Content uploaded by Mohd Halim Mohd Noor
Author content
All content in this area was uploaded by Mohd Halim Mohd Noor on Jul 16, 2024
Content may be subject to copyright.
REVIEW
A survey on sentiment analysis and its applications
Tamara Amjad Al-Qablan
1
•Mohd Halim Mohd Noor
1
•Mohammed Azmi Al-Betar
2,3
•
Ahamad Tajudin Khader
1
Received: 26 May 2022 / Accepted: 1 August 2023
ÓThe Author(s), under exclusive licence to Springer-Verlag London Ltd., part of Springer Nature 2023
Abstract
Analyzing and understanding the sentiments of social media documents on Twitter, Facebook, and Instagram has become a
very important task at present. Analyzing the sentiment of these documents gives meaningful knowledge about the user
opinions, which will help understand the overall view on these platforms. The problem of sentiment analysis (SA) can be
regarded as a classification problem in which the text is classified as positive, negative, or neutral. This paper aims to give
an intensive, but not exhaustive, review of the main concepts of SA and the state-of-the-art techniques; other aims are to
make a comparative study of their performances, the main applications of SA as well as the limitations and the future
directions for SA. Based on our analysis, researchers have utilized three main approaches for SA, namely lexicon/rules,
machine learning (ML), and deep learning (DL). The performance of lexicon/rules-based models typically falls within the
range of 55–85%. ML models, on the other hand, generally exhibit performance ranging from 55% to 90%, while DL
models tend to achieve higher performance, ranging from 70% to 95%. These ranges are estimated and may be higher or
lower depending on various factors, including the quality of the datasets, the chosen model architecture, the preprocessing
techniques employed, as well as the quality and coverage of the lexicon utilized. Moreover, to further enhance models’
performance, researchers have delved into the implementation of hybrid models and optimization techniques which have
demonstrated an ability to enhance the overall performance of SA models.
Keywords Sentiment analysis Feature selection Deep learning Machine learning Optimization
1 Introduction
Social media platforms such as Twitter, Facebook, and
Instagram are important to various sectors such as the polit-
ical, social, and economic sectors. These platforms enable
people to share their thoughts with strangers over the Internet
[1]. People spend a lot of time browsing these platforms to see
the latest news and services provided. These platforms allow
them the freedom to express their opinions and to interact
with each other. Some users even took it as a way to trade
because the facility of displaying their products and the
facility of interaction with their customers.
SA is a sub-area of natural language processing (NLP)
that studies and analyzes people’s opinions on a particular
topic to create a general view from the micro-community
represented by the subscribers, potentially helping deci-
sion-makers make appropriate decisions based on these
opinions.
According to the authors [1], the term ‘‘sentiment’’ was
used in studies in 2001 by [2,3], due to these authors’
&Tamara Amjad Al-Qablan
tqablan@student.usm.my
Mohd Halim Mohd Noor
halimnoor@usm.my
Mohammed Azmi Al-Betar
m.albetar@ajman.ac.ae; mohbetar@bau.edu.jo
Ahamad Tajudin Khader
tajudin@usm.my
1
School of Computer Sciences, Universiti Sains Malaysia,
Pulau Pinang 11800, Malaysia
2
Artificial Intelligence Research Center (AIRC), College of
Engineering and Information Technology, Ajman University,
Ajman 346, United Arab Emirates
3
Department of Information Technology, Al-Huson University
College, Al-Balqa Applied University, Irbid 50, Jordan
123
Neural Computing and Applications
https://doi.org/10.1007/s00521-023-08941-y(0123456789().,-volV)(0123456789().,-volV)
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
interest in evaluating market sentiment. These were fol-
lowed by Turney and Pang et al. studies [4,5] in 2002,
which were published in the proceedings of the Association
for Computational Linguistics (ACL) and the annual con-
ference on Empirical Methods in Natural Language Pro-
cessing (EMNLP).
The researchers of SA science are interested in analyz-
ing opinions in all ways that express them, whether they
were textual [6], audio [7] or visual data [8]. Many models
were presented and developed to improve the analysis
results. Some studies classified the opinion into two pre-
defined classes, usually positive or negative [9], while
some others classified them into three categories, namely
positive, negative, and neutral [10]. In others studies, the
opinions are distinct to multi-class classification according
to the subject of the study [11].
The science of SA was used to analyze opinions on
multiple and diverse vital fields, such as a studies in health
fields to analysis the knowledge of public awareness
regarding COVID-19 pandemic [12]. This is to create
knowledge about the public opinion of telehealth and tel-
emedicine [13], or studies in economic fields, such as
studying, analyzing, and comparing the direct impact of
news spread through the media on the development of
stock market behavior [14]. In education, for example, the
some studies tried to examine the community acceptance of
e-learning as a precaution in exceptional circumstances
[15].
Generally, three extraction levels can be performed in
SA: the document, sentence, and aspect levels. In the
document level techniques, the entire document is defined
as a single entity, and the overall polarity of the document
is calculated. SA at the sentence level investigates the
sentence and produces a single opinion. In the aspect level,
the main focus remains on a specific feature of the data
[16–18].
According to the literature review, three main approa-
ches used to handle the SA process. The first approach
relies on the lexicon-based approach which is divided into
two classes: dictionary and corpus-based approaches [19].
The second approach depends on ML methods [20], while
the third approach uses the capability of DL methods [21].
Moreover, hybrid models from these main approaches were
presented to improve the analysis result. For example, in
[22], a mixed approach of lexicon and DL was presented.
Furthermore, optimization methods like using meta-
heuristics algorithms [23] are hybridized to improve the SA
process. Figure 1depicts various SA approaches and their
subcategories.
Technically, the approaches used for SA systems are
built based on a set of basic steps as shown in Fig. 2. These
steps are: data collection, preprocessing, feature extraction,
feature reduction, classification process, then performance
evaluation.
1.1 The gap analysis in the existing SA studies
Owing to the obvious enormous number of studies that
addressed the SA approaches, we have decided to conduct
a survey of studies published between 2017 and 2022.
Therefore, the primary goal of this paper is to gain a clear
understanding of the various SA approaches used
throughout human interactive applications in these years.
To achieve this goal this paper presents a more compre-
hensive study of SA concepts by discussing many research
parts related to SA, including the pipeline, SA levels, SA
growth, approaches from various viewpoints, optimization
approaches, and techniques. Moreover, this paper provides
a critical analysis of SA approaches, the major challenges,
the different sources of application, and commonly avail-
able datasets in English and Arabic languages. In addition,
it is highlighted the major differences between the emotion
recognition domain and the SA domain, with the purpose
of paving the way for future research.
As a result, the purpose of our research is to answer the
following five research questions:
Q1: What are the trends in SA from 2017 to 2022?
Q2: What are the various applications of SA?
Q3: How the SA approaches applied to different SA-
realted fields?
Q4: What data sources have been employed in the
domains of SA?
Q5: What are the main research gaps of SA domain and
what is the possible future directions to fill up these
gaps? up these gaps?
1.2 Research motivations
This paper offers a comprehensive survey of SA, surpass-
ing previous studies by encompassing a wide range of
interconnected topics within the field. With a coverage of
175 studies during the period spanning from 2017 to 2022,
this review dives into various aspects such as approaches,
applications, tools, sources, challenges, trends in SA, and
more, providing a thorough examination of SA from mul-
tiple perspectives. It serves as a valuable resource by
consolidating a cornucopia of valuable information in a
single document, making it convenient to access a com-
prehensive understanding of the field. Our work varies
from prior surveys in that it provides a clear and compre-
hensive study of the pros and cons of different SA
approaches, which may assist researchers in determining
the best approaches and tools to their approach’s
Neural Computing and Applications
123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
challenges. The main motivation for conducting the present
survey can be summarized as follows:
1. To comprehensively describe the SA approaches and
identify the essential tools needed for its implementa-
tion, a significant number of literature has been
thoroughly researched.
2. Analyzing several methodologies and approaches to
choose the best one for a certain application.
3. Classification of the prevailing SA approaches to
provide a comprehensive understanding of the existing
techniques and tools currently in use.
4. Giving a thorough overview of SA applications,
resources, and challenges in the new studies.
5. Provide clarification regarding the differences between
SA and emotion recognition.
Fig. 1 Sentiment analysis approaches
Neural Computing and Applications
123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
6. Outlined a selection of English- and Arabic-language
datasets that are currently available.
7. Outlined critical Analysis points of SA Approaches
8. A set of open issues were discussed in depth.
The remaining sections of this overview paper are orga-
nized as follows. The timeline and the growth of the SA are
presented in Sect. 2. The SA foundation is provided in
Sect. 3. A critical analysis of SA methods is illustrated in
Sect. 4. SA methods are presented in Sect. 5. The SA has
been widely used for lots of applications, which are
extensively summarized in Sect. 6. Section 7provides a
selection of datasets that are readily available in English
and Arabic. Section 8presents the conceptual differences
between SA and emotion recognition. In Sect. 9, a set of
open issues associated with SA is presented. The paper
concludes in Sect. 10, where the answers to the research
questions are provided, and several future directions and
points of interest are highlighted for further investigation.
2 Growth of sentiment analysis
SA is one of the rapidly growing research areas in com-
puter science. In this section, we present the growth and
progression of SA studies through different views pub-
lished in the past seven years. The presented results are
based on the Scopus database.
SA has gained considerable popularity recently.
Research on SA has been published in respected journals
with high-prestige publishers such as Elsevier, Springer,
Institute of Electrical and Electronics Engineers Inc.,
MDPI, and many others, as shown in Fig. 3.
Figure 4shows the growth of SA based on the topic
versus the number of publications per topic. The highest
number of publications is in the computer science field, at
7962. The engineering field has also obtained a massive
publications number, at 3245. In Mathematics topics, 1989
documents were published. SA techniques were also pub-
lished in fields such as decision sciences, social sciences,
physics and astronomy, as shown in Fig. 4.
The top 12 authors who published several articles in SA
are given in Fig. 5. Cambria, E. has the most articles with
72 articles recorded in the Scopus database.
Nanyang Technological University in Singapore has the
most interest in SA with 104 published articles in various
domains. The institution with the second-highest interest in
SA, with 85 published articles in SA, is the Chinese
Academy of Sciences. The Vellore Institute of Technology
in India takes third place. The top 10 institutions are pre-
sented in Fig. 6.
Fig. 2 SA approaches steps
Fig. 3 Number of SA publications per publisher (Source: Scopus)
Fig. 4 Number of SA-based publications published by each subject
(Source: Scopus)
Fig. 5 Number of SA-based publications published for the top 12
authors (Source: Scopus)
Neural Computing and Applications
123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
The growth of SA is increasing rapidly year by year as
shown in Fig. 7, which shows the progress of SA based on
the number of articles published annually.
In terms of countries, India has the most interest in SA,
and it has published more than 2200 articles in different
domains while China has the second rank with more than
1900 published articles in SA. Figure 8shows 20 countries
sorted in descending order based on the number of articles
published.
The progression of the SA based on citations number
appears ascending when reviewing literature articles. Fig-
ure 9shows that the citations are doubled from year to
year. These data were extracted using the Scopus database
on July 2022.
3 Sentiment analysis
This section presents a definition of SA and gives an
overview of the basic analysis steps, including the pre-
processing, feature extraction, and feature selection steps.
3.1 SA definition
SA (also known as opinion mining) is an active research
area in NLP [24]. It aims at identifying, extracting, and
organizing sentiments from user generated texts, audio, or
visual data in social networks. The importance of SA
comes from its ability to give an overview or a general
sentiment on a specific topic, potentially helping decision
makers to make the right decisions. Many studies have
presented various approaches to SA, each dealing with the
sentiment classification problem differently. In summary,
SA can be done at three levels: document, sentence, and
aspect levels [16]. A brief description of each level is given
below:
3.1.1 Document level
The task of this level is to classify the whole document and
extract an overall view. We consider the following docu-
ment as an example, ‘‘Overall, I believe the most recent
phone I purchased is excellent. It comes with a high-res-
olution camera. The phone is extremely slim and fits easily
into a pocket. The price is a little high given the configu-
rations. However, the phone seems really good.’’ The
document as a whole appears to be more positive than
negative. So, if one word must be used to describe the
review, it is positive.
Fig. 6 Number of SA-based publications published by each institu-
tion (Source: Scopus)
Fig. 7 Number of SA publications per year (Source: Scopus)
Fig. 8 Number of SA publications of each country (Source: Scopus)
Fig. 9 Number of citations in SA publications (Source: Scopus)
Neural Computing and Applications
123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
3.1.2 Sentence level
At this level, the task is to examine the sentences and
determine whether each one reflects a positive, negative, or
neutral viewpoint. Considering the example presented in
the above subsection, the document can be divided into five
sentences, each of which can be classified as positive,
negative, or neutral. Here, opinions become slightly more
specific.
3.1.3 Aspect level
This level aims to find sentiments concerning the particular
aspects of words [25]. For example, we consider the fol-
lowing sentence, ‘‘I like the new design of the Instagram
App, but I hate its customer policy.’’, the review is on
design and customers policy, which are two aspects of the
product Instagram App, and the review is positive for the
design aspect and negative for customers policy. The task
at this level aids in determining exactly what people like or
dislike. It concentrates on the characteristics of features
rather than the sentiment of paragraphs. Implicit or explicit
aspect extraction can be considered the core task for SA
[26].
3.2 SA formulation
The generic SA process is illustrated in Fig. 2. The basic
steps involve pre-processing, feature extraction, and fea-
ture selection. Then, the classification step can be per-
formed using different SA approaches (e.g., ML, DL ). The
steps present in this section are generally used for text
analysis. However, with few modifications, we can use the
same curriculum for audio and image data. Knowing that
deep learning approaches can skip the first three steps and
go straight to classification, they can self-learn features
from the dataset. By contrast, these approaches are com-
plicated and computationally expensive. A brief illustration
of the pre-processing step, feature extraction step, and
feature selection step is presented below.
3.2.1 Pre-processing step
The next step after collecting data are a pre-processing.
The importance of this step comes from removing unnec-
essary data to lower the size of the input raw dimension,
enhancing analysis. The data are cleaned by deleting
unnecessary words, URLs, stop words, punctuations, and
other elements that are required to be reduced before fea-
ture extraction step. Sets of pre-processing techniques were
illustrated in [27]. The following are common pre-pro-
cessing operations:
1. Remove URL
2. Remove hashtag (i.e., #happy or replace it with happy)
3. Remove mentions (i.e., @BarackObama)
4. Remove white spaces and quotations
5. Remove vowels repeated (i.e., looooove )
6. Replace emoticons into tags that express their senti-
ment (i.e., replace:) with a smile or replace it with a
positive word)
7. Deal with negations (i.e., replace can’t, don’t, never
with not)
8. Deal with slang through slang dictionaries (i.e., replace
l8 with late)
9. Apply stemming process (i.e., Replace words like
‘‘great,’’ ‘‘greatly,’’ ‘‘greatest,’’ and ‘‘greater’’ with the
single word ‘‘great,’’ reducing entropy and improving
the relevance of the concept of ‘‘great’’)
Indeed, several studies have explored the effects of text
preprocessing on classifier prediction accuracy and com-
putation time [28–32]. Based on the findings from the study
[28], it is evident that text preprocessing techniques have a
positive impact on classifier accuracy for SA in an English
dataset. The study observed improvements in the accuracy
of SVM classifiers from 81.09% to 81.63% after applying
preprocessing techniques. Similarly, the accuracy of NB
classifiers increased significantly from 83.69% to 91.81%
with the use of preprocessing. These results emphasize the
importance of text preprocessing in improving the perfor-
mance of classifiers in SA tasks. Regarding the study [29],
which focused on comparing preprocessing techniques on
Twitter data, the outcomes highlight the advantages of the
stemming technique, particularly in terms of computational
speed. Stemming, which reduces words to their base form,
demonstrated superior computational efficiency compared
to other techniques such as lemmatization and spelling
correction. The study suggests that stemming not only
effectively captured sentiment information but also sim-
plified the data representation, resulting in reduced pro-
cessing time for SA on Twitter data. Overall, these studies
provide insights into the benefits of text preprocessing
techniques in SA tasks. Preprocessing helps to improve
classifier accuracy and can offer computational advantages,
such as faster processing speed. However, it is important to
consider the specific requirements of the task and the
characteristics of the dataset when choosing the most
suitable preprocessing techniques.
3.2.2 Feature extraction step
As we know the ML and DL models cannot process text, so
we need to figure out a way to convert these textual data
into numerical data. Text feature extraction can be defined
as the process of extracting a list of valuable words from
Neural Computing and Applications
123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
text data and converting them into a feature set that a
classifier can use. The approaches listed below are com-
monly used to extract and represent features from text data.
1. Bag of Words (BoW)
BOW is the most used technique for converting text to
numerical representation, where a vector of integer num-
bers is generated representing the occurrence of words
within a text. It is a simple way to represent the vector of
data. However, the order or structure of words in a text is
ignored while using this technique.
BoW can be formulate as follows: let A and B two
sentences n(A) represent set of words in sentence A, and
n(B) represent set of words in sentence B then BoW
ðA[BÞ¼nðAÞþnðBÞnðA\BÞ. Suppose a new sen-
tence C is added then BoW
ðA[B[CÞ¼nðAÞþnðBÞþnðCÞnðA\B\CÞ.
For example, we consider the following reviews:
Review 1: ‘‘This car is very slow’’
Review 2: ‘‘I like this car’’
From Table 1, we can observe that the vector length for
all reviews is constant.
2. Term frequency-inverse document frequency (TF-IDF)
Given that the BoW technique assesses the repetition of a
word in a document without taking into account its rele-
vance to the whole corpus, a modification on it called term
frequency-inverse document frequency (TF-IDF) was
developed to calculate the word’s importance for the cor-
pus as a whole.
The term frequency concept for word (i) (TFwi) means
the number of times a word (i) appears in a document d
divided by the total number of words in the document, as
seen in equation 1.
TFwi¼Pðwi;dÞ
Pðw;dÞð1Þ
The inverse document frequency (IDF) concept is used to
determine the significance of a word. The IDF equation is
IDFðtÞ¼log N
Dtf
;ð2Þ
where N denotes the number of documents and Dtf is the
number of documents that include word t [33,34].
The TF_IDF equation appears in 3
TF IDF ¼TF IDF ð3Þ
For example: assuming a document ‘‘I like that car, which
car do you mean? the white car’’ has 12 words and the
word car appears thrice times in those 12 words, the term
frequency will be 3/12=0.25. If there are 50,000 docu-
ments, only 500 of them contain car, then the IDF (car)=
log(50,000/500) = 2, and the TF-IDF (car)= 0.025*2 =
0.05.
3. Word Embedding
Another common feature extraction technique is word
embedding, which is a modeling language and feature
learning method that converts words in a document into
continuous real-number vectors. Word embedding a dis-
tributed representation system of the documents’ vocabu-
lary in which words with comparable meanings are in a
similarly represented [16]. Word2Vec, which was devel-
oped by [35], is a popular technique to learn word
embedding using neural networks. It includes two models,
namely skip-gram and continuous bag-of-words (CBOW).
Both techniques depend on the likelihood of words
occurring close together [36]. Despite Word2vec has good
lexical performance, it is limited by the characteristics of
local semantics, making it challenging to effectively utilize
the global lexical cooccurrence statistics. Another embed-
ding technique is called GloVe [37], which combines both
local and global semantic information, they capture not
only syntactic relationships between words but also
semantic relationships between words with similar con-
textual usage [38]. A comparison that compared GloVe
with the word2vec for text feature extraction was done in
[37,39], through the experiments, it was proved that GloVe
has a better effect in text classification than the word2vec
technique. Another popular word representation technique
called FastText was created by the Facebook AI Research
lab [40]. It expands on the concept of word embeddings by
supposing that each word is represented as a bag of char-
acters n-gram, which will help to provide vector repre-
sentations for words that are not part of a vocabulary list
[41].
