60.8%
ment vector based on CNN and IMDB with UPNN on Yelp 2014.
Dahou et al. (2016) – Deep learning Word embed- Convolutional neural Book, movie, restaurant LABR book reviews, Accuracy= 91.7% on and
ding built from network (CNN) etc. Arabic sentiment Accuracy = 89.6% unbal-
Arabic corpus tweet dataset, etc anced HTL and LABR
dataset, respectively.
Ahuja et al. (2019) Sentence Machine learning TF-IDF, n-gram KNN, SVM, logistic – SS-tweets Accuracy = 57% with TF-
regression, NB, ran- IDF and logistic regres-
dom forest sion and accuracy = 51%
with n-gram and random
forest.
Untawale and Choud- – Machine learning – Naïve Bayes and ran- movie reviews Rotten tomatoes, Naïve Bayes required more
hari (2019) dom forest reviews from time and memory than
Times of India, etc random forest.
Shamantha et al. (2019) - Machine learning – Naïve bayes, SVM and Twitter Twitter Accuracy = above 80% with
random forest Naïve Bayes (3features)
on 200 tweets.
Goularas and Kamis – Machine learning – Random forest and – – Accuracy = 95% with
(2019) SVM random forest.
Nandal et al. (2020) Aspect level Machine learning – SVM with different – Amazon reviews Mean square error = 0.04
kernels: linear, radial with radial basis func-
basis function (RBF), tion and 0.11 with linear
and polynomial kernel.
Social Network Analysis and Mining

Sharma and Sharma – Machine learning and – Deep artificial neural Twitter Twitter Positive emotion rate =
(2020) deep learning network and SVM 87.5 with the proposed
algorithm.
Mukherjee et al. (2021) Sentence level Machine learning and TF IDF Naïve Bayes, support Cellphone reviews Amazon reviews Accuracy = 95.30%
Deep Learning with vector machines, arti- with RNN + Nega-
(2021) 11:81

negation prediction ficial neural network tion and 95.67% with

process (ANN), and recurrent ANN+negation.
neural network (RNN)
Social Network Analysis and Mining (2021) 11:81 Page 13 of 19 81

increasing the number of words in the NRC emotion lexicon approach is based on the probability of multiple emotions
and semi-automatic translation using six online translators. present in the sentence and utilized both semantic and
Those entries obtained from at least three translators were sentiment representation for better emotion classification.
considered pre-validated and then validated by the manual Results are evaluated over their own constructed dataset
translator. Bandhakavi et al. (2017) applied a domain-spe- with tweet conversation pairs, and their model is compared
cific lexicon for the process of feature extraction in emotion with other baseline models. Xu et al. (2020) extracted fea-
analysis. The authors concluded that features derived from tures emotions using two-hybrid models named 3D con-
their proposed lexicon outperformed the other baseline fea- volutional-long short-term memory (3DCLS) and CNN-
tures. Braun et al. (2021) constructed a multilingual corpus RNN from video and text, respectively. At the same time,
called MEmoFC, which stands for Multilingual Emotional the authors implemented SVM for audio-based emotion
Football Corpus, consisting of football reports from English, classification. Authors concluded results by fusing audio
Dutch and German Web sites and match statistics crawled and video features at feature level with MKL fusion tech-
from Goal.com. The corpus was created by creating two nique and further combining its results with text-based
metadata tables: one explaining details of a match like a emotion classification results. It provides better accuracy
date, place, participation teams, etc., and the second table than every other multimodal fusion technique, intending to
consisted of abbreviations of football clubs. Authors dem- analyze the sentiments of drug reviews written by patients
onstrated the corpus with various approaches to know the on social media platforms. Basiri et al. (2020) proposed
influence of the reports on game outcomes. two models using a three-way decision theory. The first
Machine Learning-based Techniques Emotion detec- model is a 3-way fusion of one deep learning model with
tion or classification may require different types of machine the traditional learning method (3W1DT), while the other
learning models such as Naïve Bayes, support vector model is a 3-way fusion of three deep learning models with
machine, decision trees, etc. Jain et al. (2017) extracted the the conventional learning method (3W3DT). The results
emotions from multilingual texts collected from three dif- derived using the Drugs.com dataset revealed that both
ferent domains. The authors used a novel approach called frameworks performed better than traditional deep learn-
rich site summary for data collection and applied SVM and ing techniques. Furthermore, the performance of the first
Naïve Bayes machine learning algorithms for emotion clas- fusion model was noted to be much better as compared to
sification of twitter text. Results revealed that an accuracy the second model in regards to accuracy and F1-metric.
level of 71.4% was achieved with the Naïve Bayes algo- In recent days, social media platforms are flooded with
rithm. Hasan et al. (2019) evaluated the machine learning posts related to covid-19. Singh et al. (2021) applied
algorithms like Naïve Bayes, SVM, and decision trees to emotion detection analysis on covid-19 tweets collected
identify emotions in text messages. The task is divided into from the whole world and India only with Bidirectional
two subtasks: Task 1 includes a collection of the dataset Encoder Representations from Transformers (BERT)
from Twitter and automatic labeling of the dataset using model on the Twitter data sets and achieved accuracy 94%
hashtags and model training. Task 2 is developing a two- approximately.
stage EmotexStream that separates emotionless tweets at the Transfer Learning Approach In traditional approaches,
first stage and identifies emotions in the text by utilizing the the common presumption is that the dataset is from the same
models trained in the task1. The authors observed accuracy domain; however, there is a need for a new model when
of 90% in classifying emotions. Asghar et al. (2019) aimed the domain changes. The transfer learning approach allows
to apply multiple machine learning models on the ISEAR you to reuse the existing pre-trained models in the target
dataset to find the best classifier. They found that the logistic domain. For example, Ahmad et al. (2020) used a transfer
regression model performed better than other classifiers with learning technique due to the lack of resources for emotion
a recall value of 83%. detection in the Hindi language. The researchers pre-trained
Deep Learning and Hybrid Technique Deep learning a model on two different English datasets: SemEval-2018,
area is part of machine learning that processes informa- sentiment analysis, and one Hindi dataset with positive, neu-
tion or signals in the same way as the human brain does. tral, conflict, and negative labels. They achieved a score of
Deep learning models contain multiple layers of neurons. 0.53 f1 using the transfer learning and 0.47 using only base
Thousands of neurons are interconnected to each other, models CNN and Bi-LSTM with cross-lingual word embed-
which speeds up the processing in a parallel fashion. Chat- ding. Hazarika et al. (2020) created a TL-ERC model where
terjee et al. (2019) developed a model called sentiment and the model was pre-trained over source multi-turn conversa-
semantic emotion detection (SSBED) by feeding sentiment tions and then transferred over emotion classification task on
and semantic representations to two LSTM layers, respec- exchanged messages. The authors emphasized the issues like
tively. These representations are then concatenated and lack of labeled data in multi-conversations with the frame-
then passed to a mesh network for classification. The novel work based on inductive transfer learning.

