International Conference on Innovative Data Communication Technologies and Application (ICIDCA-2023)

IEEE Xplore Part Number: CFP23CR5-ART; ISBN: 979-8-3503-9720-8

A Review on Sentiment Analysis using Machine

Learning
Sumit Sindhu Dr. Sanjeev Kumar Dr. Amandeep Noliya
2023 International Conference on Innovative Data Communication Technologies and Application (ICIDCA) | 979-8-3503-9720-8/23/$31.00 ©2023 IEEE | DOI: 10.1109/ICIDCA56705.2023.10099665

Department of Computer Science and Department of Computer Science and Department of Computer Science and
Engineering Engineering Engineering
Guru Jambheshwar University of Guru Jambheshwar University of Guru Jambheshwar University of
Science and Technology Science and Technology Science and Technology
Hisar, Haryana Hisar, Haryana Hisar, Haryana
sumitsindhu786@gmail.com sanjukhambra@yahoo.co.in am.noliya@gmail.com

Abstract— S entiment Analysis [S A] is a method that is used tweets, reviews, and comments are positive or negative.
to detect the state of some information or text. It can be applied Sentiment analysis can be applied to a variety of tasks, such
to tasks such as text classification and summarization, and as market research surveys or automated customer service.
various methods can be used to extract sentiment from text, Today, NLP and sentiment analysis are both widely
including machine learning algorithms and rule-based systems. employed in a variety of sectors, from banking to healthcare.
Sentiment analysis is an important tool for businesses to By assisting businesses in better comprehending the wants
understand customer feedback and measure customer
and interests of their clients, they offer insightful data into
satisfaction and loyalty. It classifies the given text into three user behaviour [3]. They are also being used to automate
categories: positive, negative, or neutral. The statement "the
customer care responsibilit ies like pro mptly responding to
picture quality of TV is good" expresses a positive opinion
about a particular TV, while "the sound is not adequate" consumer inquiries or spotting possible issues before they
expresses a negative opinion. Recent challenges with sentiment arise.
analysis include capturing sarcasm, emotion in text and dealing Social med ia has become an invaluable source for
with long texts. The paper presents a literature survey of many sentiment analysis due to a large amount of user-generated
machine learning techniques as well as deep learning models
content available in a variety of formats. Social media posts
such as CNN, LS TM, etc. can include life perspectives, thoughts on many topics,
Keywords—Machine Learning, Deep Learning, Sentiment
current news, and internet issues. Due to a large number of
Analysis, Natural Language Processing, Artificial Intelligence opinions, sentiment analysis is used. Organizations can use
the information provided by social networks to improve the
I. INT RODUCT ION effectiveness of their products and services [4]. It may not be
necessary to conduct surveys, opinion polls, etc. to obtain
Sentiment analysis (SA) is a widely used technique in
information fro m user reviews. The quickest and easiest
natural language processing (NLP). It enables machines to approach for a business to learn what its customers think of
analyse human emotions and opinions by analysing text data.
its goods or services is to ask them on social networks and
Sentiment analysis can be used to determine the overall then evaluate the results to discover what they liked, what
sentiment expressed in a document as well as identify they did not like, and how it may be improved. Political
specific topics or aspects that are being discussed. It has parties now employ sentiment analysis to boost their appeal.
become an important tool for businesses to understand A popular example of the use of sentiment analysis in this
customer feedback to make better decisions. Sentiment
scenario is Narendra Modi's victory in the prime ministerial
analysis helps organisations measure customer satisfaction election. It is possible to determine how many voters are
and loyalty by analysing the sentiment expressed in online
positive, negative, or neutral towards Narendra Modi by
reviews and social media posts [1]. looking at the tweets about him. Therefore, sentiment
Natural language processing and sentiment analysis are analysis can benefit analysts. Sentiment analysis is generally
two related topics that have grown in popularity recently. performed using a combination of NLP and M L techniques.
The study of how computers read and process human In today’s world, sentiment analysis is proving a boon for
language is known as "natural language processing," or NLP. companies, manufacturers, politicians, etc. The value of
Sentiment analysis, on the other hand, is more narrowly sentiment analysis may be demonstrated by our need to
focused on comprehending the emotions portrayed in text. know how others react to situations and what they think.
Since the 1950s, when Alan Turing first outlined his concept Using neural networks makes this analysis more interesting.
of artificial intelligence (AI), NLP has existed. Since then, The things that matter are co mplexity and performance,
scientists have utilised AI to create algorithms that can which differ fro m approach to approach. So the best
comprehend and translate written or spoken speech [2]. With algorithms are needed that can do this task more efficiently
the introduction of machine learning (M L) models like deep while imparting high performance.
learning (DL) networks, wh ich enable machines to learn SA is performed on three levels [5], which are sentence
fro m data without being explicit ly programmed, this
level, document level, and aspect level. At the sentence level,
technology has evolved dramatically over time. the polarity of each sentence is obtained. Sentence-level
In the late 1990s, sentiment analysis first appeared as an classification is useful only when the sentence has sentiments
NLP application. It uses methods like NLP and machine related to a single object. At the document level, the main
learning to identify whether text-based discussions like focus is to obtain the sentiment of the entire document. It is