A comparison between Word2Vec, Glove, Glovepre,
and FastText techniques was done in [42]. The results
showed that FastText outperformed other methods in terms
of concerning time consumption for training the classifier.
Furthermore, a comparison between three DL models
DNN;CNN ;andRNN with TF-IDF and Word embedding
for different datasets revealed that combining DL tech-
niques with word embedding is better than TF-IDF when
performing SA [43].
Table 1 Example of BoW representation of two reviews
Review This car is very slow I like
Review
1
1111100
Review
2
1100011
Neural Computing and Applications
123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
A new concept of word embedding techniques recently
appeared called contextualized word embeddings [44].
This concept appeared simultaneously with the emergence
of large language models (LMs), which have been widely
used for various NLP tasks, including word representation
and feature extraction. These models, such as GPT
[45,46], BERT [47,48], OpenAI’s GPT [49], and many
others [50], have been trained on vast amounts of text data
and are capable of learning rich and contextualized word
representations. The embedding space of a large language
model can capture various semantic and syntactic rela-
tionships between words resulting in significant improve-
ments in various NLP tasks [51,52]. Indeed, pre-trained
LMs can be used by researchers to extract features from
words that capture their semantic and contextual informa-
tion, these features can then be used as inputs for a wide
range of downstream tasks, including SA tasks [53].
3.2.3 Feature selection step
A set of irrelevant and redundant features usually occur
among the set of features obtained in the previous step.
Thus, we need feature selection techniques to remove or
reduce these undesirable features to reduce the computa-
tional cost and improve the performance of SA classifica-
tion [54].
Many existing techniques were used. The following
paragraphs review some strategies for selecting the
features.
1. Wrapper Methods
The wrapper model uses a subset of features to
train a learning model (i.e., decision tree (DT), naive
Bayes (NB)) and uses a search strategy like: sequential
forward selection (SFS) and sequential backward
selection (SBS) [55] to search across the space of
possible feature subsets then evaluates each subset
based on the quality of the algorithm’s performance.
The basic steps of the wrapper method are:
searching for a subset of features, building a ML
model, and evaluating the model’s performance. Then,
the whole process is repeated and started again with a
new subset of features until some stopping criterion is
satisfied.
The model uses a greedy search strategy because
it wants to obtain the best suitable combination of
features that will result in the best performing model.
However, this leads to a significant computational cost
and overfitting [56,57].
2. Filter Methods
The filter model is designed in a way that makes
feature selection and model learning independent
[58,59]. Information gain, relief method, Fisher score
method, and gain ratio are examples of filter-based
approaches that have been widely used [60]. The
features are assessed using statistical measurements
without taking into account the utilization of interre-
lationships between them. This creates a risk of failure
in identifying redundant features, resulting in low
stability and performance [56].
3. Embedded Methods
In this approach, the feature selection technique is
embedded into the learning process. In more detail, the
learning and feature selection processes are carried out
simultaneously. As a result, it might be considered a
smarter approaches than filter strategies. Given that it
interacts with the prediction model, it is also faster than
the wrapper method because the learning model does
not need to be trained every time a feature subset is
picked [61]. Embedded approaches are also less
computationally expensive than wrapper approaches
since the search for the optimum subset of features
takes place during the classifier’s training (for example,
when improving weights in a neural network) [62].
LASSO and RIDGE regression are some of the most
popular examples of embedded methods [63].
4. Hybrid Methods
Generally, hybrid methods combine diverse
approaches to give results that are more accurate and
obtain the best feasible feature subset. This method
tries to combine the strengths of models (e.g., wrapper
and filters) by utilizing their different evaluation
criteria at various phases of the search process [64].
Usually, two or more feature selection methods of
various conceptual origins are combined [62]. For
instance, one could remove a set of features with a less
computationally expensive filter method then find the
best candidate subset features with a more computa-
tionally expensive wrapper method.
5. Optimization Methods
Different metaheuristic strategies to solve the
feature selection problem were implemented to obtain
the best possible feature subset. These algorithms
formulated the feature selection issue as an NP-hard
optimization, so the use of metaheuristic strategies is
very appropriate to solve the feature selection problem
[65–69]. The genetic algorithm (GA), harmony search
(HS), and particle swarm optimization (PSO) are
examples of these algorithms.
Neural Computing and Applications
123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
4 Critical analysis of sentiment analysis
approaches
The twenty-first century has seen a flood of data. Over the
last two decades, data have exploded in various fields,
leading to a huge increase in data amount [70]. The digital
world’s infusion of technology has paved the way for the
growth of big data. People throughout the world are
becoming more electronically intelligent, using equipment
such as digital sensors, communication tools such as social
media applications, actuators, and data processors [71]. In
the age of big data, the use of SA has shown to be a
beneficial for categorizing public opinions into different
sentiments and assessing public mood [72].
In this section, we will present a set of critical points
concerning different SA approaches. A summarized over-
view of the SA approaches, along with their respective
advantages and disadvantages, can be found in Table 2.
Furthermore, Table 3provides a systematic comparison
among the three approaches.
As illustrated in these tables, the lexicon/rule-based
approaches utilize predefined sentiment dictionaries or
rules to assign sentiment scores. These approaches are
simple to implement, computationally efficient, and
domain-specific. However, they may struggle in capturing
nuanced sentiments and require adding new words and
terms to lexicons for new domains. On the other hand, most
ML approaches learn from labeled data and offer higher
accuracy. However, they require substantial training for
each domain data and need to retrain models for each new
domain. DL approaches, in contrast, possess the ability to
automatically learn complex patterns, capture contextual
information, and show impressive performance. However,
they require extensive resources and can be less inter-
pretable. The selection of an appropriate approach is con-
tingent upon specific requirements and the resources at
hand. Additionally, hybrid models can be employed to
capitalize on the strengths of different approaches.
4.1 Lexicon-based approach
The lexicon-based approach is considered unsupervised
approach [73]. It has two main types: the dictionary and
corpus-based approaches. The general stages of this
approach are as follows. Data are extracted from the text
reviews. Then pre-processing steps are applied to the
Table 2 Summary of the advantages and disadvantages of the SA approaches
SA Approach Type Advantage Disadvantage
Lexicon-based
approach
Manually-based approach More accurate compared to
other lexicon-based approach
- Time and human resource consuming
Dictionary-based
approach
Does not require training or labeling data
- The learning process simple and fast
- Easy to use for multilingual SA
- Domain dependence
- Cannot deal with different slangs
- Cant deal with acronyms
- Falling in translating problem
when used for multilingual SA
Corpus-based approach Independent of a domain
- Can deal with different slangs
and acronyms
- Need a huge corpus to cover all terms
which is a challenging effort
ML-based approach Unsupervised
Semi-supervised
Supervised
Dictionary is not required
- Demonstrate high precision of
classification
-Domain classification
- Need to retrain models for each new domain
- Features extraction techniques must be used
- Supervised and semi-supervised approaches
required labeled data
DL-based approach Unsupervised
Semi-supervised
Supervised
Dictionary is not required
- Demonstrate high precision of
classification
- Features extracted automatically
- No need to use features extraction
techniques
- The extracted features are harder to
assess and understand
- Required a huge amount of data to
achieve accurate results
- Collecting and labeling vast amounts
of data can be extremely difficult and
time-consuming
Neural Computing and Applications
123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
extracted data. Finally, lexicon is applied to the processed
data to determine the polarity, as shown in Fig. 10.
These approaches can be constructed manually or
automatically. Although manually constructed lexicons are
clearly more accurate than automatically constructed ones,
they are a time-and human resource-consuming procedure.
As a result, many studies rely on the automatic lexicons,
which are constructed utilizing semantic relations of
thesauri, morphological techniques, or co-occurrence
algorithms in large corpora [74]. In the dictionary-based
approach, a small set of words is picked and subsequently
expanded with the help of information from dictionaries.
This method’s main proposition is that if a word has sen-
timent polarity, then its synonyms and antonyms do as
well. As a result, a new, more complete, set of opinion
words can be created from the initial set of words. In this
Table 3 Comparative evaluation of the SA approaches
Lexicon-based approach Machine learning-based approach Deep learning-based approach
Type Dictionary and corpus-based Supervised, Sime-supervised,
and Unsupervised
Supervised, Sime-supervised,
and Unsupervised
Labelled Data No need Supervised and semi-supervised
approaches required labelled data
Supervised, semi-supervised
approaches
required labelled data
Feature extraction Manual feature engineering Automated feature extraction Automated feature extraction
Time to train the data No need Need time Need time
Complexity Simple and fast Depending on the algorithm
employed
Due to the deep network architecture,
it is complex
Domain dependency Lexicons are usually designed for
specific domains
Domain dependence, but they are
often more malleable and adaptable
than lexicon-based methods
More Flexible and adaptable than
lexicon-based and ML approaches
Adaptability Requires adding new words and
terms to lexicons for new domains
Flexibly adapt to a wide range of
domains with adequate training
Flexibly adapt to a wide range of
domains
with adequate training
Performance Depends on the lexicon scope Depends on the quality of labelled
data
Depends on the quality of labelled data
and the deep network architecture
Applications Simple SA tasks Various SA tasks Various SA tasks with massive data
Out-of-vocabulary
words
May have trouble with words that
are not in the lexicon
Can to a limited degree handle
out-of-vocabulary words by
utilizing n-gram representations
Contextual Information, capturing
syntactic and semantic contexts provide
additional contextual clues to handle
Out-of-Vocabulary Words effectively
Amount of training
data
It is dependent on pre-defined
lexicons
Requires an adequate quantity of
labelled training data
Requires a huge quantity of
labelled training data
Fig. 10 Flowchart for Lexicon-
Based Approach
Neural Computing and Applications
123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
approach, the computational methods are performed on a
text for determining how the text affects the reader’s
emotions. The performance of a dictionary-based approach
is strongly dependent on the dictionary chosen. As a result,
we must be careful before deciding whether the dictionary
is acceptable for our dataset and our purposes. The second
approach is looking for rules and patterns in the texts. The
co-occurrence statistics of several words are compared to
determine word polarity. Its methods rely on patterns that
occur together, as well as a seed list of opinion words.
Finally, the word orientation is used to categorize reviews.
The main advantage of the lexicon based-approach is
that it does not require any training or labeling data,
making the learning process more simplified and faster. By
constrast, the limitation of the dictionary-based approach is
a domain dependence. Extracting domain-specific opinions
using it is not possible. Given that words can have multiple
meanings, a positive word in one domain might be a
negative word in another. To avoid this limitation a corpus-
based approach is used. However, the corpus-based
approach is not as efficient as the dictionary-based
approach due to the need for a huge corpus to cover all
terms, which is a challenging effort.
4.2 Machine learning-based approaches
Three types of ML approaches are commonly used for SA:
supervised, unsupervised, and semi-supervised approaches.
The key difference between these approaches is the amount
of labeled data available, which denotes the understanding
of what the model’s output should be for a given input [75].
The major task of supervised learning approaches is to
build a model that maps the input features to output using
labeled data [76,77]. Training these models often requires
a huge amount of labeled data during training. Further-
more, this process can be costly and time-consuming,
bringing challenges in terms of both financial resources and
the time required to collect and label the data [78,79]. On
the contrary, unsupervised learning models are imple-
mented without the need for labeled data during training. In
these approaches, the model autonomously explores and
analyzes the patterns and structures inherent in the given
data to discover and understand the underlying patterns,
clusters, and relationships within the data without any prior
knowledge of labels or specific output expectations
[80–82]. Different clustering techniques are used in these
approaches such as k-means, Euclidean co-clustering
(ECC), and a graph co-regularized nonnegative matrix tri-
factorization (GNMTF). Among the clustering methods,
k-means is the commonly used clustering method due to its
simplicity and efficiency [83]. The unsupervised approa-
ches have the advantage of not requiring a labeled dataset,
but the SA results, on the other hand, are less accurate and
trustworthy methods because the models cannot be evalu-
ated. Semi-supervised models lie in the middle between
supervised and unsupervised models, which are beneficial
when labeled data are scarce or expensive to obtain, where
the model can utilize the abundant unlabeled data to
improve its performance [84,85], it is operated with a
combination of labeled and unlabeled data. These models
might incorporate the unlabeled data to cause more regu-
larization while using the labeled data for supervised
training. For example, in [86], the authors manually labeled
10% of the dataset and used it to validate the automatically
generated labels by TextBlob for the remaining 90%.
Table 4summarizes the main differences between these
three approaches.
In contrast to the lexical approach, which does not
require time to train the data, the supervised and semi-
supervised ML approach does. Accordingly, in all ML
approaches, data are separated into training and testing
datasets. Furthermore, these approaches required a set of
processing steps to extract and select the more relevant
features before training the classifier. The use of powerful
feature extraction and feature selection techniques can help
these approaches perform well. Another challenge associ-
ated with ML approaches is domain classification, which
refers to the need for retraining models for each new
domain. Figure 11 shows a flowchart for a typical ML
approach that can be used for SA.
4.3 Deep learning-based approaches
Given the vast amount of text data on the web, DL models
such as LSTM and CNN are quickly gaining popularity for
sentiment classification. The fundamental benefit of a DL
model over other models is that it learns on its own. It has
the the ability to extract features automatically from the
raw text data, so it does not need the feature extraction
steps like BOW and TF-IDF. However, from a different
perspective, these extracted features are harder to assess
and understand. One of the limitations of DL approaches is
the inadequacy of labeled data. These approaches required
a huge amount of data to achieve accurate results. Col-
lecting and labeling vast amounts of data can be extremely
difficult and time-consuming. DL approaches also have the
disadvantage of being financially and computationally
costly because they require considerable computing time
when compared with approaches based on ML or lexicon-
based approaches. Figure 12 shows a flowchart for a DL
approach that can be used for SA.
Neural Computing and Applications
123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
5 Sentiment analysis approaches
Several approaches of SA are reviewed in this sec-
tion. Traditional approaches are covered in the first sub-
section, ML approaches in the second, DL approaches in
the third, optimization approaches in the fourth, and hybrid
approaches in the last subsection.
5.1 Traditional approach of SA
Traditional SA involves using a pre-prepared sentiment
lexicon to calculate document’s score through the senti-
ment scores of all the words in the document. The lexicon-
based approach can be divided into two categories: dic-
tionary and corpus-based approaches [87]. In this section,
we will look at some recent research on lexicon-based
approaches.
Table 4 Comparison between supervised, unsupervised, and semi-supervised approaches
Supervised Unsupervised Semi-supervised
Data Labeled data Unlabeled data Partially labelled
Process Classification,
Regression
Clustering, Dimensionality
reduction
Combination of supervised and
unsupervised techniques
Common algorithm NB, SVM,RF, NN K-mean, Gaussian mixture models,
Principal component analysis
Semi-Supervised SVM, MNB with
Expectation Maximization
Drawback Time-consuming to train, labels
variables require expertise
Wildly inaccurate results Insufficient labeled data can result
in highly inaccurate outcomes
Complexity Less Computational
Complexity
More Computational Complexity Relies on the quantity and quality
of the labeled data provided
Output Output is a predictive mode Output is a data structure or pattern Output In the middle ground between
supervised and unsupervised learning
Fig. 11 Flowchart for machine learning-based approach
Neural Computing and Applications
123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Assiri et al. [88] constructed a sentiment lexicon and a
weighted lexicon-based algorithm for Saudi dialects. The
proposed algorithm looks for correlations between polarity
and non-polarity terms, then weights them based on those
relationships. To improve the performance of their algo-
rithm, the authors also presented a set of rules for effi-
ciently handling negation and supplication linguistic
features.Two real datasets were used to test their algorithm:
the Saudi Twitter dataset and the El-Beltagy and Ali
dataset. In [89], a mixed lexicon of two languages, English
and Malay, was created to developed a cross-lingual SA
approach for positive and negative tweets. The proposed
approach has average accuracy of roughly 55%. The
authors justify this result because their approach cannot
analyze slang dialect or detect the extensively used short-
ened words on social media. Yurtalan et al. [90] proposed a
lexicon-based polarity determination and calculation
approach for Turkish, based on examining a Turkish tweets
dataset at the word, word group and idiom levels. The
proposed system was tested on three datasets, and an
accuracy of 88% was observed. The authors concluded that
analyzing the system at the word group level enhances its
overall performance by more than 12%. Wunderlich and
Memmert [91] presented a study that tested the feasibility
of using a lexicon-based approach to analyze sentiment
regarded to football sport on twitter platform. A new
contribution for Urdu SA was done in [92], where
researchers improved the Urdu lexicon and the Urdu SA.
The new combination was effective because they devel-
oped a wide coverage lexicon, effective handling of
negations, intensifiers and context-dependent words by the
Urdu SA. The algorithm for Urdu SA was implemented in
four phases, one after another. In the first phase, only
positive and negative words were considered, and in phase
two, negation handling was implemented. Intensifiers were
dealt with in phase three, while context-dependent terms
were dealt with effectively in phase four. In this approach,
sentences were passed at each stage as input and each
sentence was polarized after being searched in several files
such as positive, negative, intensity, negation, and context-
dependent terms. All polarities were summed at the end of
the sentence processing. Then, the sentence was regarded
positive if the total was larger than 0 and negative if the
total was less than 0. If the sum is zero, the sentence was
regarded as neutral. The study in [93] focused on the illegal
immigration issue, an enhanced lexicon-based approach
was employed using an incorporated the General Inquirer
as the sentiment lexicon with multi-level grammatical
dependencies and the role of verb. For the data gathered
from Twitter, result showed that the enhanced approach
outperformed ten online SA tools with accuracy of roughly
81% and 82% on two different datasets containing 2,500
tweets collected from Twitter for a period of three months.
Comparisons between the lexicon-based and ML
approaches were carried out in [73,94,95]. Aloqaily et al.
[94] used a lexicon-based approach for Arabic tweets
datasets concerning the Syrian civil war and crises.
Researchers outlined the differences between short and
long Arabic text SA. The results of this study demonstrated
that the ML approach performs better than does the lexi-
con-based approach. In terms of ML, five classifiers were
used: logistic model trees (LMT), simple logistic, support
vector machine (SVM), DT, voting-based, and k-nearest
neighbor (KNN). Even though this study did not address
the dialectical Arabic or other aspects that could affect SA
efficiency, such as intensifying and negating, the best
Fig. 12 Flowchart for deep learning-based approach
Neural Computing and Applications
123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
results with accuracy of 85.5% was achieved by LMT
method. By contrast Piryani etal. [95] compared the lexi-
con-based and ML approaches on Nepali tweet datasets. In
the ML approach, the polarity was classified using four ML
classifiers: NB, SVM, LR, and DT. SVM achieved the best
accurate result among all other classifiers. The hybrid
lexicon approach, which combined three lexicons, namely
NRC emotion lexicon, DSSL, and Nepali SentiWordNet,
outperformed the ML approach. The study also indicated
that the TF-IDF outperformed the BOW and the N-gram
feature extraction techniques. In [73], in the case of the
lexicon-based approach, SA was performed using a wide
coverage Urdu sentiment lexicon and efficient Urdu sen-
timent analyzer. Given that the negations, intensifiers and
context-dependent words issues were taken into consider-
ation the accuracy improved from 74% to 89%. By con-
trast, in the ML approach, three classifiers were applied,
namely SVM, DT and KNN. The lexicon-based approach
was better than the ML approach in terms of accuracy,
precision, recall, and F-measure, where the accuracy was
89% with 86% precision, 90% recall, and 88% F-measure,
compared with 67% accuracy and 68% precision, 67%
recall and 67% F-measure in the ML approach. The author
explained the results were as demonstrated due to the wide
coverage for the lexicon and the analyzer that can effi-
ciently handle data from different domains.