13
 81 Page 14 of 19 Social Network Analysis and Mining (2021) 11:81

Table 4 shows that most researchers implemented mod- for researchers. Furthermore, some of the corpora and lexi-
els by combining machine learning and deep learning tech- cons are domain specific, which limits their re-use in other
niques with various feature extraction techniques. Most of domains.
the datasets are available in the English language. However, Another common problem is usually seen on Twitter,
some researchers constructed the dataset of their regional Facebook, and Instagram posts and conversations is Web
language. For example, Sasidhar et al. (2020) created the slang. For example, the Young generation uses words like
dataset of Hindi-English code mixed with three basic emo- 'LOL,' which means laughing out loud to express laughter,
tions: happy, sad, and angry, and observed CNN-BILSTM 'FOMO,' which means fear of missing out, which says anxi-
gave better performance compared to others. ety. The growing dictionary of Web slang is a massive obsta-
cle for existing lexicons and trained models.
3.5 Model assessment People usually express their anger or disappointment in
sarcastic and irony sentences, which is hard to detect (Ghan-
Finally, the model is compared with baseline models based bari-Adivi and Mosleh 2019). For instance, in the sentence,
on various parameters. There is a requirement of model eval- “This story is excellent to put you in sleep,” the excellent
uation metrics to quantify model performance. A confusion word signifies positive sentiment, but in actual the reviewer
matrix is acquired, which provides the count of correct and felt it quite dull. Therefore, sarcasm detection has become a
incorrect judgments or predictions based on known actual tedious task in the field of sentiment and emotion detection.
values. This matrix displays true positive (TP), false nega- The other challenge is the expression of multiple emo-
tive (FN), false positive (FP), true negative (TN) values for tions in a single sentence. It is difficult to determine vari-
data fitting based on positive and negative classes. Based on ous aspects and their corresponding sentiments or emotions
these values, researchers evaluated their model with metrics from the multi-opinionated sentence. For instance, the
like accuracy, precision, and recall, F1 score, etc., mentioned sentence “view at this site is so serene and calm, but this
in Table 5. place stinks” shows two emotions, 'disgust' and 'soothing'
in various aspects. Another challenge is that it is hard to
detect polarity from comparative sentences. For example,
4 Challenges in sentiment analysis consider two sentences 'Phone A is worse than phone B'
and emotion analysis and 'Phone B is worse than Phone A.' The word ’worse’ in
both sentences will signify negative polarity, but these two
In the Internet era, people are generating a lot of data in the sentences oppose each other (Shelke 2014).
form of informal text. Social networking sites present vari-
ous challenges, as shown in Fig. 5, which includes spelling
mistakes, new slang, and incorrect use of grammar. These 5 Conclusion
challenges make it difficult for machines to perform senti-
ment and emotion analysis. Sometimes individuals do not In this paper, a review of the existing techniques for both
express their emotions clearly. For instance, in the sentence emotion and sentiment detection is presented. As per the
“Y have u been soooo late?”, 'why' is misspelled as 'y,' 'you' paper’s review, it has been analyzed that the lexicon-based
is misspelled as 'u,' and 'soooo' is used to show more impact. technique performs well in both sentiment and emotion
Moreover, this sentence does not express whether the person analysis. However, the dictionary-based approach is quite
is angry or worried. Therefore, sentiment and emotion detec- adaptable and straightforward to apply, whereas the corpus-
tion from real-world data is full of challenges due to several based method is built on rules that function effectively in a
reasons (Batbaatar et al. 2019). certain domain. As a result, corpus-based approaches are
One of the challenges faced during emotion recognition more accurate but lack generalization. The performance of
and sentiment analysis is the lack of resources. For example, machine learning algorithms and deep learning algorithms
some statistical algorithms require a large annotated dataset. depends on the pre-processing and size of the dataset. None-
However, gathering data is not difficult, but manual labeling theless, in some cases, machine learning models fail to
of the large dataset is quite time-consuming and less reliable extract some implicit features or aspects of the text. In situ-
(Balahur and Turchi 2014). The other problem regarding ations where the dataset is vast, the deep learning approach
resources is that most of the resources are available in the performs better than machine learning. Recurrent neural
English language. Therefore, sentiment analysis and emo- networks, especially the LSTM model, are prevalent in sen-
tion detection from a language other than English, primarily timent and emotion analysis, as they can cover long-term
regional languages, are a great challenge and an opportunity dependencies and extract features very well. But RNN with

13
 Table 4  Work on emotion detection
Reference Approach Feature extraction Models Datasets Emotion model No of emotions Results

Chaffar and Inkpen Machine learning Bag of words, Naïve Bayes, decision Multiple dataset Ekman with neutral 10 Acuracy = 81.16% on
(2011) N-grams, WordN- tree, and SVM class, Izard Aman’s dataset and
etAffect 71.69% on Global
dataset
Kratzwald et al. (2018) Deep learning with Customised embed- Sent2Affect Literary tales, election – – F1-score = 68.8% on
transfer learning ding GloVe tweet Isear Headlines literary dataset with
approach General tweets pre-trained Bi-LSTM
Social Network Analysis and Mining

Sailunaz and Alhajj Machine learning NAVA (Noun Adverb, SVM, Random forest, ISEAR Guilt, Joy, Shame, 6 Accuracy = 43.24% on
(2019) verb and Adjective) Naïve Bayes Fear, sadness, disgust NAVA text with Naïve
Bayes.
Shrivastava et al. Deep learning Word2Vec Convolutional neural TV shows transcript – 7 Training accuracy =
(2019) network 80.41% and 77.54%
(2021) 11:81

with CNN (7 emo-

tions)
Batbaatar et al. (2019) Deep learning Word2Vec, GloVe, SENN ISEAR, Emo Int, elec- – – Acuracy = 98.8% with
FastText, EWE toral tweets, etc GloVe+EWE and
SENN on emotion
cause dataset
Ghanbari-Adivi and Deep learning DoctoVec Ensemble classi- wonder, anger, hate, 6 OANC, Crowd- 99.49 on regular sen-
Mosleh (2019) fier, tree-structured happy , sadness, and Flower, tences
parzen estimator fear ISEAR,
(TPE) for tuning
parameters
Xu et al. (2020) Deep learning-based – 3DCLS model for Moud and IEMOCAP Happy, sad, angry, 4 Accuracy = 96.75%
Hybrid Approach visual , CNN-RNN neutral by fusing audio and
for text and SVM visual features at fea-
for text ture level on MOUD
dataset
Adoma et al. (2020) Pretrained transfer – BERT, RoBERTa, Dis- ISEAR shame, anger,fear, 7 Accuracy = 74%, 79%
models (machine tilBERT, and XLNet disgust,joy, sadness, , 69% for RoBERTa,
learning and deep and guilt BERT, respectively.
learning)
Chowanda et al. (2021) Machine learning and sentistrength, N-gram Generalised linear Affective Tweets Anger, fear, sadness, 4 Accuracy = 92% and
Deep learning and TF IDF model, Naïve Bayes, joy recall = 90% with
fast-large margins, the generalized linear
etc. model
Dheeraj and Ram- Deep learning Glove Multi-head attention Patient doctor interac- Anxiety, addiction, 6 Accuracy = 97.8% using
akrishnudu (2021) with bidirectional tions from Webmd obsessive cleaning MHA-BCNN with
long short-term and Healthtap disorder (OCD), Adam optimizer
Page 15 of 19

memory and convo- platforms depression, etc

lutional neural net-
81

work (MHA-BCNN)