We conclude from previous studies that a lexicon-based
method relies on the language structure and relationships
between words. It is based on determining how distinctly a
word is skewed toward positive or negative significance
[96]. Furthermore, to provide satisfactory results, the issues
of negations, intensifiers, and context-dependent words
must be considered. As far as we know, no general struc-
ture has been developed for lexicons to deal with SA of
languages.
5.2 Machine learning-based SA approaches
ML can be defined as the automatic improvement of a
computer’s learning process based on previous computer
experiences. For SA, three types of ML techniques are
generally used: supervised, unsupervised, and semi-super-
vised learning approaches [97]. Supervised approaches,
which are the most common methods, were used in the
following studies [28,98–113]. Data are split into two sets:
training data and testing data. In unsupervised approaches
[114,115], prior training is not needed to extract the data
[96], so training data are unnecessary. Semi-supervised
approaches [116,117] may learn from labeled and unla-
beled data. Although unlabeled data have no notion about
classes, they can include information about joint distribu-
tion across the classification features. As a result, when
there is a scarcity of labeled data in the targeted data
domain, employing a semi-supervised technique with
unlabeled data can outperform supervised learning [118].
Regarding the aforementioned studies, SVM, NB, Random
Forest (RF), and KNN are the most broadly used classifi-
cation techniques in the supervised approaches. SVM is
estimated to be the most efficient of these techniques.
However, the unsupervised learning technique gives a less
accurate classification result than the supervised technique.
It also needs a large amount of data to be trained correctly,
but they do not need labeled data to train their models,
which is considered expensive and time-consuming to
annotate. We take a look at some of these studies as
follows.
Four classifiers were used in [98] for SA, namely NB,
BFTree, J48, and OneR. OneR is more accurate, but NB
had a faster training rate. The experiments were applied on
three datasets, one of which from the benchmark IMDB SA
dataset, which was utilised for the first time in this research
[119]. This dataset has 50,000 movie reviews for binary
SA. In [108], the three classifiers NB, KNN, and RF were
implemented to identify the polarity of movie reviews’
tweets. The data of 2000 reviews were collected form an
IMDB benchmark dataset and imported to the WEKA tool
to perform text pre-processing steps before SA classifica-
tion. NB classifier gave the best results with 81.45%
accuracy compared with 78.65% of the RF classifier and
55.3% of KNN classifier. In [101], the SVM classifier was
selected and applied to 6000 gathered and manually labeled
tweets relating to HPV vaccines. The authors indicated that
SVM’s superior results compared with algorithms NB and
RF in their earlier experiment were the reason why they
chose it.
Mukhtar et al. [102] presented the Urdu SA technique
using the KNN, DT, and SVM classifiers. The classifiers’
performance was not sufficient with less than 50% accu-
racy. The findings were thoroughly reviewed and modifi-
cations were made, including the extraction of features that
enhanced the rating performance to a satisfactory result. In
terms of accuracy, precision, retrieval and f-measurement,
KNN was better than SVM and DT. Mukhtar et al. [73]as
stated in the preceding subsection, presented another SA
study that compared the supervised ML approach with a
lexicon-based approach, concluding that the lexicon
approach outperformed the ML approach not only with
regard to accuracy, precision, and recall measures, but also
with regard to time savings and efforts used. The authors in
[106] applied three feature selection strategies, namely
information gain (IG), Chi-square, and Gini index, to create
different feature subsets on three datasets: IMDb, elec-
tronics and kitchen product. Then, the three feature subsets
were combined using UNION, INTERSECTION, and
revised UNION statistical methods to obtain a top-ranked
feature list for each dataset in order to improve the
Neural Computing and Applications
123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
performance of four ML classifiers: SVM, multinomial
naive Bayes (MNB), RF, and logistic regression (LR). The
results showed that the SVM outperformed MNB, RF, and
LR classifiers on the movie (IMDb) review dataset, with
the highest accuracy of 92.31%.
Alam and Yao conducted an intriguing study in [28]to
investigate the effect of text pre-processing on the accuracy
of the NB, SVM, and Maximum Entropy (MaxE) ML
algorithms. The study was conducted on a collection of 359
reviews gathered through the KNIME Twitter API. Their
experiments showed that the accuracy of the NB algorithm
is greatly increased after pre-processing steps with an
improvement of 8.12% from 83.69% before pre-processing
steps to 91.81%, while the accuracy of the SVM algorithm
is slightly improved by 0.54% and no increase in accuracy
was observed with the MaxE method. ML was also applied
to predict cryptocurrency price market movements in
[105]. The authors used two resources of data: Twitter data
and market data. Neural Network (NN), SVM and RF were
used to estimate the prices of Bitcoin, Ethereum, Ripple,
and Litecoin cryptocurrency market movements. Three
models were used to conduct their research. In the first
model, each algorithm was trained only with Twitter data,
in the second with market data, and in the third with both
Twitter and market data. The authors claimed that using
both Twitter and market data for the training process
improved the performance of all algorithms. The results
also indicated that in the cases of Bitcoin, Ethereum, and
Ripple, NN outperformed the other models, while SVM
outperformed the others in the case of Litecoin.
Furthermore, a technique for analyzing sentiment in
scientific text was presented in [112]. The uni-gram, bi-
gram, and tri-gram features were applied to aid the SVM,
NB, DT, LR, KNN and RF classifiers in achieving the
maximum levels of accuracy. SVM outperformed the other
classifiers, with an accuracy of 88%, followed by NB. In
[99], the authors constructed a predictive education
framework to assess student satisfaction with 249 MOOC
courses from Class Central, as well as the opinions of 6,393
students on these courses. The authors used gradient
boosting trees for applying SA and hierarchical linear
models in their framework, and demonstrated the impor-
tance of SA in predicting student satisfaction. The authors
in [107] highlighted the aspect level analysis for Amazon
product reviews to identify the opinions about specific
aspects that were either preferred or not. They employed
the SVM classifier with three kernels: linear, polynomial,
and radial basis function (RBF). The authors in [109]
employed SA of tweets in Malay to classify them into
positive and negative sentiment. Three ML algorithms,
namely NB, SVM, and RF, as well as four different doc-
ument features, namely BOW, TF-IDF, unigram with
Senti-Wordnet, and unigram with Senti-Wordnet with
negation words were used. The authors found that the RF
outperformed other ML classifiers, with the best result
being 95.6% accuracy using unigram Senti-Wordnet
including negation words. In [111], three ML algorithms,
namely SVM, RF, and MNB, were utilized to develop a
model of SA on land transportation infrastructure. SVM
outperformed all other classifiers, with the maximum
accuracy of 76.12%.
A semi-supervised model for predicting sentiment dis-
tributions based on recursive neural network (RecursveNN)
was proposed in [117]. The approach used the recursive
nature of sentences to learn vector space representations for
multi-word phrases. The authors in [114] looked into a
number of unsupervised SA approaches. They constructed
word clusters and used them as features in a SA problem
based on aspects. They also employed HPS, which is an
unsupervised stemming technique that aided the Czech
language’s complex morphology. Also, an unsupervised
ML technique was applied in [115] to help locate relevant
keywords related to the main topic to analyze opinions and
detect tweet polarity. The authors used the Apache Storm
tool to created a real-time system to track Twitter
sentiment.
5.3 Deep learning-based SA approaches
DL is a subset of ML. They are used in many artificial
intelligent fields such as image processing, speech recog-
nition, pattern recognition, and NLP. However, when large
amounts of training data are available, DL methods out-
perform ML, according to the literature [120].
Many researchers have been using DL methods in the
SA process [121–123]. As a result, many models have been
proposed such as Convolution Neural Networks (CNN),
Recurrent Neural Networks (RNN), Long Short-Term
Memory (LSTM) and Bidirectional LSTM (Bi-LSTM). In
this section, we will present a set of these studies.
SA employing DL methods has been implemented for
different purposes, such as health reviews [124–126],
financial and product reviews [127,128], services moni-
toring system reviews [129] and Movie Reviews
[21,130,131]. They have also been applied to many lan-
guages such as English [132], Chinese [133], Hindi [21],
Spanish [134], Lithuanian [135], Arabic [136,137], Bam-
bara-French [138], Persian [139,140], and Malay [141].
SA of these studies was proposed based on the CNN,
LSTM and Bi-LSTM models. The CNN model outper-
forms the traditional ML methods such as SVM and NB for
Chinese, Spanish, Bambara-French, and Hindi languages
[21,133,134,138]. By contrast, SVM slightly outperforms
the CNN model in Lithuanian, where, the best DL result
was 70.6% using CNN and 72.4% using SVM [135]. A
comparison between three DL models, namely CNN,
Neural Computing and Applications
123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
LSTM, and RCNN, was done in [137] on a corpus of
40,000 Arabic labeled tweets. LSTM obtained the best
outcome with an average accuracy of 81.31%.
Other studies used more than one model to improve their
outcomes. For example, an integrating structure of CNN
and Bi-LSTM model was proposed by [142], a word
embedding model was utilized to transform tweets into
numerical values, then the CNN model received features
embedding as input and produced a lower dimension of
features. Then, the Bi-LSTM model took the input from the
CNN layer and produced a binary classification results. The
effect of the word embedding result on the proposed model
was investigated using Word2Vec and GloVe separately on
retrieved tweets and SST-2 datasets. The ConvBiLSTM
model with Word2Vec had an accuracy of 91.13% on
retrieved tweets dataset according to the findings. In [136],
a DL model for Arabic language SA was created using one
layer of CNN architecture for local feature extraction,
followed by two layers of LSTM to maintain long-term
dependencies. The architecture was supported by FastText
skip-gram word embedding model as the input layer, then
the extract feature maps learned by CNN and LSTM were
passed to a SVM classifier to give the final classification,
the experiments revealed that this model performed
exceptionally well, with an accuracy rate of 90.75%. A
hybrid approach of CNN and LSTM was proposed by
Nezhad and Deihimi in [140], The CNN model was used to
extract features and the LSTM model was used to learn the
long term dependencies. They utilized the word2Vec
model for word representation and applied their approach
on a Persian dataset. The effectiveness of the proposed
model had an accuracy of 85%. A new architecture (Conv-
LSTM-Conv) was proposed in [131]. The authors imple-
mented it on a movie review dataset, and the words were
represented using a word embedding model. In this
approach, the CNN model was used to create a set of
features, then these features were sent to LSTM for
learning long-term feature dependencies. The results were
then resent to the CNN layer to provide abstract features
before final dense layers. The proposed approach achieved
an accuracy of 89.02%. Yuan Lin et al. [143] developed a
Comparison Enhanced Bi-LSTM with multi-head attention
(CE-B-MHA) model. The model combines the ability of
multi-head attention to obtain global information with the
ability of Bi-LSTM to obtain local sequence information.
The proposed model was implemented on IMDB, Seme-
val2017, and Stanford Sentiment Treebank datasets. The
experimental findings reveal that CE-B-MHA outper-
formed several existing models on IMBD and Seme-
val2017 datasets with accuracies of 89.5% and 80.4%,
respectively. A new approach for aspect-based SA was
proposed in [144], which combined the output of four DL
models, namely CNN, LSTM, Bi-LSTM, and gated
recurrent unit (GRU) using the stacking ensemble approach
as one input to multinomial logistic regression for final
classification. The results of experimental showed an
increase in the accuracy from 5% to 20% in comparison to
basic DL models.
Recent DL researches have focused on the use of
transformers like bidirectional encoder representations
from transformers (BERT) [145] for SA classification,
where the encoding and decoding architecture in trans-
formers assists in language modeling [146]. For example,
Mishev et al. [147] evaluated SA using a variety of finance
datasets labeled by financial experts. The authors used a
variety of methods, starting with lexicon-based approach
and ending with those that use transformers, such as BERT
and the Robustly optimised BERT approach (RoBERTa)
[148], Despite the short datasets, the authors conclude that
NLP transformers outperform alternative techniques. Zeng
et al. [149] developed a Local Context Focus (LCF) model
based on the multi-head self-attention (MHSA) mecha-
nism, which used the Context Features Dynamic Mask
(CDM) and Context Features Dynamic Weighted (CDW)
layers to extract local context features. The study employed
three datasets, namely, SemEval-2014 laptop dataset,
restaurant dataset, and ACL twitter dataset. The GloVe
word embeddings and the BERT-shared layer are both
employed to improve the performance of the LCF model.
In all three datasets, the results show that the LCF-BERT
outperforms the state of the art. Moreover, [150] proposed
a fine-tuned BERT model for predicting customer senti-
ments on customer reviews from Twitter, IMDB, Yelp, and
Amazon datasets. Their results were compared to LSVM,
fastText, BiLSTM, and hybrid fastText-BiLSTM models.
The proposed model outperforms the other models on a
variety of performance parameters.
5.4 Optimization approaches of SA
In this category, we will talk about the optimization
approaches, which use meta-heuristics algorithms either to
reduce the number of features to improve the accuracy, for
example, GA [23,151], hybrid grass bee algorithm [152],
or multi-objective swarm optimization algorithm [153]; or
to obtain the optimal cluster like cuckoo search technique
[83], hybrid spider monkey technique [154], and hybrid
multilingual fuzzy technique [155].
For example, the authors in [151] proposed a feature
reduction technique to select the best set of features using
the GA. This technique reduced the number of features up
to 42% without compromising the accuracy of the results.
The results showed that this technique was superior to both
principal component analysis (PCA) and latent semantic
analysis (LSA) by 15.4% and 40.2%, respectively, in terms
of accuracy. The GA was employed in conjunction with a
Neural Computing and Applications
123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
CNN classifier as a feature optimization strategy in [23].
Their study’s major goal is to improve accuracy. The
results of their tests revealed that this model is promising.
Kumar and Khorwal [156] used a binary representation for
the firefly technique to optimize feature subset selection
using the SVM classifier. For the classification, four sep-
arate datasets were employed in two different languages.
They compared their findings to those obtained by
employing a SVM classifier without feature selection and
those obtained by applying a GA-based feature selection.
The results demonstrated that this method was effective in
optimizing the feature set and increasing the system’s
accuracy by 5% compared with outcomes given without
applying feature selection technique and by roughly 3%
compared with the GA technique. Regarding Arabic SA, a
hybrid model of the SVM classifier with Information Gain
(IG) filter feature reduction technique and the Improved
whale optimization algorithm (IWOA) was presented in
[157]. In the first step, IG feature selection was utilized as a
pre-feature reduction strategy to reduce the scope of the
search space by rating all features according to IG and
deleting irrelevant features. In the second step, an IWOA
algorithm was utilized to find the best feature combination
from the data collected and chosen in the first phase to
increase SA classification performance. On four Arabic SA
datasets, six well-known optimization techniques and two
DL algorithms are compared to the suggested algorithm.
The suggested technique exceeds all previous algorithms in
terms of accuracy by finding the best solutions while
minimizing the number of selected features.
A hybrid model for SA on domain-independent datasets
using a hybrid CNN with improved SVM was developed in
[153], by replacing the soft-max function as the CNN
output layer with an SVM classifier. The hybrid model
obtained the benefits of classification efficiency with the
SVM classifier to enhance the CNN’s classification capa-
bility. SVM training was further aided by the optimal
feature set obtained from the multi-swarm particle swarm
optimization (MSPSO)-based feature selection
approach.Furthermore, in [152], a novel hybrid grass bee
optimization (HGBEE) method was used to suggest an
optimal feature selection for multimodal SA in social
media movies. A diverse collection of features was
extracted, and then the final optimal features chosen by
using the HGBEE method, yielding a feature set with an
optimal value for greater precision and reduced computing
time. Madani et al. [155] presented a hybrid technique for
classifying tweets by integrating fuzzy logic system with a
lexicon-based approach. The authors first used the Senti-
WordNet dictionary to determine the crisp value of posi-
tivity and negativity for each tweet. Then a fuzzy logic
system that comprised three steps, namely fuzzification,
rule inference, and defuzzification, was implemented to
obtain the final crisp value and determine the final polarity
for each tweet.
Optimization methods were employed to find the best
cluster, like in [83], where the authors proposed a novel
meta-heuristic method called CSK, which is based on
K-means and cuckoo search algorithm. The suggested
method’s efficacy was evaluated in comparison with the
PSO algorithm, differential evolution (DE), cuckoo search
(CS), improved cuckoo search (ICS), Gauss-based cuckoo
search (GCS), and two n-gram techniques on four twitter
datasets. The suggested approach outperformed the other
alternatives, according to experimental findings and sta-
tistical analysis. A hybrid spider-monkey optimization with
k-means clustering technique (SMOK) was developed in
[154] to extract sentiments from datasets by determining
the ideal cluster heading. The proposed method was com-
pared to four algorithms: SMO, PSO, DE, and GA. The
experiment results on the proposed technique revealed that
the created method outperformed traditional methods in
terms of average accuracy and average computing time.
This strategy, however, proved unsuitable for sarcastic and
conflicting tweets.
The majority of these studies have used different opti-
mization techniques to improve the SA classification per-
formance. A list of these studies is shown in Table 5.
5.5 Hybrid SA approaches
The hybrid approach aims to merge two or more approa-
ches to improve the performance of SA. In recent years,
many hybrid architectures have been created, some of them
mixed lexicon/rule-based method with ML models, as in
[161–165]. For example, Elshakankery et al. in [161]
proposed a hybrid incremental learning approach for Ara-
bic SA (HILATSA). The authors employed two ML clas-
sifiers: SVM and LR, as well as RNN DL classifier for the
classification task. They also built four lexicons: Arabic
words lexicon, emoticon lexicon, idioms lexicon, and
special intensified words lexicon. Six datasets were used to
test the suggested approach. The first five were used to
create lexicons for training, testing, and verifying the
classifier model, while the sixth was used to evaluate and
the proposed approach. The proposed approach accuracy
was 73.67% for three class classification and 83.73% for
two class classification. Ersahin et al. [162] used a com-
bination of Turkish sentiment dictionary and ML algo-
rithms to develop a Turkish SA algorithm on three datasets.
STN was enlarged with ASDICT on the lexicon-based side,
and a lexicon score was produced based on the polarity of
the words in eSTN, which was then fed as one of the
features to ML classifiers. The authors utilized three clas-
sifiers in ML: SVM, NB, and J48. The proposed approach
enhanced accuracy by 7% on average, where the NB
Neural Computing and Applications
123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
outperformed SVM and J48 in the movie and Twitter
datasets, whereas the SVM excelled in a hotel dataset
according to the testing data. The authors in [165] proposed
a hybrid approach of lexicon-based and an ensemble of
stacked ML algorithms. The Sentiment Score, which is
determined from a dictionary-based classifier, was included
to the feature set as an extra feature. The stacked ensemble
classifier was built utilizing three ML algorithms, namely,
SVM, KNN, and C5.0, as well as two meta learners,
namely, RF and GLM. The result shows that the RF meta
classifier perform better than does the GLM meta classifier
by achieving the highest accuracy of 90.66% for five fold
cross validation and 91.25% accuracy for ten fold cross
validation. Moreover, the proposed approach outperformed
the SVM, NB, DT, RF, and ME ML classifiers. The authors
in [164] suggested a hybrid approach to SA using ML and
lexicons based on Twitter data of natural disasters. TF-IDF
and K-means were chosen for sentiment classification, and
in a pipeline of Doc2Vec and K-means, the latent dirich-
let allocation (LDA) topic modeling technique was
employed to collect topics. The data were collected from
243,746 tweets concerning natural disasters in India. Sen-
timent classification was then performed using similarity
and polarity indices, as well as subject identification among
Twitter discussions.