13
 81 Page 16 of 19 Social Network Analysis and Mining (2021) 11:81

Table 5  Evaluation metrics

Evaluation metric Description Equation

Accuracy It’s a statistic that sums up how well the model performs in all classes. It’s help- (TP+TN)/(TP+TN+FP+FN)
ful when all types of classes are equally important. It is calculated as the ratio
between the number of correct judgments to the total number of judgments.
Precision It measures the accuracy of the model in terms of categorizing a sample as posi- TP/(TP+FP )
tive. It is determined as the ratio of the number of correctly categorized Positive
samples to the total number of positive samples (either correctly or incorrectly).
Recall This score assesses the model’s ability to identify positive samples. It is deter- TP/(TP+FN)
mined by dividing the number of Positive samples that were correctly catego-
rized as Positive by the total number of Positive samples.
F-measure It is determined by calculating the harmonic mean of precision and recall. (2*Precision*Recall)/
(Precision+Recall) =
(2*TP)/((2*TP)+FP+FN)
Sensitivity It refers to the percentage of appropriately detected actual positives and it quanti- TP/((TP+FN))
fies how effectively the positive class was anticipated.
Specificity It is the complement of sensitivity, the true negative rate which sums up how TN/(FP+TN)
effectively the negative class was anticipated. The sensitivity of an imbalanced
categorization may be more interesting than specificity.
Geometric-mean (G-mean) It is a measure that combines sensitivity and specificity into a single value that
√
(Specificity ∗ Sensitivity)
balances both objectives.

References
Abdaoui A, Azé J, Bringay S, Poncelet P (2017) Feel: a French
expanded emotion lexicon. Lang Resour Eval 51(3):833–855
Abdi A, Shamsuddin SM, Hasan S, Piran J (2019) Deep learning-based
sentiment classification of evaluative text based on multi-feature
fusion. Inf Process Manag 56(4):1245–1259
Adoma AF, Henry N-M, Chen W (2020) Comparative analyses of
bert, roberta, distilbert, and xlnet for text-based emotion rec-
ognition. In: 2020 17th International Computer Conference on
Wavelet Active Media Technology and Information Processing
(ICCWAMTIP), IEEE, pp 117–121. https://​doi.​org/​10.​1109/​
ICCWA​MTIP5​1612.​2020.​93173​79
Agbehadji IE, Ijabadeniyi A (2021) Approach to sentiment analysis and
business communication on social media. In: FongS, Millham R
(eds) Bio-inspired algorithms for data streaming and visualiza-
tion, big data management, and fog computing, Springer Tracts
in Nature-Inspired Computing. Springer, Singapore. https://​doi.​
org/​10.​1007/​978-​981-​15-​6695-0_9
Agrawal A, An A (2012) Unsupervised emotion detection from text
using semantic and syntactic relations. In: 2012 IEEE/WIC/
ACM international conferences on web intelligence and intel-
ligent agent technology, pp 346–353. https://​doi.​org/​10.​1109/​
WI-​IAT.​2012.​170.
Ahmad Z, Jindal R, Ekbal A, Bhattachharyya P (2020) Borrow from
rich cousin: transfer learning for emotion detection using cross
Fig. 5  Challenges in sentiment analysis and emotion detection lingual embedding. Expert Syst Appl 139:112851
Ahmed WM (2020) Stock market reactions to domestic sentiment:
panel CS-ARDL evidence. Res Int Bus Finance 54:101240
Ahuja R, Chug A, Kohli S, Gupta S, Ahuja P (2019) The impact of
attention networks performs very well. At the same time, it features extraction on the sentiment analysis. Procedia Comput
is important to keep in mind that the lexicon-based approach Sci 152:341–348
and machine learning approach (traditional approaches) are Akilandeswari J, Jothi G (2018) Sentiment classification of tweets with
also evolving and have obtained better outcomes. Also, pre- non-language features. Procedia Comput Sci 143:426–433
Al Ajrawi S, Agrawal A, Mangal H, Putluri K, Reid B, Hanna G, Sarkar
processing and feature extraction techniques have a signifi- M (2021) Evaluating business yelp’s star ratings using sentiment
cant impact on the performance of various approaches of analysis. Materials Today: Proceedings. https://d​ oi.o​ rg/1​ 0.1​ 016/j.​
sentiment and emotion analysis. matpr.​2020.​12.​137