Although other studies merged lexicon/rule-based
method with DL models [22,166–168], Dashtipour et al. in
[166] proposed a hybrid framework that combined depen-
dency-based rules with DL classifiers including CNN and
LSTM for polarity detection in the Persian language. Their
approach outperformed ML methods (SVM, LR) with a
margin of 10–15% and DL methods (LSTM, CNN) with
margin of 3–4%. In [167], the authors developed an
unsupervised technique for opinion target extraction and
aspect term extraction tasks that integrated chunk-level
linguistic rules with DL approach to make logical predic-
tions and attain higher-level aspect representations through
a deep GRU. Ray et al. [168] offered a hybrid approach
combining DL and a rule-based for the aspect SA. The data
were classified using three approaches: CNN, rule-based
approach, and CNN with rule-based approach. The hybrid
approach outperformed the others, with an accuracy of
87%. The authors employed POS as feature extraction
technique, and word embedding as a vectorization tech-
nique. Another SA model, SLCABG, was proposed in [22],
which is based on the sentiment lexicon and includes CNN
and Bidirectional GRU (BiGRU). To improve the senti-
ment features, the sentiment lexicon was employed. The
CNN and BiGRU networks were then utilized to extract the
major sentiment features and context features, after which
the weighted sentiment features were classified.
Some studies built up ML models with DL models like
[169–171]. For example, Srinidhi et al. [169] suggested a
hybrid model for sentiment categorization, which incor-
porated Manhattan LSTM (MaLSTM) and SVM. The
hidden representation was learned using LSTM, while the
sentiments were determined using SVM. The classification
was done on the IMDB dataset. In [170], a hybrid model
was proposed that combined the DL (CNN) with ML
(SVM) to deal with multimodal data (textual and visual)
using decision multimodal fusion for fine-grained predic-
tion of sentiments. The suggested model contains four
phases: discretization to separate image and text, text
analytics using CNN, image analytics using SVM, and
boolean decision module with OR operation to determine
the final classification.
Based on these studies, hybrid approaches outperform
state-of-the-art approaches(ML, DL, and lexicon\rule-
based approaches).
Table 5 Optimization methods
of SA Techniques Objectives References Year
Binary rat swarm optimization algorithm Feature reduction [158] 2022
Random evolutionary whale optimization algorithm Feature reduction [159] 2022
Improved african vulture optimization algorithm Feature reduction [160] 2022
Hybrid grass bee optimization Feature reduction [152] 2021
Multi-objective swarm optimization Feature reduction [153] 2020
Genetic algorithm Feature reduction [23] 2020
Hybrid spider monkey optimization method Optimal cluster [154] 2020
Fuzzy-based approach Optimal cluster [155] 2020
Whale optimization algorithm Feature Reduction [157] 2019
Genetic algorithm Feature reduction [151] 2019
K-MEAN& Cuckoo search Optimal cluster [83] 2017
Firefly algorithm Feature reduction [156] 2017
Neural Computing and Applications
123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
6 SA applications
Several online sources, such as Twitter, Facebook, and
YouTube, have been used for SA to examine the quality of
a new technique used to improve the SA result like [172],
or to obtain a general overview of a trending issue, such as
‘‘knowing the opinion on Egypt’s New Real Estate
Registration Law’’ [173].
In this section, we will present the latest studies that
dealt with these sources.
6.1 SA in Twitter
Twitter has gained the most attention from researchers in
checking the success of their suggested SA models com-
pared Facebook or YouTube. In the following section, we
will mention some of these studies.
By using data pre-processing strategies and a hybrid
approach that combines ML, statistical learning, and lexi-
con-based approaches, the authors in [164] devised a
framework to assess people’s sentiments on twitter about
natural catastrophes. Tweets were represented using the
TF-IDF technique. The lexicon-based approach from the
Orange Data Mining toolkit was used in [174]todoSAon
a corpus of tweets to forecast the favorable, negative, and
neutral feelings of the tweets for the syrian chemical attack.
Moreover, because the study only looked at tweets written
in English, it could not capture the distinctions and feelings
of non-English tweets. The lexicon-based approach was
used in [12] to analyze 109,990 tweets using the NRC
lexicon.
Villavicencio et al. [175] utilized the RapidMiner data
science software to apply NB model to categorize English
and Filipino tweets related to COVID-19 vaccination.
Given that the models in RapidMiner do not include the
Filipino language, the authors manually annotated the
collected tweets. Through classifying tweets into their
appropriate sentiment and emotion classes, Sailunaz and
Alhajj [176] employed a ML technique, specifically by the
NB algorithm, to generate user effect scores based on user-
based and Twitter-based factors. Tweets and their replies
concerning a specific topic were collected and turned into a
database. Only simple texts were utilized in their studies,
and they retrieved proper terminology from it. Social net-
work content includes many abbreviations. The CNN-
based technique was introduced in [9] for SA, where the
features were collected from user behavior data. The
authors revealed how they utilized the CNN DL model to
improve the accuracy of their respective SA. Wang et al.
[177] found that integrating diffusion patterns with Twitter
texts enhanced Twitter SA by applying the SentiDiff
algorithm, but they did not take into consideration the
subject specifics of these tweets or how sentiment diffusion
patterns change between subjects when combining textual
and sentiment diffusion data.
In [178], the SentiStrength App, which is a lexical
approach, was utilized for SA of tweets about diabetes.
This technology, however, has yet to be verified for use in
the medical SA field, also due to the restrictions of the
standard API they used, which prevented them from ana-
lyzing additional information on likes or replies. Ozturk
and Ayvaz [179] looked over Turkish and English tweets to
see how people felt about the Syrian refugee situation. For
Turkish tweets, they created a lexicon, while for English
tweets, they utilized the RSentiment dictionary.
Another study introduced an approach to process multi-
sentiment classification problem was in [180], where the
authors expanded the idea of binary and ternary classifi-
cation and created the SENTA tool, which allows users to
choose the features that are most suited for the application
used to classify tweets into seven sentiment classes from a
large list of features. This method was also demonstrated to
be adequately accurate in both binary and ternary classifi-
cation. The classification was carried out using the RF ML
method, and if the user does not want their features to be
exported, they may utilize the Weka tool to do so.
By utilizing financial community tweets on Twitter,
Daniel et al. [181] employed SA to identify event popu-
larity for a financial market. TextBlob, Sentistrength, Affin,
and MySentimentApi were used to classify the tweets, the
first three produced by separate authors and the fourth
developed by [181] using dictionary terms that include the
polarity of words.
Negation processing in SA received attention in [182].
Given that we know how the negation operates to modify
the polarity of sentiment terms that are within its scope,
[182] dealt with this issue by using an unsupervised clas-
sification system based on a set of rules. To assess the
impact of negation, a corpus of tweets written in Spanish
was used to test the system both with and without negation.
The results show that treating negation improves the
accuracy of the final system significantly.
Table 6presents some of these studies and the objectives
of each one.
6.2 SA in Facebook
Facebook is the world’s most active social media. How-
ever, it is not often used for SA because the data are
chaotic, poorly organized, and individuals frequently uti-
lize short forms with numerous spelling errors. This makes
examining the data more difficult [188]. Many researcher
have used Facebook posts to address a variety of problems.
The use of this platform posts was motivated by the fact
Neural Computing and Applications
123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
that it is the most popular social media site for the topics of
these studies [189].
Different approaches have been used to deal with
Facebook data. For example, lexical-based approach was
implemented in [173,190]. The authors in [173] used
Facebook posts to obtain a general view about the street
opinions of Egypt’s New Real Estate Registration Law.
Notwithstanding the conclusion obtained in this study is
that this approach has not been able to handle all of the data
structures collected for the study, because the lexical-based
approach did not take into account the context. Moreover,
it depends on detecting the polarity of words separately
from one another, which is inappropriate for the Arabic
language, where the use of words and expressions is highly
subjective. As a result, more advanced language techniques
are required to deal with these limitations. In [190] the
authors employed a lexicon-based approach to analyze
Arabic text sentiment. Three types of language were con-
sidered: Modern Standard Arabic, Mixed Arabic, and
informal Arabic texts. The study used the pre-processing
on the text, paying special attention to negation words,
emotions, and intensifier words that have a substantial
effect on the polarity of the sentence.
Whereas ML and DL approaches was used in
[191–195]. SA based on the BiLSTM DL approach was
used in [191] to identify and assess vaccine-deniers’
arguments against children vaccination. The acquired data
was manually categorized, which is a time-consuming
operation. The study also had a set of limitations, including
omitting pictures, which relates to a significant portion of
the emotional impact of personal opinions. The collected
data were also limited to a single Facebook page. Alvarez
et al. [192] investigated data from Russian IRA Facebook
ads to analyze emotions in the ad text. They applied sen-
tence-level SA, and justified that because the typical
Facebook ad text presented one to three sentences. A linear
regression method was used to train the classifier using
data from the text that represent the Facebook ad, not the
text that appeared in the ad picture. Even though we would
expect the emotions to be consistent between the ad text
and the text in the picture, the text in the advertisement
picture could have the opposite interpretation, affecting the
sentiment accuracy results. In [193], four ML models were
trained on a corpus of Iraqi Arabic dialects to identify
sentiments. A word embedding approach was used to
represent the input Features. LR and SVM performed better
than DT and NB classifiers, according to the findings.
Considering the nature of the gaming community, which
uses many slang phrases that are not categorized using
dictionaries, the authors in [194] used the KNN learning
approach for the sentiment classification of the data set,
which was labeled manually. The accuracy of the model
was 82.3%. However, retraining the data-set using other
models (e.g., NB, SVM, or dictionaries) will give deeper
insight to the current findings.
A hybrid technique that combines lexicon-based and
ML approaches was presented in [196] to predict cos-
tumers’ satisfaction with Jordanian telecommunication
providers’ services. The data set was gathered from Face-
book comments and labeled using a user-defined lexicon,
before being used to construct a ML model as well as
Arabic language grammar standards. Moreover, sentence
structure was applied on it, in which three ML models were
used (SVM, kNN, and NB) for the classification. Notably,
Table 6 Sentiment analysis in
twitter Objectives References Year
Examining corona vaccination tweets from India’s Twitter social media [183] 2022
Developing a model to detect depression for Arabic tweets [184] 2022
Analyze public concerns regarding COVID-19 vaccines in Iran [185] 2022
Exploring netizen’s opinions on cryptocurrency [186] 2022
Developing a framework to analyze users’ sentiments on natural disasters [164] 2021
Sentiment analysis of tweets related to Syria Chemical Attack [174] 2021
Twitter SA towards COVID-19 Vaccines in the Philippines [175] 2021
Discover ADRs from digital social media [187] 2021
Increase understanding of public awareness of COVID-19 pandemic [12] 2020
Proposed a CNN model that incorporates users’ behavioral information [9] 2019
Categorizing tweets into their appropriate sentiment and emotion category [176] 2019
Utilize sentiment diffusion patterns to improve Twitter sentiment analysis [177] 2019
Analyze the sentiments expressed in messages about diabetes on Twitter [178] 2019
Investigating the public opinions towards the Syrian refugee crisis [179] 2018
Propose an approach that classifies Twitter texts into seven different classes [180] 2017
Detecting events popularity through financial market tweets [181] 2017
Studying the scope of negation for the Spanish sentiment analysis [182] 2017
Neural Computing and Applications
123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
due to issues with the Arabic language, the user-defined
lexicon would fail to meet its goals if it did not support
Arabic language grammatical rules and sentence structure.
The studies in [189,197,198] implemented SA using a
set of available tools. The authors in [189] pointed out that
there are a number of accessible resources for applying SA,
including the University of Pittsburgh’s OpinionFinder,
Apache OpenNLP, and StanfordNLP. In their study,
OpinionFinder was used for SA to assess the patients’
opinions on a particular medical prescription. Google
Cloud’s Natural Language API was also used to analyze
the sentiment of police agency Facebook pages in [197],
They justified this use by pointing to their programming
interface’s simplicity and their industry-leading position in
the disciplines of search and language processing. Simi-
larly, in [191,192] posts and comments that did not contain
any text, such as images or videos, were not considered.
The study in [198] investigated the effect of a health-re-
lated issue on consumer responses on social media for a
well-known and trusted business. The Lexalytics Intelli-
gence Platform was used to analyze the sentiment of cus-
tomer comments and postings.
In Table 7, we present these studies and the objectives of
each one.
6.3 SA for video
SA was also applied on videos’ comments to obtain an
overview of what viewers think.
The authors in [202] combined a dictionary and NB
algorithm to create a hybrid SA approach. Researchers
built a sentiment dictionary to extract features from Dan-
maku text, which is a form of live network video com-
mentary that focuses on video content. The essential task of
a sentiment dictionary in a Danmaku reviews is to extract
feature terms and convert them to word vectors. Following
feature extraction, a NB model was used to classify the
Sentiment of the Danmaku (SD) comments. However, it is
inappropriate for Danmuku data. The results revealed that
the suggested SD-NB outperforms N-gram-NB and
N-gram-SVM in in terms of effectiveness.
Another study merged particle swarm optimization
strategy with SVM to estimate SA outputs over upwards of
1000 text comments gathered from YouTube medical
videos in the [203] search. Furthermore, the study used a
comparison of NB, KNN, DT, SVM, and SVM with the
optimisation approach (SVM-PSO). The strategy that
combined optimization techniques with ML classifiers
outperformed the others.
With minimal work and no extra training data, the
authors in [204] updated and enlarged a predetermined
lexicon-based SA approach that was initially created for
movie reviews to detect toxicity and hostility between
players in video games. They justified not using a super-
vised ML approach because the data set is substantially
smaller than what would be required. Moreover, set of
alterations were made to the lexicon because of the unique
characteristics of game chat, such as brief, truncated, and
offensive messages containing emoticons and game-
specific language, which are considered main challenges to
traditional lexicon-based approaches or ML models
designed or trained on conventional data-sets such as
movie reviews or customer feedback. Additionally, the
goal of [205] was to analyze the sentiment of comments of
YouTube videos about cleft lip and palate. SentiStrength, a
free SA tool that measures the strength of sentiments in
short texts and can deal with slang, was used to classify the
comments on the videos. Another study [206] in the
medical field used the titles of videos published on You-
Tube. Its authors tried to determine if and how the public’s
Table 7 Sentiment analysis in
facebook Objectives References Year
Detecting and analyzing unstructured data of Hate Speech [199] 2022
Sentiment analysis regarding agritech startups topics in the Thai language [200] 2022
Detecting and identifying symptoms of influenza-like illness [201] 2022
knowing the opinions on Egypt’s New Real Estate Registration Law [173] 2021
Identify and analyze vaccine-deniers arguments against child vaccinations [191] 2021
Sentiment analysis use for predicting mobile applications breakout [195] 2021
A Sentiment Analysis of the 2016 Elections Russian Facebook Ads [192] 2020
Sentiments analysis of comments for jordanian telecom companies [196] 2020
sentiment analysis of police agency media pages [197] 2020
Sentiments analysis of Iraqi Arabic dialect [193] 2019
Sentiment analysis to measure the success of three YouTube gamers [194] 2019
Brand crisis-sentiment analysis [198] 2019
An Adopted Sentiment Analysis Model for Arabic Comments [190] 2018
Investigate the patients perspectives on a given medical prescription [189] 2017
Neural Computing and Applications
123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
perception has changed before and after a vaccine program.
They used the co-occurrence network (CON) to analyze the
data, and the Jaccard similarity coefficient was used to
determine the degree of co-occurrence. They employed
lexical coding criteria to rule out logic operators, demon-
strative adjectives, personalities information, and nonsen-
sical words when conducting the CON. To determine
sentiment, the authors used the NRC sentiment dictionary,
which is a compilation of English terms and their corre-
lations with eight basic emotions (anger, fear, anticipation,
trust, surprise, sadness, joy, and disgust) and two moods
(negative and positive).
We summarized some of these studies and the objectives
of each one in Table 8.
6.4 Multimodal SA
Some recent studies focus on SA from multimodal signals
which include visual, audio and textual data [209,210],
video blogs or spoken reviews that are posted on YouTube,
Instagram, or any other social media platforms contain an
expression of sentiment, e.g., a blogger talking about a new
product.
A new quantum-theoretic multimodal fusion approach
was presented in [211]. The approach was designed to
detect the sentiment of a multimodal sentence composed of
word-aligned visual, auditory, and textual formulations.
The visual and auditory characteristics retrieved in this
study were of poor quality. Therefore, improving tech-
niques to extract sentiment-sensitive features from non-
textual formats might be a promising avenue for improving
sentiment accuracy in multimodal SA. The hybrid
approach proposed in [152] sought to find correlations
between text, audio, and video, followed by multimodal
SA. To achieve the best feature set, a novel HGBEE
optimization method was used, followed by multimodular
fusion of text, acoustic, and video features. Eventually,
they classified sentiment using their proposed multikernal
extreme learning classifier.
Tzirakis et al. [212] proposed an emotion recognition
system that utilizes text, audio and visual modality in an
end-to-end manner. They used CNNs for this system and
presented a new transformer-based text modality architec-
ture that can reliably capture the sentence meaning. They
created an audio model for the audio channel and a high-
resolution network for the visual model, and they suggested
new attention-based approaches to combine the modalities.
They employed LSTM networks to capture the temporal
dynamics of the signal.
The authors in [213] presented a multimodal approach to
Arabic SA, in which a new dataset was prepared for
analysis. SA was performed using several fusion tech-
niques and various combinations of data modalities, with a
focus on the speaker’s face, audio, and spoken text. Using a
hybrid approach of CNN and fuzzy logic, [214] introduced
a convolutional fuzzy sentiment classifier. Video, ECG,
and text datasets were used to assess the method’s per-
formance because the video dataset’s input is a series of
images, and the label is the polarity of facial expression.In
terms of classification accuracy, this technique outper-
formed the baseline strategies. By contrast, it dealt with
each modality separately, with no fusion procedures used.
The authors in [215] reported an increased accuracy for
both multimodal and unimodal on the CMU-MOSI dataset
through proposed a model to learn the intra-modality and
inter-modality dynamics of the three data types (text,
visual, and audio) using a tensor fusion network.
In Table 9, we present some recent studies that are
related to multimodal SA.
7 Sentiment analysis datasets
SA datasets serve as the foundation for developing and
evaluating various SA approaches. Several publicly avail-
able and regularly used SA datasets in several languages
have proven to be valuable tools, even though some SA
studies employed datasets acquired by the researchers
themselves. In this section, we will offer an overview of
publicly available datasets in English and Arabic
languages.
Table 8 Sentiment analysis in
video Objectives References Year
Predicting web series video views [207] 2022
Sentiment analysis of YouTube videos about cleft lip and palate (CLP) [205] 2021
Sentiment analysis to analyse YouTube Italian videos concerning vaccination [206] 2020
present a sentiment analysis model for the recommendation of video clips [208] 2020
Sentiment analysis of live comments related to the video being screened [202] 2020
Sentiment analysis for medical experts’ comments on telemedicine videos [203] 2019
Sentiment analysis of player chat messaging in the video game StarCraft 2 [204] 2017
Neural Computing and Applications
123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
7.1 English datasets
7.1.1 IMDB movie reviews dataset
This dataset includes movie evaluations gathered from the
well-known website for movie reviews known as Internet
Movie Database (IMDB). This Dataset has approximately
50K movie reviews for binary SA, half of the dataset was
classed as positive, and the other half as negative, in an
equal-sized split. The dataset is available on the Kaggle
platform website as a CSV file containing two columns:
‘‘review’’ and ‘‘sentiment’’ [223].