13
Social Network Analysis and Mining (2021) 11:81 Page 17 of 19 81

Al Amrani Y, Lazaar M, El Kadiri KE (2018) Random forest and intelligence. Canadian AI 2011. Lecture notes in computer sci-
support vector machine based hybrid approach to sentiment ence, vol 6657. Springer, Berlin, Heidelberg. https://​doi.​org/​10.​
analysis. Procedia Comput Sci 127:511–520 1007/​978-3-​642-​21043-3_8
Alqaryouti O, Siyam N, Monem AA, Shaalan K (2020) Aspect-based Chatterjee A, Gupta U, Chinnakotla MK, Srikanth R, Galley M,
sentiment analysis using smart government review data. Appl Agrawal P (2019) Understanding emotions in text using deep
Comput Inf. https://​doi.​org/​10.​1016/j.​aci.​2019.​11.​003 learning and big data. Comput Hum Behav 93:309–317
Alswaidan N, Menai MEB (2020) A survey of state-of-the-art Chen T, Xu R, He Y, Wang X (2017) Improving sentiment analysis
approaches for emotion recognition in text. Knowl Inf Syst via sentence type classification using BILSTM-CRF and CNN.
62(8):1–51 Expert Syst Appl 72:221–230
Archana Rao PN, Baglodi K (2017) Role of sentiment analysis in Cho H, Kim S, Lee J, Lee JS (2014) Data-driven integration of multiple
education sector in the era of big data: a survey. Int J Latest sentiment dictionaries for lexicon-based sentiment classification
Trends Eng Technol 22–24 of product reviews. Knowl-Based Syst 71:61–71
Arora M, Kansal V (2019) Character level embedding with deep Chowanda A, Sutoyo R, Tanachutiwat S et al (2021) Exploring text-
convolutional neural network for text normalization of unstruc- based emotions recognition machine learning techniques on
tured data for twitter sentiment analysis. Soc Netw Anal Min social media conversation. Procedia Comput Sci 179:821–828
9(1):1–14 Dahou A, Xiong S, Zhou J, Haddoud MH, Duan P (2016) Word embed-
Arulmurugan R, Sabarmathi K, Anandakumar H (2019) Classification dings and convolutional neural network forarabic sentiment clas-
of sentence level sentiment analysis using cloud machine learn- sification. In: Proceedings of coling 2016, the 26th international
ing techniques. Cluster Comput 22(1):1199–1209 conference on computational linguistics: Technical papers, pp
Asghar MZ, Subhan F, Imran M, Kundi FM, Shamshirband S, Mosavi 2418–2427
A, Csiba P, Várkonyi-Kóczy AR (2019) Performance evaluation Dashtipour K, Gogate M, Li J, Jiang F, Kong B, Hussain A (2020)
of supervised machine learning techniques for efficient detection A hybrid Persian sentiment analysis framework: integrating
of emotions from online content. arXiv preprint arXiv:​19080​ dependency grammar based rules and deep neural networks.
1587 Neurocomputing 380:1–10
Bakker I, Van Der Voordt T, Vink P, De Boon J (2014) Pleasure, Devi Sri Nandhini M, Pradeep G (2020) A hybrid co-occurrence and
arousal, dominance: Mehrabian and Russell revisited. Curr Psy- ranking-based approach for detection of implicit aspects in
chol 33(3):405–421 aspect-based sentiment analysis. SN Comput Sci 1:1–9
Balahur A, Turchi M (2014) Comparative experiments using super- Dheeraj K, Ramakrishnudu T (2021) Negative emotions detection on
vised learning and machine translation for multilingual sentiment online mental-health related patients texts using the deep learn-
analysis. Comput Speech Lang 28(1):56–75 ing with MHA-BCNN model. Expert Syst Appl 182:115265
Bandhakavi A, Wiratunga N, Padmanabhan D, Massie S (2017) Lexi- Dixon T (2012) “Emotion”: the history of a keyword in crisis. Emot
con based feature extraction for emotion text classification. Pat- Rev 4(4):338–344
tern Recogn Lett 93:133–142 Ekman P (1992) An argument for basic emotions. Cognit Emot
Basiri ME, Abdar M, Cifci MA, Nemati S, Acharya UR (2020) A 6(3–4):169–200
novel method for sentiment classification of drug reviews using Esuli A, Sebastiani F (2006) Sentiwordnet: a publicly available lexical
fusion of deep and machine learning techniques. Knowl Based resource for opinion mining. LREC, Citeseer 6:417–422
Syst 198:105949 Gamon M (2004) Sentiment classification on customer feedback data:
Batbaatar E, Li M, Ryu KH (2019) Semantic-emotion neural noisy data, large feature vectors, and the role of linguistic analy-
network for emotion recognition from text. IEEE Access sis. In: COLING 2004: Proceedings of the 20th international
7:111866–111878 conference on computational linguistics, pp 841–847
Becker K, Moreira VP, dos Santos AG (2017) Multilingual emotion Garcia K, Berton L (2021) Topic detection and sentiment analysis in
classification using supervised learning: comparative experi- twitter content related to covid-19 from brazil and the USA. Appl
ments. Inf Process Manag 53(3):684–704 Soft Comput 101:107057
Bernabé-Moreno J, Tejeda-Lorente A, Herce-Zelaya J, Porcel C, Her- Ghanbari-Adivi F, Mosleh M (2019) Text emotion detection in social
rera-Viedma E (2020) A context-aware embeddings supported networks using a novel ensemble classifier based on Parzen tree
method to extract a fuzzy sentiment polarity dictionary. Knowl- estimator (tpe). Neural Comput Appl 31(12):8971–8983
Based Syst 190:105236 Goularas D, Kamis S (2019) Evaluation of deep learning techniques
Bhardwaj A, Narayan Y, Dutta M et al (2015) Sentiment analysis for in sentiment analysis from twitter data. In: 2019 International
Indian stock market prediction using sensex and nifty. Procedia conference on deep learning and machine learning in emerging
Comput Sci 70:85–91 applications (Deep-ML), IEEE, pp 12–17
Bhaskar J, Sruthi K, Nedungadi P (2015) Hybrid approach for emotion Gräbner D, Zanker M, Fliedl G, Fuchs M, et al. (2012) Classification
classification of audio conversation based on text and speech of customer reviews based on sentiment analysis. In: ENTER,
mining. Procedia Comput Sci 46:635–643 Citeseer, pp 460–470
Braun N, van der Lee C, Gatti L, Goudbeek M, Krahmer E (2021) Hasan M, Rundensteiner E, Agu E (2014) Emotex: detecting emotions
Memofc: introducing the multilingual emotional football corpus. in twitter messages. In: 2014 ASE BIGDATA/SOCIALCOM/
Lang Resour Eval 55(2):389–430 CYBERSECURITY Conference. Stanford University, Academy
Bučar J, Žnidaršič M, Povh J (2018) Annotated news corpora and a of Science and Engineering (ASE), USA, ASE, pp 1–10
lexicon for sentiment analysis in Slovene. Lang Resour Eval Hasan M, Rundensteiner E, Agu E (2019) Automatic emotion detec-
52(3):895–919 tion in text streams by analyzing twitter data. Int J Data Sci Anal
Buechel S, Hahn U (2017) Emobank: Studying the impact of anno- 7(1):35–51
tation perspective and representation format on dimensional Hazarika D, Poria S, Zimmermann R, Mihalcea R (2020) Conver-
emotion analysis. In: Proceedings of the 15th conference of the sational transfer learning for emotion recognition. Inf Fusion
european chapter of the association for computational linguistics: 65:1–12
volume 2, Short Papers, pp 578–585 Hosseini AS (2017) Sentence-level emotion mining based on com-
Chaffar S, Inkpen D (2011) Using a heterogeneous dataset for emotion bination of adaptive meta-level features and sentence syntactic
analysis in text. In: Butz C, Lingras P (eds) Advances in artificial features. Eng Appl Artif Intell 65:361–374