7.1.2 SemEval-2017 task 4 dataset
The SemEval-2017 Task 4 dataset stands out from previous
SemEval-2016 Task 4 datasets with new updates including
adding Arabic as a new language for all subtasks and
making data from Twitter users’ profiles available [224].
This dataset serves as a comprehensive benchmark for
evaluating SA approaches, encompassing five subtasks for
both Arabic and English languages that address different
aspects of SA. It has been divided into three classification
subtasks, namely subtasks A, B, and C. Additionally, there
are two quantification subtasks, namely subtasks D and E.
The SA scale employed ranges from a two-point to a five-
point scale.
7.1.3 YELP dataset
The Yelp open dataset includes reviews written by users
about different businesses and services available on the
Yelp platform for a range of companies such as restaurants,
hotels, shopping centers, fitness centers salons, and more.
The dataset is available for use for personal, educational,
and academic purposes where Only the reviews that Yelp
recommended at the time of data collection are included, it
is accessible for free, but first, the user must register and
accept the conditions of usage. The latest version of the
dataset in 2023 contains approximately 6.99 million
reviews of 1.98 million users according to the Yelp official
website [225]. However, where the online text dataset is
offered in JSON (JavaScript Object Notation) format, a
straightforward Python script may be used to quickly
convert it to a CSV format.
7.1.4 Amazon review dataset
The Amazon review dataset is a collection of customer
reviews and product ratings for items sold on the Amazon
online store [226]. This dataset contains around 82 million
reviews spanning numerous product categories, including
electronics, books, movies, and more, collected from
approximately 20 million people between 1996 and 2014.
The Amazon review dataset is a large collection of reviews
that include important information such as the reviewer’s
ID, the product ID, the review text, the reviewer’s rating on
a scale of 1 to 5, and extra metadata such as the product
image and price. This dataset is beneficial for assessing
customer opinions, analyzing sentiment, and constructing
prediction models based on user comments.
7.1.5 UCI reviews dataset
The reviews dataset from the UCI ML repository [227]
contains 3000 reviews, each labeled with 1 or 0 to indicate
whether the review is positive or negative. These reviews
were gathered randomly from three major review sites:
IMDB, Yelp, and Amazon.
Table 9 Multi-modal sentiment
analysis Objectives References Year
Develop a real-time Sentiment Analysis model of COVID-19 in Indonesia [216] 2022
Presented a novel image-text interaction network (ITIN) [217] 2022
Hybrid contrastive learning of tri-modal representations [218] 2022
Proposed a hybrid MSA model based on weighted CNN [219] 2022
Topic Recognition in Video transcriptions [220] 2021
Sentiment analysis of YouTube videos on Mohs surgery: Quality of content [221] 2021
Analysis of YouTube video contents and comments on Koha and DSpace [222] 2021
Proposed a novel text-audio-video multi-modal sentiment analysis method [152] 2021
Fusing different modalities of features for multimodal sentiment analysis [211] 2021
Proposed a new approach for multimodal sentiment analysis [212] 2021
Present an approach to predict the speaker’s sentiment of dialect Arabic [213] 2020
Presented an emotion-based sentiment classifier for video and text sequences [214] 2019
Proposed a model deals with text,visual, audio using a tensor fusion network [215] 2017
Neural Computing and Applications
123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
7.2 Arabic datasets
7.2.1 Large-scale arabic book review dataset (LABR)
This dataset was created by [228]. There are more than
63K book reviews in this dataset, each rated from 1 to 5
stars in one of three categories: positive, negative, or
neutral. They were considered ratings of 4 or 5 for positive
reviews which are 42,689 reviews, and ratings of 1 or 2 for
negative reviews which are 8012 reviews. Reviews that
receive a 3 are regarded as neutral reviews.
7.2.2 Bahraini dialects dataset
Omran et al. [229] presented a dataset that encompasses
Amazon product reviews written in Bahraini dialects.
known as ‘‘E-MSA-BDs-PR-SA’’ (English Modern Stan-
dard Arabic Bahraini Dialects product reviews for SA),
which is available at [230]. This dataset was generated
utilizing two translation cascade phases, one machine and
one manual. The English Amazon product reviews were
translated into Standard Arabic using Google Translate.
The generated Arabic reviews were then manually con-
verted to Bahraini reviews by certified native speakers
using specific forms established by them. The dataset is
balanced, with 2500 positive and 2500 negative reviews.
7.2.3 Arabic sentiment tweets dataset (ASTD)
The authors of [231] provided a SA dataset from Twitter.
The dataset consists of 2479 tweets, including 1684 tweets
in the negative category and 795 tweets in the positive
category.
7.2.4 Arabic health services dataset (Main-AHS and Sub-
AHS)
Two classes make up this dataset (positive and negative),
which was first given in [232]. It is an unbalanced dataset
with 2026 tweets, 1398 of which are negative and 628 of
which are positive. This dataset is called Main-AHS, and
another subset of this dataset, called Sub-AHS [233] con-
tains 1732 tweets, with 502 positive tweets and 1230
negative tweets.
8 Sentiment analysis versus emotion
recognition: a comparative perspective
In this section, we will give an overview of the relation,
similarities, and major differences between SA and emo-
tion recognition. Three types of information can be found
in human communication: linguistic, paralinguistic, and
nonlinguistic. Linguistic information consists of rules and
symbols used in communication. Paralinguistic informa-
tion encompasses additional cues like tone, speech manner,
and rhythm provided by the speaker. Nonlinguistic infor-
mation includes factors such as the age and gender of the
speaker, which are unrelated to the core content of speech
[234,235]. The process of recognizing, analyzing, under-
standing, and reacting to people’s emotional states and
emotions from happiness to fear is known as emotion
recognition [236,237]. This process used technical abilities
like facial recognition, speech recognition, voice recogni-
tion, biosensors, deep learning, and pattern identification to
identify and recognize human emotions [238,239].
Emotion recognition systems can be categorized as
either unimodal or multimodal. Unimodal systems use a
single modality, such as a face, signals, or text, while
multimodal systems combine multiple sources of data to
create a global view and make the recognition more
accurate [240,241]. Although SA and emotion recognition
are closely related areas, the prevailing perspective among
researchers views sentiment as part of emotion. emotion
recognition includes a wider and more in-depth under-
standing of emotional states and sensitivities whereas SA
concentrates on a more basic categorization of sentiment
polarity [242,243], for instance, when analyzing customer
feedback, one might encounter expressions like ‘‘I love
your product’’ (emotion) or ‘‘Your customer service is
good’’ (sentiment).
Several approaches have been investigated in emotion
recognition field which can be categorized as traditional
ML approaches like SVM [244–247], KNN
[244,245,248], extreme learning machine (ELM)
[249,250], and DL approaches like RNN-LSTM
[251–253], CNN [254,255], Moreover, some researchers
used hybrid techniques that integrated ML and DL
[256,257].
When comparing SA and speech emotion recognition,
several similarities, as well as differences emerge. Speech
emotion recognition methods rely on audio data such as
conversations, recorded speeches, or live speech, whereas
SA depends mostly on textual data gathered from posts,
comments, or reviews. Both SA and speech emotion
recognition utilize a total of three approaches: lexicon/rule-
based, machine learning (ML), and deep learning (DL).
However, in SA, the lexicon/rule-based technique employs
a polarity dictionary, whereas lexicon/rule-based models in
speech emotion recognition employ an emotions dic-
tionary. Furthermore, features in SA models can be
described as either lexical or linguistic, whilst features in
speech emotion recognition include both linguistic and
acoustic parts. These models employ distinct techniques for
feature extraction due to the different types of data they
Neural Computing and Applications
123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
handle. SA encounters challenges related to contextual
understanding, sarcasm detection, and negation. On the
other hand, speech emotion recognition faces challenges
such as speaker dependence and varying noise levels. A
comparison between SA and speech emotion recognition is
presented in Table 10.
9 Open issues
Despite the advancement and development of many SA
approaches, there are fundamental obstacles that still need
to be addressed. Future SA work will require a new per-
spective to address major issues in the field, where some of
which are summarized here.
9.1 Spam detection and fake reviews
Reading people’s reviews about various products and ser-
vices has become a common practice among customers in
the shopping process. Some fake review manipulators see
this as a chance to write spam reviews to improve or harm
a brand. Distinguishing between spam and true reviews is a
highly challenging process because it necessitates linguis-
tic and grammatical knowledge. Moreover, the major
limitation in the classification approach to distinguish spam
reviews is the unavailability of the labeled dataset [258].
One study [259] focused on building a set of features to
employ with different classifiers as trustworthy input for
the sake of opinion spam detection. The results showed that
the best classifiers for detecting Persian spam opinions are
DT and AdaBoost classifiers. Another study [258] focused
on utilizing three features: review sentiment and com-
ments, content-based factor, and rating deviation. This
approach used comment data to determine whether the
review is spam. The authors used those labeled data with
the ML model to categorize the remaining unlabeled data.
9.2 Sarcastic reviews
Sarcasm is challenging to SA. Sarcasm is a way of
expressing feelings in which people write or say something
that is not exactly what they mean [260]. People who write
sarcastic text use positive words to express their negative
feelings or vice versa. Therefore, sarcasm can easily fool
SA models unless they are built to deal with it. Sarcasm
can be hard to understand not only for a machine but also
for a human. Many approaches have been proposed for
sarcasm detection [261–263]. One study [132] attempted to
identify sarcastic sentences used in English conversation.
The results indicated that the interjection and unigram
features had a positive increase in organizing the classifi-
cation of the tweeting tone with sarcasm.
9.3 Negations
Given that negations affect the polarity of words, they are
highly crucial in determining the correct tone of sentiment.
Table 10 Comparison between sentiment analysis and speech emotion recognition
Sentiment analysis Speech emotion recognition
Definition Analysis of sentiment expressed in text Recognition and classification of speech-based emotions
Data Textual data (reviews, posts, customer
feedback)
Audio data (conversations, speeches-recorded or live
speech)
Approaches Polarity lexical-based approach, ML approach,
DL approach, and hybrid approaches
Emotion lexicons-based approach, ML approach,
DL approach, and Hybrid approaches
Key Features Lexical features, Linguistic features Linguistic features, acoustic features
(prosody, voice quality, tone)
Feature Extraction
Techniques
Bag-of-words, TF-IDF, word embeddings Linear Predictive Coding, Mel-Frequency Cepstrum,
Relative spectral, principal component analysis,
Linear discriminant analysis
Objective Determining sentiment polarity
(positive, negative, neutral)
Identifying emotional states (happiness, anger, sadness, etc.)
Common Applications Social media sentiment analysis,
brand reputation management
Human-computer interaction, Affective computing,
Voice-driven systems, Mental health analysis
Challenges and
Considerations
Contextual understanding, Sarcasm detection,
Data noise, Spam detection and fake reviews,
Negations
Speech pattern variety, speaker dependence, and noise levels
Neural Computing and Applications
123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Words like ‘‘no’’, ‘‘not’’, and ‘‘can’t’’ are examples of
negations. When a negation appears in a sentence, identify
which words are affected by this negation (i.e., the scope of
the negation). For instance, in the sentence ‘‘The camera is
not perfect but the phone is nice’’, the scope of negation is
restricted to the next word after the negation because the
phrase ‘‘but’’ divides the sentence into two sentences. Also,
in the sentence ‘‘The phone has not worked properly for a
long time’’, the scope of negation reaches the end of the
sentence. Authors in [264] have given three categories of
negation: the linguistic, structural and valance shifters. In
[265] the authors proposed using a tailored reinforcement
learning method, then conducted a thorough comparison
using a two-pronged evaluation. They first used a manually
labeled dataset to compare predictive performance. In this
case, reinforcement learning outperforms other approaches
in the literature, resulting in a balanced classification
accuracy of up to 70.17%. Second, they examined how
detecting negation scopes can improve SA for financial
news, resulting in a 10.63% increase in the correlation
between news sentiment and stock market returns. This
highlights the importance of negation scope detection in
sentiment decision support. In [266], a novel method for
determining the scope of negation in a sentence was pre-
sented. The authors proposed a hybrid architecture with the
ability to capture salient information in order to determine
whether a token is in scope. They used a Bi-LSTM network
and a CNN in this method.
9.4 Independent domains
Sentiment classification is well known as a domain-de-
pendent task [267]. In many cases, a sentiment classifier
that has been trained in one domain may not perform fully
in another because the sentiment expressions used in var-
ious domains are usually different [268]. Furthermore, the
same word can convey different reactions in different
contexts. For example, in the electronics domain, the word
‘‘easy’’ is frequently used in positive reviews, as in ‘‘This
tablet is easy to use.’’ However, in the security domain, it is
frequently used as a negative term. ‘‘This application is
easy to hack.’’ As a result, we will need to create a more
robust sentiment classifier that can work across multiple
domains. This is known as multi-domain or cross-domain
SA [268].
9.5 Chronological order of reviews
Reviews are listed in chronological order. There is a crucial
issue to consider when conducting SA of reviews based on
the time in which they were shared [269]. In other words,
people’s attitudes toward a given issue or product may alter
over time, it is critical to consider the element of changing
sentiments over time. Personal experiences, fresh knowl-
edge, product upgrades, changing cultural trends, and
shifting preferences can all cause people’s ideas, attitudes,
and perceptions to change.
10 Conclusion and future work
In this review paper, SA and its relevant approaches were
presented. Researchers seek to develop SA approaches to
deal with accurate analysis in different vital majors for
different purposes. The goal of SA is to provide a precise
analysis of sentiments. To do so, researchers have
employed a variety of techniques, including lexicon/rules-
based models, ML, DL, and hybrid models, this overview
briefly discussed the benefits and drawbacks of these
strategies. According to the literature review, hybrid tech-
niques outperformed other approaches in terms of accu-
racy. Furthermore, researchers applied optimization
approaches to these models to determine the best clusters
among the unlabeled data or to reduce the number of fea-
tures to reduce the computing time. The literature reviews
additionally showed that multimodal SA had become a
popular search topic that focuses on analyzing data, which
includes visual, audio, and textual content. Furthermore,
the literature reviews showed that the majority of studies
that focused on SA were in English, with some newer
studies focusing on other languages such as Arabic and
Chinese, and some other studies that emphasize addressing
the issues associated with multilingual SA.
In the forthcoming paragraphs, we will delve into the
research questions that were initially raised in this paper
and provide comprehensive and detailed responses to each
of them.
According to Q1 and Q2, the literature review indicates
that there were three basic ways employed to treat the SA
process. The first technique is lexicon-based, which is
classified into two types: dictionary-based and corpus-
based approaches. The second technique is based on ML
approaches, whilst the third approach is based on DL
approaches methods. Furthermore, hybrid models based on
these primary approaches were provided to improve the
analysis outcome. A hybrid strategy that merged a lexicon-
based approach with a DL approach is an example of a
presented methodology. Furthermore, this review discussed
studies that dealt with optimization strategies such as the
use of meta-heuristics algorithms that were combined to
optimize the SA process. Contextualized word embeddings
are a new manner of word embedding technique. This
concept arose alongside huge language models (LMs) such
as GPT and BERT. A large language model’s embedding
space can capture multiple semantic and syntactic relations
between words, leading to significant improvements in SA
Neural Computing and Applications
123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
tasks. For example, researchers can employ pre-trained
LMs to extract features from words that capture semantic
and contextual information; these features can then be used
as inputs for a variety of SA tasks.
SA is becoming more popular in a variety of sectors.
More political, industrial, and other sectors have come to
understand the importance of analyzing public opinion to
gain insights and make educated decisions. A tendency for
more in-depth mental analysis has also been found,
according to studies, where the researchers focused on
developing models that go deeper into SA classification.
This involved classifying them into multiple categories or
levels of polarity, such as positive, negative, neutral,
strongly positive, and strongly negative. It’s critical to
distinguish between the SA and the realm of emotional
recognition. where emotions are classified into several
forms in this domain, including joy, sadness, rage, etc.
Moreover, due to the huge amount of data that is
available, many researchers have employed DL approaches
to deal with this amount of data. Several models, such as
CNN, RNN, LSTM, and Bi-LSTM, have been presented.
These approaches have yielded encouraging results in
terms of capturing complicated linguistic patterns and
enhancing sentiment classification accuracy. In addition,
domain adaptation and transfer learning techniques have
gained popularity as solutions to the problem of scarce
labeled data in specialized areas. These approaches seek to
adapt knowledge gained in one area to another, hence
increasing SA performance. Recent DL research have
focused on the use of transformers like Bidirectional
Encoder Representations from Transformers (BERT) for
SA classification, where the encoding and decoding
architecture in transformers assists in language modeling.
Another trend that evolved during this period was aspect
analysis. There is a rising interest in evaluating feelings
toward specific features or entities within a text by splitting
it down into several aspects. The purpose of aspect-based
SA researches is to find and evaluate sentiment expressions
associated with numerous parts or features of a product,
service, or topic. The result of these researches is com-
monly utilized in a variety of sectors such as consumer
feedback analysis, brand monitoring, market research,
social media analytics, and product creation. Furthermore,
as the popularity of multimedia content has expanded, so
has the study of multimedia SA, which encompasses tex-
tual, audio, and visual data. As an example, consider a
video of a blogger reviewing a new product. This com-
prises assessing sentiments from text, photos, videos, and
audio, resulting in a more thorough understanding of
emotional expression.
According to Q3, many different companies and sectors,
including those in healthcare, news reporting, e-commerce,
public opinion research, presidential election surveys, and
many more, use SA. Moreover, several online resources,
including Twitter, Facebook, and YouTube, have been
utilized to assess the quality of new methods employed to
enhance the SA result or to get a broad overview of a
trending issue. For example, one of the applications of SA
results is understanding customers’ opinions and identify-
ing opportunities for improvement through the analysis of
consumer feedback, which will allow for managing brand
reputation and dealing with unfavorable opinions by being
aware of consumer preferences, market trends, and com-
petitor attitudes to make smarter decisions and have more
accurate insights. Furthermore, the real-time evaluation of
client feedback will enhance customer service and proac-
tive response to their feedback, which has economic and
political dimensions. Another application of SA science
was used in the political sphere, where evaluating voter
attitudes about political candidates, policies, and elections
helps to develop effective campaign plans. SA is also used
in the financial markets for analyzing the general vibes of
financial news to forecast movements in the markets and
guide investment decisions, as well as estimate risk and
fraud detection.
In response to Q4, we listed the significant sources that
have been used in the field of SA in Sect. 6, and we also
presented a set of available datasets used in SA studies in
Sect. 7.
According to Q5, there are a set of research gaps in the
SA domain that still need more in-depth investigation. One
of these gaps is related to current domain adaptation
techniques, which it is usually dependent on supervised
approaches. However, as we all know, the labeled data are
not always available. As a result, future research may
investigate unsupervised and semi-supervised adaptation
approaches that can use unlabeled data to efficiently
improve SA adaptation models. So, finding more effective
ways for transferring SA models from one domain to
another with minimal labeled data is the research gap in
domain adaptation and transfer learning for SA. Another
issue is related to LMs models, where large language
models (LMs) and contextualized word embedding
approaches have made tremendous progress in capturing
contextual information, but context understanding still has
space for improvement. Furthermore, the availability of
high-quality labeled data that captures different scenarios
of language is critical for training reliable SA models.