13
 81 Page 18 of 19 Social Network Analysis and Mining (2021) 11:81

Hutto C, Gilbert E (2014) Vader: a parsimonious rule-based model for sentiment analysis and polarity detection. Procedia Comput
sentiment analysis of social media text. In: Proceedings of the Sci 185:370–379
international AAAI conference on web and social media, vol 8 Munezero M, Montero CS, Sutinen E, Pajunen J (2014) Are they dif-
Itani M, Roast C, Al-Khayatt S (2017) Developing resources for senti- ferent? affect, feeling, emotion, sentiment, and opinion detec-
ment analysis of informal Arabic text in social media. Procedia tion in text. IEEE Trans Affect Comput 5(2):101–111
Comput Sci 117:129–136 Nagamanjula R, Pethalakshmi A (2020) A novel framework based
Izard CE (1992) Basic emotions, relations among emotions, and emo- on bi-objective optimization and lan 2 fis for twitter sentiment
tion-cognition relations. Psychol Rev 99(3):561–565 analysis. Soc Netw Anal Min 10:1–16
Jain VK, Kumar S, Fernandes SL (2017) Extraction of emotions from Nagarajan SM, Gandhi UD (2019) Classifying streaming of twitter
multilingual text using intelligent text processing and computa- data based on sentiment analysis using hybridization. Neural
tional linguistics. J Comput Sci 21:316–326 Comput Appl 31(5):1425–1433
Jang HJ, Sim J, Lee Y, Kwon O (2013) Deep sentiment analy- Nandal N, Tanwar R, Pruthi J (2020) Machine learning based aspect
sis: mining the causality between personality-value-attitude level sentiment analysis for amazon products. Spat Inf Res
for analyzing business ads in social media. Expert Syst Appl 28(5):601–607
40(18):7492–7503 Onyenwe I, Nwagbo S, Mbeledogu N, Onyedinma E (2020) The
Jha V, Savitha R, Shenoy PD, Venugopal K, Sangaiah AK (2018) A impact of political party/candidate on the election results from
novel sentiment aware dictionary for multi-domain sentiment a sentiment analysis perspective using# anambradecides2017
classification. Comput Electr Eng 69:585–597 tweets. Soc Netw Anal Min 10(1):1–17
Jian Z, Chen X, Wang Hs (2010) Sentiment classification using the Pasupa K, Ayutthaya TSN (2019) Thai sentiment analysis with
theory of ANNs. J China Univ Posts Telecommun 17:58–62 deep learning techniques: a comparative study based on word
Jurek A, Mulvenna MD, Bi Y (2015) Improved lexicon-based senti- embedding, pos-tag, and sentic features. Sustain Cities Soc
ment analysis for social media analytics. Secur Inform 4(1):1–13 50:101615
Kratzwald B, Ilić S, Kraus M, Feuerriegel S, Prendinger H (2018) Deep Plutchik R (1982) A psychoevolutionary theory of emotions. https://​
learning for affective computing: text-based emotion recognition doi.​org/​10.​1177/​05390​18820​21004​003
in decision support. Decis Support Syst 115:24–35 Poria S, Gelbukh A, Cambria E, Hussain A, Huang GB (2014)
Laubert C, Parlamis J (2019) Are you angry (happy, sad) or aren’t you? Emosenticspace: a novel framework for affective common-sense
Emotion detection difficulty in email negotiation. Group Decis reasoning. Knowl-Based Syst 69:108–123
Negot 28(2):377–413 Prabowo R, Thelwall M (2009) Sentiment analysis: a combined
Le Q, Mikolov T (2014) Distributed representations of sentences and approach. J Inform 3(2):143–157
documents. In: International conference on machine learning, Pu X, Wu G, Yuan C (2019) Exploring overall opinions for document
pp 1188–1196 level sentiment classification with structural SVM. Multim Syst
Li Z, Xie H, Cheng G, Li Q (2021) Word-level emotion distribution 25(1):21–33
with two schemas for short text emotion classification. Knowl- Rabeya T, Ferdous S, Ali HS, Chakraborty NR (2017) A survey on
edge-Based Syst 227:107163 emotion detection: a lexicon based backtracking approach for
Liu Y, Wan Y, Su X (2019) Identifying individual expectations in ser- detecting emotion from Bengali text. In: 2017 20th international
vice recovery through natural language processing and machine conference of computer and information technology (ICCIT),
learning. Expert Syst Appl 131:288–298 IEEE, pp 1–7
Liu F, Zheng J, Zheng L, Chen C (2020a) Combining attention-based Rao G, Huang W, Feng Z, Cong Q (2018) LSTM with sentence rep-
bidirectional gated recurrent neural network and two-dimensional resentations for document-level sentiment classification. Neuro-
convolutional neural network for document-level sentiment clas- computing 308:49–57
sification. Neurocomputing 371:39–50 Ray P, Chakrabarti A (2020) A mixed approach of deep learning
Liu S, Lee K, Lee I (2020b) Document-level multi-topic sentiment method and rule-based method to improve aspect level senti-
classification of email data with bilstm and data augmentation. ment analysis. Appl Comput Inform
Knowl-Based Syst 197:105918 Roberts K, Roach MA, Johnson J, Guthrie J, Harabagiu SM (2012)
Lövheim H (2012) A new three-dimensional model for emotions and Empatweet: annotating and detecting emotions on twitter. Lrec,
monoamine neurotransmitters. Med Hypoth 78(2):341–348 Citeseer 12:3806–3813
Ma X, Zeng J, Peng L, Fortino G, Zhang Y (2019) Modeling multi- Russell JA (1980) A circumplex model of affect. J Pers Soc Psychol
aspects within one opinionated sentence simultaneously for 39(6):1161
aspect-level sentiment analysis. Future Gener Comput Syst Sailunaz K, Alhajj R (2019) Emotion and sentiment analysis from twit-
93:304–311 ter text. J Comput Sci 36:101003
Meena A, Prabhakar TV (2007) Sentence level sentiment analysis in Salinca A (2015) Business reviews classification using sentiment analy-
the presence of conjuncts using linguistic analysis. In: Amati sis. In: Proceedings of the 2015 17th international symposium
G, Carpineto C, Romano G (eds) Advances in information on symbolic and numeric algorithms for scientific computing
retrieval. ECIR 2007. Lecture notes in computer science, vol (SYNASC), IEEE, pp 247–250
4425. Springer, Berlin, Heidelberg. https://​doi.​org/​10.​1007/​978- Sangeetha K, Prabha D (2020) Sentiment analysis of student feed-
3-​540-​71496-5_​53 back using multi-head attention fusion model of word and-
Mladenović M, Mitrović J, Krstev C, Vitas D (2016) Hybrid sentiment context embedding for LSTM. J Ambient Intell Hum Comput
analysis framework for a morphologically rich language. J Intell 12:4117–4126
Inf Syst 46(3):599–620 Sasidhar TT, Premjith B, Soman K (2020) Emotion detection in
Mohammad SM, Turney PD (2013) Crowdsourcing a word-emotion hinglish (hindi+ english) code-mixed social media text. Pro-
association lexicon. Comput Intell 29(3):436–465 cedia Comput Sci 171:1346–1352
Moraes R, Valiati JF, Gavião Neto WP (2013) Document-level senti- Schouten K, Frasincar F (2015) Survey on aspect-level sentiment
ment classification: an empirical comparison between SVM and analysis. IEEE Trans Knowl Data Eng 28(3):813–830
ANN. Expert Syst Appl 40(2):621–633 Seal D, Roy UK, Basak R (2020) Sentence-level emotion detec-
Mukherjee P, Badr Y, Doppalapudi S, Srinivasan SM, Sangwan tion from text based on semantic rules. In: Information and
RS, Sharma R (2021) Effect of negation in sentences on