However, acquiring such statistics might be difficult,
especially for specific circumstances or minority lan-
guages. Also, there is a gap in multimodal SA models,
where it is crucially dependent on integrating data from
several modalities. Therefore, creating efficient fusion
approaches that may properly mix elements from many
modalities while maintaining their complementary char-
acter is still a difficult task. Another point related to
Neural Computing and Applications
123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
multimodal SA is the dynamic analysis of multimodal
streams, like live video content, is becoming increasingly
necessary to track how sentiment and trends change over
time.
Moreover, ambiguity, knowing the cultural context, and
constructing models that can dynamically update their
knowledge of sentiment over time are some of the issues
that researchers continue to encounter. Furthermore, there
has been a growing interest among researchers in utilizing
LMs models for SA tasks. Moreover, recent research has
given substantial attention to the use of large models like
GPT in the field of NLP [46,270–272]. However, despite
its widespread usage in research, there exists a notable re-
search gap in applying these models to SA. The lack of
comprehensive investigations exploring the full potential
of GPT’s capabilities for SA tasks highlights the need for
further exploration and study. Consequently, this will open
up possibilities for further research into the application of
these methods in the area of creating and enhancing SA
models.
This research showed that, despite the significant
advancements in the field of SA, it is evident that future
work in this domain will require a fresh perspective to
effectively address the substantial challenges that persist.
Funding This work is unfunded.
Data availability This study involved the re-analysis of existing data,
which is available at locations cited in the reference section. Fur-
thermore, the data generated and analyzed during the current study
are available in the Scopus database at https://scopus.com.
Declarations
Conflict of interest On behalf of all authors, the corresponding author
states that there is no conflict of interest.
References
1. Pang B, Lee L et al (2008) Opinion mining and sentiment
analysis. Found Trends Inf Retr 2(1–2):1–135
2. Das S, Chen M (2001) Yahoo! for amazon: Extracting market
sentiment from stock message boards. In: proceedings of the
Asia Pacific finance association annual conference (APFA), vol.
35, p 43. Bangkok, Thailand
3. Tong RM (2001) An operational system for detecting and
tracking opinions in on-line discussion. In: working notes of the
ACM SIGIR 2001 workshop on operational text classification,
vol. 1
4. Turney PD (2002) Thumbs up or thumbs down? semantic ori-
entation applied to unsupervised classification of reviews. arXiv
preprint arXiv:cs/0212032
5. Pang B, Lee L, Vaithyanathan S (2002) Thumbs up? sentiment
classification using machine learning techniques. arXiv preprint
arXiv:cs/0205070
6. Xu G, Meng Y, Qiu X, Yu Z, Wu X (2019) Sentiment analysis
of comment texts based on bilstm. Ieee Access 7:51522–51532
7. Luo Z, Xu H, Chen F (2019) Audio sentiment analysis by
heterogeneous signal features learned from utterance-based
parallel neural network. In: AffCon@ AAAI
8. Poria S, Cambria E, Hazarika D, Majumder N, Zadeh A, Mor-
ency L-P (2017) Context-dependent sentiment analysis in user-
generated videos. In: proceedings of the 55th annual meeting of
the association for computational linguistics (vol. 1: Long
Papers), pp 873–883
9. Alharbi ASM, Doncker E (2019) Twitter sentiment analysis
with a deep neural network: An enhanced approach using user
behavioral information. Cogn Syst Res 54:50–61
10. Wei J, Liao J, Yang Z, Wang S, Zhao Q (2020) Bilstm with
multi-polarity orthogonal attention for implicit sentiment anal-
ysis. Neurocomputing 383:165–173
11. Bouazizi M, Ohtsuki T (2019) Multi-class sentiment analysis on
twitter: classification performance and challenges. Big Data
Mining Anal 2(3):181–194
12. Boon-Itt S, Skunkan Y (2020) Public perception of the covid-19
pandemic on twitter: sentiment analysis and topic modeling
study. JMIR Public Health Surveill 6(4):21978
13. Champagne-Langabeer T, Swank MW, Manas S, Si Y, Roberts
K (2021) Dramatic increases in telehealth-related tweets during
the early covid-19 pandemic: a sentiment analysis. Healthcare,
vol 9. Multidisciplinary Digital Publishing Institute, Basel,
p 634
14. Biswas S, Ghosh A, Chakraborty S, Roy S, Bose R (2020) Scope
of sentiment analysis on news articles regarding stock market
and gdp in struggling economic condition. Int J 8(7):3594–609
15. Almurtadha Y, Ghaleb M (2021) Sentiment analysis to measure
public response to online education during coronavirus pan-
demic. In: 2021 national computing colleges conference
(NCCC), pp 1–5. IEEE
16. Zhang L, Wang S, Liu B (2018) Deep learning for sentiment
analysis: a survey. Wiley Interdiscip Rev Data Mining Knowl
Discov 8(4):1253
17. Shirsat VS, Jagdale RS, Deshmukh S (2017) Document level
sentiment analysis from news articles. In: 2017 international
conference on computing, communication, control and
automation (ICCUBEA), pp1–4. IEEE
18. Pradhan VM, Vala J, Balani P (2016) A survey on sentiment
analysis algorithms for opinion mining. Int J Comput Appl
133(9):7–11
19. Khoo CS, Johnkhan SB (2018) Lexicon-based sentiment anal-
ysis: comparative evaluation of six sentiment lexicons. J Inf Sci
44(4):491–511
20. Naresh A, Venkata Krishna P (2021) An efficient approach for
sentiment analysis using machine learning algorithm. Evol Intel
14:725–731
21. Rani S, Kumar P (2019) Deep learning based sentiment analysis
using convolution neural network. Arab J Sci Eng
44(4):3305–3314
22. Yang L, Li Y, Wang J, Sherratt RS (2020) Sentiment analysis
for e-commerce product reviews in chinese based on sentiment
lexicon and deep learning. IEEE Access 8:23522–23530
23. Sravya G, Sreedevi M (2020) Genetic optimization in hybrid
level sentiment analysis for opinion classification. Int J Adv
Trends Comput Sci Eng 9:1440–1445
24. Tang D, Qin B, Liu T (2015) Deep learning for sentiment
analysis: successful approaches and future challenges. Wiley
Interdiscip Rev Data Mining Knowl Discov 5(6):292–303
25. Liu B (2012) Sentiment analysis and opinion mining. Synth Lect
Human Lang Technol 5(1):1–167
Neural Computing and Applications
123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
26. Tubishat M, Idris N, Abushariah MA (2018) Implicit aspect
extraction in sentiment analysis: review, taxonomy, oppportu-
nities, and open challenges. Inf Process Manage 54(4):545–563
27. Angiani G, Ferrari L, Fontanini T, Fornacciari P, Iotti E,
Magliani F, Manicardi S (2016) A comparison between pre-
processing techniques for sentiment analysis in twitter. In:
KDWeb
28. Alam S, Yao N (2019) The impact of preprocessing steps on the
accuracy of machine learning algorithms in sentiment analysis.
Comput Math Organ Theory 25(3):319–335
29. Pradha S, Halgamuge MN, Vinh NTQ (2019) Effective text data
preprocessing technique for sentiment analysis in social media
data. In: 2019 11th international conference on knowledge and
systems engineering (KSE), pp 1–8. IEEE
30. Oliveira DN, Merschmann LHC (2021) Joint evaluation of
preprocessing tasks with classifiers for sentiment analysis in
Brazilian portuguese language. Multimed Tools Appl
80:15391–15412
31. Cirqueira D, Pinheiro MF, Jacob A, Lobato F, Santana A (2018)
A literature review in preprocessing for sentiment analysis for
Brazilian Portuguese social media. In: 2018 IEEE/WIC/ACM
international conference on web intelligence (WI), pp 746–749.
IEEE
32. Deniz A, Kiziloz HE (2017) Effects of various preprocessing
techniques to Turkish text categorization using n-gram features.
In: 2017 international conference on computer science and
engineering (UBMK), pp 655–660. IEEE
33. Ahuja R, Chug A, Kohli S, Gupta S, Ahuja P (2019) The impact
of features extraction on the sentiment analysis. Procedia
Comput Sci 152:341–348
34. Waykole RN, Thakare AD (2018) A review of feature extraction
methods for text classification. Int J Adv Eng Res Dev
5(04):351–4
35. Mikolov T, Sutskever I, Chen K, Corrado GS, Dean J (2013)
Distributed representations of words and phrases and their
compositionality. Adv Neural Inf Process Syst 26:3111–3119
36. Aoumeur NE, Li Z, Alshari EM (2023) Improving the polarity
of text through word2vec embedding for primary classical
Arabic sentiment analysis. Neural Process Lett 55:1–16
37. Pennington J, Socher R, Manning CD (2014) Glove: global
vectors for word representation. In: proceedings of the 2014
conference on empirical methods in natural language processing
(EMNLP), pp 1532–1543
38. Xiaoyan L, Raga RC, Xuemei S et al (2022) Glove-cnn-bilstm
model for sentiment analysis on text reviews. J Sens 2022:12
39. Yanan Z, Dagang T (2018) Research on text classification based
on glove and svm. Softw Guide 17(6):45–48
40. Bojanowski P, Grave E, Joulin A, Mikolov T (2017) Enriching
word vectors with subword information. Trans Assoc Comput
Linguist 5:135–146
41. Kowsari K, Jafari Meimandi K, Heidarysafa M, Mendu S,
Barnes L, Brown D (2019) Text classification algorithms: a
survey. Information 10(4):150
42. Nguyen HN, Teerakanok S, Inomata A, Uehara T (2021) The
comparison of word embedding techniques in rnns for vulner-
ability detection. In: ICISSP, pp 109–120
43. Dang NC, Moreno-Garcı
´a MN, Prieta F (2020) Sentiment
analysis based on deep learning: a comparative study. Elec-
tronics 9(3):483
44. Sasaki S, Heinzerling B, Suzuki J, Inui K (2023) Examining the
effect of whitening on static and contextualized word embed-
dings. Inf Process Manage 60(3):103272
45. Radford A, Narasimhan K, Salimans T, Sutskever I et al (2018)
Improving language understanding by generative pre-training.
Preprint, 1–12
46. Nath S, Marie A, Ellershaw S, Korot E, Keane PA (2022) New
meaning for nlp: the trials and tribulations of natural language
processing with gpt-3 in ophthalmology. Br J Ophthalmol
106(7):889–892
47. Lee J, Toutanova K (2018) Pre-training of deep bidirectional
transformers for language understanding. arXiv preprint arXiv:
1810.04805
48. Barbouch M, Verberne S, Verhoef T (2021) Wn-bert: integrat-
ing wordnet and bert for lexical semantics in natural language
understanding. Comput Linguist Neth J 11:105–124
49. Lee J-S, Hsiang J (2020) Patent claim generation by fine-tuning
openai gpt-2. World Patent Inf 62:101983
50. Mars M (2022) From word embeddings to pre-trained language
models: a state-of-the-art walkthrough. Appl Sci 12(17):8805
51. Wang J, Yu L-C, Lai KR, Zhang X (2019) Tree-structured
regional cnn-lstm model for dimensional sentiment analysis.
IEEE/ACM Trans Audio Speech Lang Process 28:581–591
52. Gao Z, Feng A, Song X, Wu X (2019) Target-dependent sen-
timent classification with bert. Ieee Access 7:154290–154299
53. Li X, Fu X, Xu G, Yang Y, Wang J, Jin L, Liu Q, Xiang T
(2020) Enhancing bert representation with context-aware
embedding for aspect-based sentiment analysis. IEEE Access
8:46868–46876
54. Nguyen TH, Shirai K, Velcin J (2015) Sentiment analysis on
social media for stock movement prediction. Expert Syst Appl
42(24):9603–9611
55. Panthong R, Srivihok A (2015) Wrapper feature subset selection
for dimension reduction based on ensemble learning algorithm.
Procedia Comput Sci 72:162–169
56. Omuya EO, Okeyo GO, Kimwele MW (2021) Feature selection
for classification using principal component analysis and infor-
mation gain. Expert Syst Appl 174:114765
57. Sa
´nchez-Maron
˜o N, Alonso-Betanzos A, Calvo-Este
´vez RM
(2009) A wrapper method for feature selection in multiple
classes datasets. In: international work-conference on artificial
neural networks, pp 456–463. Springer
58. Solorio-Ferna
´ndez S, Carrasco-Ochoa JA, Martı
´nez-Trinidad JF
(2020) A review of unsupervised feature selection methods.
Artif Intell Rev 53(2):907–948
59. Birjali M, Beni-Hssane A, Erritali M (2017) Machine learning
and semantic sentiment analysis based algorithms for suicide
sentiment prediction in social networks. Procedia Comput Sci
113:65–72
60. Chandrashekar G, Sahin F (2014) A survey on feature selection
methods. Comput Electr Eng 40(1):16–28
61. Jime
´nez-Cordero A, Morales JM, Pineda S (2021) A novel
embedded min-max approach for feature selection in nonlinear
support vector machine classification. Eur J Oper Res
293(1):24–35
62. Remeseiro B, Bolon-Canedo V (2019) A review of feature
selection methods in medical applications. Comput Biol Med
112:103375
63. Shah S, Shabbir H, Rehman S, Waqas M (2020) A comparative
study of feature selection approaches: 2016–2020. Int J Sci Eng
Res 11(2):469
64. Liu H, Yu L (2005) Toward integrating feature selection algo-
rithms for classification and clustering. IEEE Trans Knowl Data
Eng 17(4):491–502
65. Zhao H, Liu Z, Yao X, Yang Q (2021) A machine learning-
based sentiment analysis of online product reviews with a novel
term weighting and feature selection approach. Inf Process
Manage 58(5):102656
66. Yusta SC (2009) Different metaheuristic strategies to solve the
feature selection problem. Pattern Recogn Lett 30(5):525–534
Neural Computing and Applications
123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
67. Al-Qablan TA, Noor MHM, Al-Betar MA, Khader AT (2023)
Improved binary gray wolf optimizer based on adaptive b-hill
climbing for feature selection. IEEE Access
68. Tahir M, Tubaishat A, Al-Obeidat F, Shah B, Halim Z, Waqas
M (2020) A novel binary chaotic genetic algorithm for feature
selection and its utility in affective computing and healthcare.
Neural Comput Appl, 1–22
69. Ahmad SR, Rodzi MZM, Nurhafizeh N, Yusop M, Ismail S
(2019) A review of feature selection and sentiment analysis
technique in issues of propaganda. Int J Adv Comput Sci Appl
10(11):240–245
70. Addo-Tenkorang R, Helo PT (2016) Big data applications in
operations/supply-chain management: a literature review.
Comput Ind Eng 101:528–543
71. Zaslavsky A, Perera C, Georgakopoulos D (2013) Sensing as a
service and big data. arXiv preprint arXiv:1301.0159
72. Shayaa S, Jaafar NI, Bahri S, Sulaiman A, Wai PS, Chung YW,
Piprani AZ, Al-Garadi MA (2018) Sentiment analysis of big
data: methods, applications, and open challenges. IEEE Access
6:37807–37827
73. Mukhtar N, Khan MA, Chiragh N (2018) Lexicon-based
approach outperforms supervised machine learning approach for
urdu sentiment analysis in multiple domains. Telemat Inform
35(8):2173–2183
74. Catelli R, Pelosi S, Esposito M (2022) Lexicon-based vs. bert-
based sentiment analysis: a comparative study in Italian. Elec-
tronics 11(3):374
75. Reddy Y, Viswanath P, Reddy BE (2018) Semi-supervised
learning: a brief review. Int J Eng Technol 7(1.8):81
76. Jasmir J, Nurmaini S, Tutuko B (2021) Fine-grained algorithm
for improving knn computational performance on clinical trials
text classification. Big Data Cogn Comput 5(4):60
77. Didi Y, Walha A, Wali A (2022) Covid-19 tweets classification
based on a hybrid word embedding method. Big Data Cogn
Comput 6(2):58
78. Zaks G, Katz G (2022) Recom: a deep reinforcement learning
approach for semi-supervised tabular data labeling. Inf Sci
589:321–340
79. Riyadh M, Shafiq MO (2021) Towards multi-class sentiment
analysis with limited labeled data. In: 2021 IEEE international
conference on big data (Big Data), pp 4955–4964. IEEE
80. Ebrahimi P, Basirat M, Yousefi A, Nekmahmud M, Gholampour
A, Fekete-Farkas M (2022) Social networks marketing and
consumer purchase behavior: the combination of sem and
unsupervised machine learning approaches. Big Data Cogn
Comput 6(2):35
81. Garcı
´a-Pablos A, Cuadros M, Rigau G (2018) W2vlda: almost
unsupervised system for aspect based sentiment analysis. Expert
Syst Appl 91:127–137
82. Yadav A, Jha C, Sharan A, Vaish V (2020) Sentiment analysis
of financial news using unsupervised approach. Procedia Com-
put Sci 167:589–598
83. Pandey AC, Rajpoot DS, Saraswat M (2017) Twitter sentiment
analysis using hybrid cuckoo search method. Inf Process Man-
age 53(4):764–779
84. Kaur G, Kaushik A, Sharma S (2019) Cooking is creating
emotion: a study on hinglish sentiments of youtube cookery
channels using semi-supervised approach. Big Data Cogn
Comput 3(3):37
85. Macrohon JJE, Villavicencio CN, Inbaraj XA, Jeng J-H (2022)
A semi-supervised approach to sentiment analysis of tweets
during the 2022 philippine presidential election. Information
13(10):484
86. Cruz Paulino JL, Antoja Almirol LC, Cruz Favila JM, Lor-
ia Aquino KAG, Hernandez De La Cruz A, Roxas RE (2021)
Multilingual sentiment analysis on short text document using
semi-supervised machine learning. In: 2021 5th international
conference on E-society, E-education and E-technology,
pp 164–170
87. Gupta N, Agrawal R (2020) Application and techniques of
opinion mining. Hybrid computational intelligence. Elsevier,
Amsterdam, pp 1–23
88. Assiri A, Emam A, Al-Dossari H (2018) Towards enhancement
of a lexicon-based approach for Saudi dialect sentiment analysis.
J Inf Sci 44(2):184–202
89. Zabha NI, Ayop Z, Anawar S, Hamid E, Abidin ZZ (2019)
Developing cross-lingual sentiment analysis of Malay twitter
data using lexicon-based approach. Int J Adv Comput Sci Appl
10(1):346–351
90. YURTALAN G, Koyuncu M, TURHAN C¸ (2019) A polarity
calculation approach for lexicon-based turkish sentiment anal-
ysis. Turk J Electr Eng Comput Sci 27(2):1325–1339
91. Wunderlich F, Memmert D (2020) Innovative approaches in
sports science-lexicon-based sentiment analysis as a tool to
analyze sports-related twitter communication. Appl Sci
10(2):431
92. Mukhtar N, Khan MA (2020) Effective lexicon-based approach
for Urdu sentiment analysis. Artif Intell Rev 53(4):2521–2548
93. Mehmood Y, Balakrishnan V (2020) An enhanced lexicon-
based approach for sentiment analysis: a case study on illegal
immigration. Online information review
94. Aloqaily A, Alhassan M, Salah K, Elshqeirat B, Almashagbah
M, Al Hussein Bin Abdullah P (2020) Sentiment analysis for
arabic tweets datasets: lexicon-based and machine learning
approaches. J Theor Appl Inf Technol 98(4):612–623
95. Piryani R, Piryani B, Singh VK, Pinto D (2020) Sentiment
analysis in nepali: exploring machine learning and lexicon-
based approaches. J Intell Fuzzy Syst (Preprint), 1–12
96. Chaovalit P, Zhou L (2005) Movie review mining: a comparison
between supervised and unsupervised classification approaches.