13
Social Network Analysis and Mining (2021) 11:81 Page 19 of 19 81

communication technology for sustainable development. Tang D, Qin B, Liu T (2015) Learning semantic representations of
Springer, pp 423–430 users and products for document level sentiment classification.
Shamantha RB, Shetty SM, Rai P (2019) Sentiment analysis using In: Proceedings of the 53rd annual meeting of the association
machine learning classifiers: evaluation of performance. In: for computational linguistics and the 7th international joint
Proceedings of the 2019 IEEE 4th international conference conference on natural language processing (Volume 1: Long
on computer and communication systems (ICCCS), IEEE, pp Papers), pp 1014–1023
21–25 Tao J, Fang X (2020) Toward multi-label sentiment analysis: a trans-
Sharma P, Sharma A (2020) Experimental investigation of automated fer learning based approach. J Big Data 7(1):1–26
system for twitter sentiment analysis to predict the public emo- Tiwari P, Mishra BK, Kumar S, Kumar V (2020) Implementation of
tions using machine learning algorithms. Mater Today Proc n-gram methodology for rotten tomatoes review dataset senti-
Shaver P, Schwartz J, Kirson D, O’connor C (1987) Emotion knowl- ment analysis. In: Cognitive analytics: concepts, methodolo-
edge: further exploration of a prototype approach. J Pers Soc gies, tools, and applications, IGI Global, pp 689–701
Psychol 52(6):1061 Tomkins SS, McCarter R (1964) What and where are the primary
Shelke NM (2014) Approaches of emotion detection from text. Int J affects? Some evidence for a theory. Percept Mot Skills
Comput Sci Inf Technol 2(2):123–128 18(1):119–158
Shirsat VS, Jagdale RS, Deshmukh SN (2019) Sentence level senti- Untawale TM, Choudhari G (2019) Implementation of sentiment
ment identification and calculation from news articles using classification of movie reviews by supervised machine learn-
machine learning techniques. In: Computing, communication ing approaches. In: Proceedings of the 2019 3rd international
and signal processing. Springer, pp 371–376 conference on computing methodologies and communication
Shrivastava K, Kumar S, Jain DK (2019) An effective approach (ICCMC), IEEE, pp 1197–1200
for emotion detection in multimedia text data using sequence Viegas F, Alvim MS, Canuto S, Rosa T, Gonçalves MA, Rocha L
based convolutional neural network. Multim Tools Appl (2020) Exploiting semantic relationships for unsupervised expan-
78(20):29607–29639 sion of sentiment lexicons. Inf Syst 94:101606
Singh M, Jakhar AK, Pandey S (2021) Sentiment analysis on the Xu G, Li W, Liu J (2020) A social emotion classification approach
impact of coronavirus in social life using the Bert model. Soc using multi-model fusion. Future Gen Comput Syst 102:347–356
Netw Anal Min 11(1):1–11 Yang X, Macdonald C, Ounis I (2018) Using word embeddings in twit-
Songbo T, Jin Z (2008) An empirical study of sentiment analysis ter election classification. Inf Retriev J 21(2–3):183–207
for Chinese documents. Expert Syst Appl 34(4):2622–2629 Ye Q, Zhang Z, Law R (2009) Sentiment classification of online
Souma W, Vodenska I, Aoyama H (2019) Enhanced news senti- reviews to travel destinations by supervised machine learning
ment analysis using deep learning methods. J Comput Soc Sci approaches. Expert Syst Appl 36(3):6527–6535
2(1):33–46 Zhang D, Si L, Rego VJ (2012) Sentiment detection with auxiliary
Soumya S, Pramod KV (2020) Sentiment analysis of malayalam data. Inf Retriev 15(3–4):373–390
tweets using machine learning techniques. ICT Express
6(4):300–305 Publisher's Note Springer Nature remains neutral with regard to
Strapparava C, Valitutti A, et al. (2004) Wordnet affect: an affective jurisdictional claims in published maps and institutional affiliations.
extension of wordnet. In: Lrec, Citeseer, vol 4, pp 1083–1086
Sun S, Luo C, Chen J (2017) A review of natural language processing
techniques for opinion mining systems. Inf Fusion 36:10–25
Symeonidis S, Effrosynidis D, Arampatzis A (2018) A comparative
evaluation of pre-processing techniques and their interactions
for twitter sentiment analysis. Expert Syst Appl 110:298–310

13