In: proceedings of the 38th annual hawaii international confer-
ence on system sciences, pp 112–112. IEEE
97. Liu Z, Dong X, Guan Y, Yang J (2013) Reserved self-training: a
semi-supervised sentiment classification method for chinese
microblogs. In: proceedings of the sixth international joint
conference on natural language processing, pp 455–462
98. Singh J, Singh G, Singh R (2017) Optimization of sentiment
analysis using machine learning classifiers. HCIS 7(1):1–12
99. Hew KF, Hu X, Qiao C, Tang Y (2020) What predicts student
satisfaction with moocs: a gradient boosting trees supervised
machine learning and sentiment analysis approach. Comput Edu
145:103724
100. Ghiassi M, Lee S (2018) A domain transferable lexicon set for
twitter sentiment analysis using a supervised machine learning
approach. Expert Syst Appl 106:197–216
101. Du J, Xu J, Song H, Liu X, Tao C (2017) Optimization on
machine learning based approaches for sentiment analysis on
hpv vaccines related tweets. J Biomed Semant 8(1):1–7
102. Mukhtar N, Khan MA (2018) Urdu sentiment analysis using
supervised machine learning approach. Int J Pattern Recognit
Artif Intell 32(02):1851001
103. Basha CB, Somasundaram K (2019) A comparative study of
twitter sentiment analysis using machine learning algorithms in
big data. Int J Rec Technol Eng 8(1):591–599
104. Hausler J, Ruscheinsky J, Lang M (2018) News-based sentiment
analysis in real estate: a machine learning approach. J Prop Res
35(4):344–371
105. Valencia F, Go
´mez-Espinosa A, Valde
´s-Aguirre B (2019) Price
movement prediction of cryptocurrencies using sentiment anal-
ysis and machine learning. Entropy 21(6):589
Neural Computing and Applications
123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
106. Ghosh M, Sanyal G (2018) An ensemble approach to stabilize
the features for multi-domain sentiment analysis using super-
vised machine learning. J Big Data 5(1):1–25
107. Nandal N, Tanwar R, Pruthi J (2020) Machine learning based
aspect level sentiment analysis for amazon products. Spat Inf
Res 28(5):601–607
108. Baid P, Gupta A, Chaplot N (2017) Sentiment analysis of movie
reviews using machine learning techniques. Int J Comput Appl
179(7):45–49
109. Soumya S, Pramod K (2020) Sentiment analysis of Malayalam
tweets using machine learning techniques. ICT Express
6(4):300–305
110. Al-Moslmi T, Omar N, Albared M, Alshabi A (2017) Enhanced
Malay sentiment analysis with an ensemble classification
machine learning approach. J f Eng Appl Sci 12(20):5226–5232
111. Buladaco M, Buladaco J, Cantero L (2020) Sentiments analysis
on public land transport infrastructure in Davao region using
machine learning algorithms. Int J Adv Trends Comput Sci Eng
9(1):685–690
112. Raza H, Faizan M, Hamza A, Mushtaq A, Akhtar N (2019)
Scientific text sentiment analysis using machine learning tech-
niques. Int J Adv Comput Sci Appl 10(12):157–165
113. Bhargav PS, Reddy GN, Chand RR, Pujitha K, Mathur A (2019)
Sentiment analysis for hotel rating using machine learning
algorithms. Int J Innov Technol Explor Eng (IJITEE)
8(6):1225–1228
114. Hercig T, Brychcı
´n T, Svoboda L, Konkol M, Steinberger J
(2016) Unsupervised methods to improve aspect-based senti-
ment analysis in czech. Comput Sistemas 20(3):365–375
115. Azzouza N, Akli-Astouati K, Oussalah A, Bachir SA (2017) A
real-time twitter sentiment analysis using an unsupervised
method. In: proceedings of the 7th international conference on
web intelligence, mining and semantics, pp 1–10
116. Odbal Wang Z (2016) A semi-supervised method for phrase-
level sentiment analysis. Moshi Shibie yu Rengong Zhineng/
Pattern Recognit Artif Intell 29(4):289–297
117. Socher R, Pennington J, Huang EH, Ng AY, Manning CD
(2011) Semi-supervised recursive autoencoders for predicting
sentiment distributions. In: EMNLP 2011-conference on
empirical methods in natural language processing, proceedings
of the conference, pp 151–161
118. Madhoushi Z, Hamdan AR, Zainudin S (2015) Sentiment
analysis techniques in recent works. In: 2015 science and
information conference (SAI), pp 288–291. IEEE
119. Rehman AU, Malik AK, Raza B, Ali W (2019) A hybrid cnn-
lstm model for improving accuracy of movie reviews sentiment
analysis. Multimed Tools Appl 78(18):26597–26613
120. Joseph J, Vineetha S, Sobhana N (2022) A survey on deep
learning based sentiment analysis. Mater Today Procee
58:456–60
121. Araque O, Corcuera-Platas I, Sa
´nchez-Rada JF, Iglesias CA
(2017) Enhancing deep learning sentiment analysis with
ensemble techniques in social applications. Expert Syst Appl
77:236–246
122. Devipriya K, Prabha D, Pirya V, Sudhakar S (2020) Deep
learning sentiment analysis for recommendations in social
applications. Int J Sci Technol Res 9(1):3812–3815
123. Yadav A, Vishwakarma DK (2020) A deep learning architecture
of ra-dlnet for visual sentiment analysis. Multimed Syst
26:431–451
124. Park S, Woo J (2019) Gender classification using sentiment
analysis and deep learning in a health web forum. Appl Sci
9(6):1249
125. Rosa RL, Schwartz GM, Ruggiero WV, Rodrı
´guez DZ (2018) A
knowledge-based recommendation system that includes
sentiment analysis and deep learning. IEEE Trans Industr Inf
15(4):2124–2135
126. Colo
´n-Ruiz C, Segura-Bedmar I (2020) Comparing deep
learning architectures for sentiment analysis on drug reviews.
J Biomed Inform 110:103539
127. Sohangir S, Wang D, Pomeranets A, Khoshgoftaar TM (2018)
Big data: deep learning for financial sentiment analysis. J Big
Data 5(1):1–25
128. Feizollah A, Ainin S, Anuar NB, Abdullah NAB, Hazim M
(2019) Halal products on twitter: data extraction and sentiment
analysis using stack of deep learning algorithms. IEEE Access
7:83354–83362
129. Terra Vieira S, Lopes Rosa R, Zegarra Rodrı
´guez D, Arjona
Ramı
´rez M, Saadi M, Wuttisittikulkij L (2021) Q-meter: quality
monitoring system for telecommunication services based on
sentiment analysis using deep learning. Sensors 21(5):1880
130. Li L, Goh T-T, Jin D (2020) How textual quality of online
reviews affect classification performance: a case of deep learn-
ing sentiment analysis. Neural Comput Appl 32(9):4387–4415
131. Ghorbani M, Bahaghighat M, Xin Q, O
¨zen F (2020) Convl-
stmconv network: a deep learning approach for sentiment
analysis in cloud computing. J Cloud Comput 9(1):1–12
132. Garg A, Duhan N (2020) Sarcasm detection on twitter data using
support vector machine. ICTACT J Soft Comput
10(4):2165–2170
133. Xu F, Zhang X, Xin Z, Yang A (2019) Investigation on the
Chinese text sentiment analysis based on convolutional neural
networks in deep learning. Comput Mater Contin 58(3):697–709
134. Paredes-Valverde MA, Colomo-Palacios R, Salas-Za
´rate MdP,
Valencia-Garcı
´a R et al (2017) Sentiment analysis in Spanish for
improvement of products and services: a deep learning
approach. Sci Program 2017:1–6
135. Kapoc
ˇiu
¯t_
e-Dzikien_
e J, Damas
ˇevic
ˇius R, Woz
´niak M (2019)
Sentiment analysis of Lithuanian texts using traditional and deep
learning approaches. Computers 8(1):4
136. Ombabi AH, Ouarda W, Alimi AM (2020) Deep learning cnn-
lstm framework for Arabic sentiment analysis using textual
information shared in social networks. Soc Netw Anal Min
10(1):1–13
137. Mohammed A, Kora R (2019) Deep learning approaches for
Arabic sentiment analysis. Soc Netw Anal Min 9(1):1–12
138. Konate A, Du R (2018) Sentiment analysis of code-mixed
Bambara-French social media text using deep learning tech-
niques. Wuhan Univ J Nat Sci 23(3):237–243
139. Dastgheib MB, Koleini S, Rasti F (2020) The application of
deep learning in Persian documents sentiment analysis. Int J Inf
Sci Manage (IJISM) 18(1):1–15
140. Nezhad ZB, Deihimi MA (2019) A combined deep learning
model for Persian sentiment analysis. IIUM Eng J
20(1):129–139
141. Ong JY, Mun’im Ahmad Zabidi M, Ramli N, Sheikh UU (2020)
Sentiment analysis of informal Malay tweets with deep learning.
IAES Int J Artif Intell 9(2):212
142. Tam S, Said RB, Tanrio
¨ver O
¨O
¨(2021) A convbilstm deep
learning model-based approach for twitter sentiment classifica-
tion. IEEE Access 9:41283–41293
143. Lin Y, Li J, Yang L, Xu K, Lin H (2020) Sentiment analysis
with comparison enhanced deep neural network. IEEE Access
8:78378–78384
144. Mohammadi A, Shaverizade A (2021) Ensemble deep learning
for aspect-based sentiment analysis. Int J Nonlinear Anal Appl
12:29–38
145. Devlin J, Chang M-W, Lee K, Toutanova K (2018) Bert: Pre-
training of deep bidirectional transformers for language under-
standing. arXiv preprint arXiv:1810.04805
Neural Computing and Applications
123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
146. Mathew L, Bindu V (2022) Efficient classification techniques in
sentiment analysis using transformers. In: international confer-
ence on innovative computing and communications,
pp 849–862. Springer
147. Mishev K, Gjorgjevikj A, Vodenska I, Chitkushev LT, Trajanov
D (2020) Evaluation of sentiment analysis in finance: from
lexicons to transformers. IEEE Access 8:131662–131682
148. Liu Y, Ott M, Goyal N, Du J, Joshi M, Chen D, Levy O, Lewis
M, Zettlemoyer L, Stoyanov V (2019) Roberta: a robustly
optimized bert pretraining approach. arXiv preprint arXiv:1907.
11692
149. Zeng B, Yang H, Xu R, Zhou W, Han X (2019) Lcf: a local
context focus mechanism for aspect-based sentiment classifica-
tion. Appl Sci 9(16):3389
150. Durairaj AK, Chinnalagu A (2021) Transformer based contex-
tual model for sentiment analysis of customer reviews: a fine-
tuned bert a sequence learning bert model for sentiment analysis.
Int J Adv Comput Sci Appl 12(11):474–480
151. Iqbal F, Hashmi JM, Fung BC, Batool R, Khattak AM, Aleem S,
Hung PC (2019) A hybrid framework for sentiment analysis
using genetic algorithm based feature reduction. IEEE Access
7:14637–14652
152. Alqahtani AS, Saravanan P, Maheswari M, Alshmrany S,
Alsarrayrih H (2021) Hybrid grass bee optimization-multikernal
extreme learning classifier: multimodular fusion strategy and
optimal feature selection for multimodal sentiment analysis in
social media videos. Concurr Comput Pract Exp 33(16):e6259
153. Raviya K, Mary Vennila S (2020) A hybrid deep learning
approach for sentiment analysis using cnn and improved svm
with multi objective swarm optimization for domain indepen-
dent datasets. Int J Adv Trends Comput Sci Eng 9:3200–3206
154. Shekhawat SS, Shringi S, Sharma H (2021) Twitter sentiment
analysis using hybrid spider monkey optimization method.
Evolut Intell 14:1307–1316
155. Madani Y, Erritali M, Bengourram J, Sailhan F (2020) A hybrid
multilingual fuzzy-based approach to the sentiment analysis
problem using sentiwordnet. Internat J Uncertain Fuzziness
Knowl Based Syst 28(03):361–390
156. Kumar A, Khorwal R (2017) Firefly algorithm for feature
selection in sentiment analysis. Computational intelligence in
data mining. Springer, Berlin, pp 693–703
157. Tubishat M, Abushariah MA, Idris N, Aljarah I (2019) Improved
whale optimization algorithm for feature selection in Arabic
sentiment analysis. Appl Intell 49(5):1688–1707
158. Rahab H, Haouassi H, Souidi MEH, Bakhouche A, Mahdaoui R,
Bekhouche M (2022) A modified binary rat swarm optimization
algorithm for feature selection in Arabic sentiment analysis.
Arab J Sci Eng, 1–28
159. Mehbodniya A, Rao MV, David LG, Nigel KGJ, Vennam P
(2022) Online product sentiment analysis using random evolu-
tionary whale optimization algorithm and deep belief network.
Pattern Recogn Lett 159:1–8
160. Shaddeli A, Soleimanian Gharehchopogh F, Masdari M, Solouk
V (2022) An improved African vulture optimization algorithm
for feature selection problems and its application of sentiment
analysis on movie reviews. Big Data Cogn Comput 6(4):104
161. Elshakankery K, Ahmed MF (2019) Hilatsa: a hybrid incre-
mental learning approach for Arabic tweets sentiment analysis.
Egypt Inf J 20(3):163–171
162. Ers¸ahI
˙n B, Aktas¸O
¨, Kilinc D, Ers¸ahI
˙n M (2019) A hybrid
sentiment analysis method for Turkish. Turk J Electr Eng
Comput Sci 27(3):1780–1793
163. Sinha S, Saxena K, Joshi N (2019) Sentiment analysis of face-
book posts using hybrid method. Int J Recent Technol Eng
8:2421–2428
164. Mendon S, Dutta P, Behl A, Lessmann S (2021) A hybrid
approach of machine learning and lexicons to sentiment analy-
sis: enhanced insights from twitter data of natural disasters. Inf
Syst Front. https://doi.org/10.1007/s10796-021-10107-x
165. Learning SM (2020) Hybrid model for twitter data sentiment
analysis based on ensemble of dictionary based classifier and
stacked machine learning classifiers-svm, knn and c50. J Theor
Appl Inf Technol 98(04):624–635
166. Dashtipour K, Gogate M, Li J, Jiang F, Kong B, Hussain A
(2020) A hybrid Persian sentiment analysis framework: inte-
grating dependency grammar based rules and deep neural net-
works. Neurocomputing 380:1–10
167. Wu C, Wu F, Wu S, Yuan Z, Huang Y (2018) A hybrid unsu-
pervised method for aspect term and opinion target extraction.
Knowl-Based Syst 148:66–73
168. Ray P, Chakrabarti A (2020) A mixed approach of deep learning
method and rule-based method to improve aspect level senti-
ment analysis. Appl Comput Inf. https://doi.org/10.1016/J.ACI.
2019.02.002
169. Srinidhi H, Siddesh G, Srinivasa K (2020) A hybrid model using
Malstm based on recurrent neural networks with support vector
machines for sentiment analysis. Eng Appl Sci Res
47(3):232–240
170. Kumar A, Srinivasan K, Cheng W-H, Zomaya AY (2020)
Hybrid context enriched deep learning model for fine-grained
sentiment analysis in textual and visual semiotic modality social
data. Inf Process Manage 57(1):102141
171. Raviya K, Mary Vennila S (2021) An implementation of hybrid
enhanced sentiment analysis system using spark ml pipeline: a
big data analytics framework. Int J Adv Comput Sci Appl
12:323–329
172. Hassonah MA, Al-Sayyed R, Rodan A, Ala’M A-Z, Aljarah I,
Faris H (2020) An efficient hybrid filter and evolutionary
wrapper approach for sentiment analysis of various topics on
twitter. Knowl Based Syst 192:105353
173. Omar A, Hamouda WI (2021) A sentiment analysis of Egypt’s
new real estate registration law on facebook. Int J Adv Comput
Sci Appl 12(4):656–663
174. Bashir S, Bano S, Shueb S, Gul S, Mir AA, Ashraf R, Noor N
et al (2021) Twitter chirps for syrian people: sentiment analysis
of tweets related to syria chemical attack. Int J Disaster Risk
Red 62:102397
175. Villavicencio C, Macrohon JJ, Inbaraj XA, Jeng J-H, Hsieh J-G
(2021) Twitter sentiment analysis towards covid-19 vaccines in
the Philippines using naı
¨ve bayes. Information 12(5):204
176. Sailunaz K, Alhajj R (2019) Emotion and sentiment analysis
from twitter text. J Comput Sci 36:101003
177. Wang L, Niu J, Yu S (2019) Sentidiff: combining textual
information and sentiment diffusion patterns for twitter senti-
ment analysis. IEEE Trans Knowl Data Eng 32(10):2026–2039
178. Gabarron E, Dorronzoro E, Rivera-Romero O, Wynn R (2019)
Diabetes on twitter: a sentiment analysis. J Diabetes Sci Technol
13(3):439–444
179. O
¨ztu
¨rk N, Ayvaz S (2018) Sentiment analysis on twitter: a text
mining approach to the Syrian refugee crisis. Telematics Inform
35(1):136–147
180. Bouazizi M, Ohtsuki T (2017) A pattern-based approach for
multi-class sentiment analysis in twitter. IEEE Access
5:20617–20639
181. Daniel M, Neves RF, Horta N (2017) Company event popularity
for financial markets using twitter and sentiment analysis.
Expert Syst Appl 71:111–124
182. Zafra SMJ, Valdivia MTM, Camara EM, Lopez LAU (2017)
Studying the scope of negation for Spanish sentiment analysis
on twitter. IEEE Trans Affect Comput 10(1):129–141
Neural Computing and Applications
123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
183. Prabha VD, Rathipriya R (2022) Competitive capsule network
based sentiment analysis on twitter covid’19 vaccines. J Web
Eng 21:1583–1602
184. Musleh DA, Alkhales TA, Almakki RA, Alnajim SE, Almar-
shad SK, Alhasaniah RS, Aljameel SS, Almuqhim AA (2022)
Twitter Arabic sentiment analysis to detect depression using
machine learning. Comput Mater Contin 71:3463–3477
185. Nezhad ZB, Deihimi MA (2022) Twitter sentiment analysis
from Iran about covid 19 vaccine. Diabetes Metab Syndr Clin
Res Rev 16(1):102367
186. Hassan MK, Hudaefi FA, Caraka RE (2022) Mining netizen’s
opinion on cryptocurrency: sentiment analysis of twitter data.
Stud Econ Financ 39(3):365–385
187. Ribeiro LA, Cinalli D, Garcia ACB (2021) Discovering adverse
drug reactions from twitter: A sentiment analysis perspective.
In: 2021 IEEE 24th international conference on computer sup-
ported cooperative work in design (CSCWD), pp 1172–1177.
IEEE
188. Drus Z, Khalid H (2019) Sentiment analysis in social media and
its application: systematic literature review. Procedia Comput
Sci 161:707–714
189. Roccetti M, Marfia G, Salomoni P, Prandi C, Zagari RM,
Kengni FLG, Bazzoli F, Montagnani M (2017) Attitudes of
crohn’s disease patients: infodemiology case study and senti-
ment analysis of facebook and twitter posts. JMIR Public Health
Surveill 3(3):7004
190. Rouby I, Badawy M, Nour M, Hegazi N (2018) Performance
evaluation of an adopted sentiment analysis model for Arabic
comments from the facebook. J Theor Appl Inf Technol
96(21):7098–7112
191. Klimiuk K, Czoska A, Biernacka K, Balwicki Ł (2021) Vaccine
misinformation on social media-topic-based content and senti-
ment analysis of polish vaccine-deniers’ comments on facebook.
Hum Vaccines Immunother 17(7):2026–2035
192. Alvarez G, Choi J, Strover S (2020) Good news, bad news: a
sentiment analysis of the 2016 election Russian facebook ads.
Good Syst Pub Res 14:3027–3053
193. Alnawas A, Arici N (2019) Sentiment analysis of Iraqi Arabic
dialect on facebook based on distributed representations of
documents. ACM Trans Asian Low Resour Lang Inf Process
(TALLIP) 18(3):1–17
194. Poecze F, Ebster C, Strauss C (2019) Let’s play on facebook:
using sentiment analysis and social media metrics to measure
the success of youtube gamers’ post types. Pers Ubiquitous
Comput 1–10
195. Hajhmida MB, Oueslati O (2021) Predicting mobile application
breakout using sentiment analysis of facebook posts. J Inf Sci
47(4):502–516
196. Nahar KM, Jaradat A, Atoum MS, Ibrahim F (2020) Sentiment
analysis and classification of Arab Jordanian facebook com-
ments for Jordanian telecom companies using lexicon-based
approach and machine learning. Jordanian J Comput Inf Technol
(JJCIT) 6(3):247–262
197. Hand LC, Ching BD (2020) Maintaining neutrality: a sentiment
analysis of police agency facebook pages before and after a fatal
officer-involved shooting of a citizen. Gov Inf Q 37(1):101420
198. Mishra MS, Sharma RW (2019) Brand crisis-sentiment analysis
of user-generated comments about@ maggi on facebook. Corp
Reput Rev 22(2):48–60
199. Rodriguez A, Chen Y-L, Argueta C (2022) Fadohs: framework
for detection and integration of unstructured data of hate speech
on facebook using sentiment and emotion analysis. IEEE Access
10:22400–22419
200. Kewsuwun N, Kajornkasirat S (2022) A sentiment analysis
model of agritech startup on facebook comments using naive
bayes classifier. Int J Electr Comput Eng (2088-8708)
12(3):2829–2838
201. Boulesnane A, Meshoul S, Aouissi K (2022) Influenza-like ill-
ness detection from Arabic facebook posts based on sentiment
analysis and 1d convolutional neural network. Mathematics
10(21):4089
202. Li Z, Li R, Jin G (2020) Sentiment analysis of danmaku videos
based on naı
¨ve bayes and sentiment dictionary. IEEE Access
8:75073–75084
203. Arunkumar P, Chandramathi S, Kannimuthu S (2019) Sentiment
analysis-based framework for assessing internet telemedicine
videos. Int J Data Anal Tech Strateg 11(4):328–336
204. Thompson JJ, Leung BH, Blair MR, Taboada M (2017) Senti-
ment analysis of player chat messaging in the video game
starcraft 2: extending a lexicon-based model. Knowl Based Syst
137:149–162
205. Bozkurt AP, Aras I (2021) Cleft lip and palate youtube videos:
content usefulness and sentiment analysis. Cleft Palate Cranio-
fac J 58(3):362–368
206. Porreca A, Scozzari F, Di Nicola M (2020) Using text mining
and sentim https://www.overleaf.com/project/
60bcb05421a73355c933f099ent analysis to analyse youtube
Italian videos concerning vaccination. BMC Public Health
20(1):1–9
207. Mi C, Li M, Wulandari AF (2022) Predicting video views of
web series based on comment sentiment analysis and improved
stacking ensemble model. Electron Commer Res, 1–28
208. Pan Z, Li X, Cui L, Zhang Z (2020) Video clip recommendation
model by sentiment analysis of time-sync comments. Multimed
Tools Appl 79(45):33449–33466
209. Soleymani M, Garcia D, Jou B, Schuller B, Chang S-F, Pantic M
(2017) A survey of multimodal sentiment analysis. Image Vis
Comput 65:3–14
210. Akhtar MS, Chauhan DS, Ghosal D, Poria S, Ekbal A, Bhat-
tacharyya P (2019) Multi-task learning for multi-modal emotion
recognition and sentiment analysis. arXiv preprint arXiv:1905.
05812
211. Li Q, Gkoumas D, Lioma C, Melucci M (2021) Quantum-in-
spired multimodal fusion for video sentiment analysis. Inf
Fusion 65:58–71
212. Tzirakis P, Chen J, Zafeiriou S, Schuller B (2021) End-to-end
multimodal affect recognition in real-world environments. Inf
Fusion 68:46–53
213. Al-Azani S, El-Alfy E-SM (2020) Enhanced video analytics for
sentiment analysis based on fusing textual, auditory and visual
information. IEEE Access 8:136843–136857
214. Chaturvedi I, Satapathy R, Cavallari S, Cambria E (2019) Fuzzy
commonsense reasoning for multimodal sentiment analysis.
Pattern Recogn Lett 125:264–270
215. Zadeh A, Chen M, Poria S, Cambria E, Morency L-P (2017)
Tensor fusion network for multimodal sentiment analysis. arXiv
preprint arXiv:1707.07250
216. Ermatita E, Abdiansah A, Rini DP, Febry F (2022) Sentiment
analysis of covid-19 using multimodal fusion neural networks.
TEM J 11(3):1316–1321
217. Zhu T, Li L, Yang J, Zhao S, Liu H, Qian J (2022) Multimodal
sentiment analysis with image-text interaction network. IEEE
Trans Multimed, 1–12
218. Mai S, Zeng Y, Zheng S, Hu H (2022) Hybrid contrastive
learning of tri-modal representation for multimodal sentiment
analysis. IEEE Trans Affect Comput, 1–12
219. Ghorbanali A, Sohrabi MK, Yaghmaee F (2022) Ensemble
transfer learning-based multimodal sentiment analysis using
weighted convolutional neural networks. Inf Process Manage
59(3):102929
Neural Computing and Applications
123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
220. Stappen L, Baird A, Cambria E, Schuller BW (2021) Sentiment
analysis and topic recognition in video transcriptions. IEEE
Intell Syst 36(2):88–95
221. Iglesias-Puzas A
´, Conde-Taboada A, Lo
´pez-Bran E (2022) A
cross-sectional study of youtube videos on mohs surgery: quality
of content and sentiment analysis. J Am Acad Dermatol
86(3):649–651
222. Deori M, Kumar V, Verma MK (2021) Analysis of youtube
video contents on koha and dspace, and sentiment analysis of
viewers’ comments. Library Hi Tech
223. IMDB Movie Reviews Datase: https://www.kaggle.com/code/
lakshmi25npathi/sentiment-analysis-of-Imdb-movie-reviews/
input
224. SemEval-2017 Task 4 Datase: http://alt.qcri.org/semeval2017/
Task4/index.php?id=data-and-tool
225. Yelp Open Dataset: https://www.yelp.com/dataset
226. Amazon review data: http://jmcauley.ucsd.edu/data/amazon/
227. UCI Reviews Datase: https://archive.ics.uci.edu/ml/datasets/
Sentiment?Labelled?Sentences
228. Aly M, Atiya A (2013) Labr: a large scale arabic book reviews
dataset. In: proceedings of the 51st annual meeting of the
association for computational linguistics (Vol. 2: Short Papers),
pp 494–498
229. Omran T, Sharef B, Grosan C, Li Y (2023) Sentiment analysis
of multilingual dataset of Bahraini dialects, Arabic, and English.
Data 8(4):68
230. Bahraini Dialects Datase: https://data.mendeley.com/datasets/
5rhw2srzjj/1
231. Nabil M, Aly M, Atiya A (2015) Astd: Arabic sentiment tweets
dataset. In: proceedings of the 2015 conference on empirical
methods in natural language processing, pp 2515–2519
232. Taleb C, Khachab M, Mokbel C, Likforman-Sulem L (2017)
IEEE 1st international workshop on Arabic script analysis and
recognition (ASAR): feature selection for an improved Parkin-
son’s disease identification based on handwriting. IEEE
233. Alayba AM, Palade V, England M, Iqbal R (2018) Improving
sentiment analysis in arabic using word representation. In: 2018
IEEE 2nd international workshop on Arabic and derived script
analysis and recognition (ASAR), pp 13–18. IEEE
234. Fujisaki H (2003) Prosody, information, and modeling-with
emphasis on tonal features of speech. In: workshop on spoken
language processing
235. Sheth T (2017) Non-verbal communication: a significant aspect
of proficient occupation. OSR J Hum Soc Sci 22(11):69–72
236. Halim Z, Rehan M (2020) On identification of driving-induced
stress using electroencephalogram signals: a framework based
on wearable safety-critical scheme and machine learning. Inf
Fusion 53:66–79
237. Ahmed N, Al Aghbari Z, Girija S (2023) A systematic survey on
multimodal emotion recognition using learning algorithms.
Intell Syst Appl 17:200171
238. Intelligence SM (2021) Disponı
´vel em: https://www.mordor
intelligence.com/industry-reports/global-clove-market. Acesso
em 29
239. Sahoo S, Routray A (2016) Emotion recognition from audio-
visual data using rule based decision level fusion. In: 2016 IEEE
students’ technology symposium (TechSym), pp 7–12. IEEE
240. Aadam Tubaishat A, Al-Obeidat F, Halim Z, Waqas M, Qayum
F (2022) Emopercept: eeg-based emotion classification through
perceiver. Soft Comput, 1–8
241. Tashu TM, Hajiyeva S, Horvath T (2021) Multimodal emotion
recognition from art using sequential co-attention. J Imaging
7(8):157
242. Halim Z, Waqar M, Tahir M (2020) A machine learning-based
investigation utilizing the in-text features for the identification
of dominant emotion in an email. Knowl Based Syst 208:106443
243. Atmaja BT, Sasou A (2022) Sentiment analysis and emotion
recognition from speech using universal speech representations.
Sensors 22(17):6369
244. Lee H-Y, Hu T-Y, Jing H, Chang Y-F, Tsao Y, Kao Y-C, Pao
T-L (2013) Ensemble of machine learning and acoustic segment
model techniques for speech emotion and autism spectrum
disorders recognition. In: INTERSPEECH, pp 215–219
245. Tahir M, Halim Z, Rahman AU, Waqas M, Tu S, Chen S, Han Z
(2022) Non-acted text and keystrokes database and learning
methods to recognize emotions. ACM Trans Multimed Comput
Commun Appl (TOMM) 18(2):1–24
246. Bharti D, Kukana P (2020) A hybrid machine learning model for
emotion recognition from speech signals. In: 2020 international
conference on smart electronics and communication (ICOSEC),
pp 491–496. IEEE
247. Zhang Y, Srivastava G (2022) Speech emotion recognition
method in educational scene based on machine learning. EAI
Endorsed Trans Scalable Inf Syst 9(5):9–9
248. Muthusamy H, Polat K, Yaacob S (2015) Improved emotion
recognition using gaussian mixture model and extreme learning
machine in speech and glottal signals. Math Probl Eng. https://
doi.org/10.1155/2015/394083
249. Guo L, Wang L, Dang J, Liu Z, Guan H (2019) Exploration of
complementary features for speech emotion recognition based
on kernel extreme learning machine. IEEE Access
7:75798–75809
250. Liu Z-T, Wu M, Cao W-H, Mao J-W, Xu J-P, Tan G-Z (2018)
Speech emotion recognition based on feature selection and
extreme learning machine decision tree. Neurocomputing
273:271–280
251. Rahman AU, Halim Z (2023) Identifying dominant emotional
state using handwriting and drawing samples by fusing features.
Appl Intell 53(3):2798–2814
252. Huang J, Liu B, Tao J (2021) Learning long-term temporal
contexts using skip rnn for continuous emotion recognition.
Virtual Real Intell Hardw 3(1):55–64
253. Baird A, Amiriparian S, Milling M, Schuller BW (2021) Emo-
tion recognition in public speaking scenarios utilising an lstm-
rnn approach with attention. In: 2021 IEEE spoken language
technology workshop (SLT), pp 397–402. IEEE
254. Mohammadrezaei P, Aminan M, Soltanian M, Borna K (2023)
Improving cnn-based solutions for emotion recognition using
evolutionary algorithms. Results Appl Math 18:100360
255. Bautista JL, Lee YK, Shin HS (2022) Speech emotion recog-
nition based on parallel cnn-attention networks with multi-fold
data augmentation. Electronics 11(23):3935
256. Singh YB, Goel S (2022) A systematic literature review of
speech emotion recognition approaches. Neurocomputing
492:245–63
257. Zhong Z (2023) Speech emotion recognition based on svm and
cnn using mfcc feature extraction. In: international conference
on statistics, data science, and computational intelligence
(CSDSCI 2022), vol. 12510, pp 445–452. SPIE
258. Saumya S, Singh JP (2022) Spam review detection using lstm
autoencoder: an unsupervised approach. Electron Commer Res
22(1):113–133
259. Nejad SJ, Ahmadi-Abkenari F, Bayat P (2020) Opinion spam
detection based on supervised sentiment analysis approach. In:
2020 10th international conference on computer and knowledge
engineering (ICCKE), pp 209–214. IEEE
260. Lunando E, Purwarianti A (2013) Indonesian social media
sentiment analysis with sarcasm detection. In: 2013 international
conference on advanced computer science and information
systems (ICACSIS), pp. 195–198. IEEE
261. Rao MV, Sindhu C (2021) Detection of sarcasm on amazon
product reviews using machine learning algorithms under
Neural Computing and Applications
123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
sentiment analysis. In: 2021 sixth international conference on
wireless communications, signal processing and networking
(WiSPNET), pp 196–199. IEEE
262. Adarsh M, Ravikumar P (2019) Sarcasm detection in text data to
bring out genuine sentiments for sentimental analysis. In: 2019
1st international conference on advances in information tech-
nology (ICAIT), pp 94–98. IEEE
263. Prasad AG, Sanjana S, Bhat SM, Harish B (2017) Sentiment
analysis for sarcasm detection on streaming short text data. In:
2017 2nd international conference on knowledge engineering
and applications (ICKEA), pp 1–5. IEEE
264. Singh N, Tomar D (2019) Comprehensive analysis of scope of
negation for sentiment analysis over social media. J Theor Appl
Inf Technol 97(6):1704–1719
265. Pro
¨llochs N, Feuerriegel S, Neumann D (2016) Negation scope
detection in sentiment analysis: decision support for news-dri-
ven trading. Decis Support Syst 88:67–75
266. Lazib L, Qin B, Zhao Y, Zhang W, Liu T (2020) A syntactic
path-based hybrid neural network for negation scope detection.
Front Comp Sci 14(1):84–94
267. Wu F, Huang Y (2016) Sentiment domain adaptation with
multiple sources. In: proceedings of the 54th annual meeting of
the association for computational linguistics (Vol. 1: Long
Papers), pp 301–310
268. Yuan Z, Wu S, Wu F, Liu J, Huang Y (2018) Domain attention
model for multi-domain sentiment classification. Knowl-Based
Syst 155:1–10
269. Amujo O, Ibeke E, Fuzi R, Ogara U, Iwendi C (2023) Sentiment
computation of uk-originated covid-19 vaccine tweets: a
chronological analysis and news effect. Sustainability
15(4):3212
270. Maeda K, Kurata H (2023) Automatic generation of sbml kinetic
models from natural language texts using gpt. Int J Mol Sci
24(8):7296
271. Sammani F, Mukherjee T, Deligiannis N (2022) Nlx-gpt: a
model for natural language explanations in vision and vision-
language tasks. In: proceedings of the IEEE/CVF conference on
computer vision and pattern recognition, pp 8322–8332
272. Trummer I (2022) Codexdb: synthesizing code for query pro-
cessing from natural language instructions using gpt-3 codex.
Procee VLDB Endow 15(11):2921–2928
Publisher’s Note Springer Nature remains neutral with regard to
jurisdictional claims in published maps and institutional affiliations.
Springer Nature or its licensor (e.g. a society or other partner) holds
exclusive rights to this article under a publishing agreement with the
author(s) or other rightsholder(s); author self-archiving of the
accepted manuscript version of this article is solely governed by the
terms of such publishing agreement and applicable law.
Neural Computing and Applications
123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
1.
2.
3.
4.
5.
6.
Terms and Conditions
Springer Nature journal content, brought to you courtesy of Springer Nature Customer Service Center GmbH (“Springer Nature”).
Springer Nature supports a reasonable amount of sharing of research papers by authors, subscribers and authorised users (“Users”), for small-
scale personal, non-commercial use provided that all copyright, trade and service marks and other proprietary notices are maintained. By
accessing, sharing, receiving or otherwise using the Springer Nature journal content you agree to these terms of use (“Terms”). For these
purposes, Springer Nature considers academic use (by researchers and students) to be non-commercial.
These Terms are supplementary and will apply in addition to any applicable website terms and conditions, a relevant site licence or a personal
subscription. These Terms will prevail over any conflict or ambiguity with regards to the relevant terms, a site licence or a personal subscription
(to the extent of the conflict or ambiguity only). For Creative Commons-licensed articles, the terms of the Creative Commons license used will
apply.
We collect and use personal data to provide access to the Springer Nature journal content. We may also use these personal data internally within
ResearchGate and Springer Nature and as agreed share it, in an anonymised way, for purposes of tracking, analysis and reporting. We will not
otherwise disclose your personal data outside the ResearchGate or the Springer Nature group of companies unless we have your permission as
detailed in the Privacy Policy.
While Users may use the Springer Nature journal content for small scale, personal non-commercial use, it is important to note that Users may
not:
use such content for the purpose of providing other users with access on a regular or large scale basis or as a means to circumvent access
control;
use such content where to do so would be considered a criminal or statutory offence in any jurisdiction, or gives rise to civil liability, or is
otherwise unlawful;
falsely or misleadingly imply or suggest endorsement, approval , sponsorship, or association unless explicitly agreed to by Springer Nature in
writing;
use bots or other automated methods to access the content or redirect messages
override any security feature or exclusionary protocol; or
share the content in order to create substitute for Springer Nature products or services or a systematic database of Springer Nature journal
content.
In line with the restriction against commercial use, Springer Nature does not permit the creation of a product or service that creates revenue,
royalties, rent or income from our content or its inclusion as part of a paid for service or for other commercial gain. Springer Nature journal
content cannot be used for inter-library loans and librarians may not upload Springer Nature journal content on a large scale into their, or any
other, institutional repository.
These terms of use are reviewed regularly and may be amended at any time. Springer Nature is not obligated to publish any information or
content on this website and may remove it or features or functionality at our sole discretion, at any time with or without notice. Springer Nature
may revoke this licence to you at any time and remove access to any copies of the Springer Nature journal content which have been saved.
To the fullest extent permitted by law, Springer Nature makes no warranties, representations or guarantees to Users, either express or implied
with respect to the Springer nature journal content and all parties disclaim and waive any implied warranties or warranties imposed by law,
including merchantability or fitness for any particular purpose.
Please note that these rights do not automatically extend to content, data or other material published by Springer Nature that may be licensed
from third parties.
If you would like to use or distribute our Springer Nature journal content to a wider audience or on a regular basis or in any other manner not
expressly permitted by these Terms, please contact Springer Nature at
onlineservice@springernature.com
A preview of this full-text is provided by Springer Nature.
Content available from Neural Computing and Applications
This content is subject to copyright. Terms and conditions apply.