979-8-3503-9720-8/23/$31.00 ©2023 IEEE 138

Authorized licensed use limited to: Zhejiang University. Downloaded on March 05,2024 at 04:41:45 UTC from IEEE Xplore. Restrictions apply.
 International Conference on Innovative Data Communication Technologies and Application (ICIDCA-2023)
IEEE Xplore Part Number: CFP23CR5-ART; ISBN: 979-8-3503-9720-8

generally used when the document has sentiments related to content. This enables us to comprehend how people feel
a single object. Both sentence- and document-level about various subjects based on what they write or say in
classification work well when the whole sentence or speeches.
document focuses on a single type of sentiment. At aspect
level, fine-grained analysis is performed, i.e., first, the The ability of lexical analysis to accurately extract
significant insights from textual data sources without
sentiments of an entity are calculated, then the sentiment of
the object is obtained. For applications ranging from requiring any additional external resources beyond those
already present within existing datasets themselves has
customer service to marketing and medical, sentiment
allowed it to establish itself over t ime as an effective method
analysis is commonly used in voice of the customer
materials, including reviews of products and services, for performing sentiment classification tasks. This makes it a
highly cost-effective alternative to other traditional methods
surveys, the Web and social networks, and healthcare
resources. currently on the market [8].

Sentiment analysis uses a technique called "sentiment B. Sentiment Analysis using Machine Learning
scoring" to measure the sentiment of a given text. Based on Arthur Samuel introduced machine learning in 1959.
the emotion a sentence conveys, it is given a numerical Machine learning is a statistical algorithm that can learn
score. A score of 0 indicates a neutral feeling, and the scale automatically without any help fro m a human. With every
can go from negative to positive. Many methods, including experience, the system becomes smarter and smarter without
rule-based algorithms, machine learning, and natural the interference of humans. It is a self-learn ing algorithm that
language processing, can be used to imp lement sentiment learns from experience [9]. The model is supplied with the
scoring. data, and the model learns the pattern and structure of the
data and provides the output from its learning-based
A text sample is classified and assigned a sentiment label experience.
(positive, negative, or neutral) as part of sentiment analysis.
A classifier is trained for classification using a labelled Machine learning has completely changed the field of
collection of documents. The sentiment of te xts that haven't natural language processing, especially when it comes to
been seen is then predicted using the classifier. Word counts, issues with sentiment analysis. Recent developments have
n-grams, part-of-speech tags, and other linguistic variables made it possible to use computer vision and deep neural
are frequently used by the classifier to produce its network-based solutions that were previously too challenging
predictions. to address using traditional programming paradigms alone.
Studies generally consider positive and negative labels C. Sentiment Analysis using Deep Learning
for the classification of tweets, ignoring the neutral label, Deep learn ing is a sub-section of machine learning with
citing the reason that neutral is less likely to learn sentiment an algorithm that tries to imitate the human brain's structure
polarity fro m neutral texts compared to positive or negative and function. DL learns fro m experience but requires a large
texts. However, it is also necessary to consider the "neutral" information base as input. A network model with neurons
label, as some tweets or texts may not convey a positive or includes several hidden layers between the model’s input and
negative sentiment [6]. Additionally, considering only output. The beauty of DL is that it automatically learns about
positive and negative tweets will not provide an accurate features and feature representation. DL is best suited for very
classification of neutral tweets. For example, the text "This large and complex data.
TV show is unpredictable" shows no emotion. It cannot be
said without context that the show is good or bad. Depending Certain neural network algorithms bas ed on the notion of
on the label, future predictions will not be accurate when this dense vector representations have previously demonstrated
text is classified into two classes only. The neutral class, state-of-the-art performances in a variety of NLP -related
which represents the absence of emotion, should not be applications. Deep learning NNs originally showed superior
regarded as a state that exists between positive and negative. performance in the computer vision and pattern recognition
Instead, it should be seen as a separate class. subfields. As a result of this shift, several deep learning
algorithms are being used to manage complicated NLP
II. SENT IMENT A NALYSIS TECHNIQUES assertions such as sentiment analysis [10].

III. LIT ERAT URE REVIEW

Lal Khan et al. [11] proposed a SA model for classifying
Ro man-Urdu and English texts. The paper proposed a
framework by applying a hybrid approach using
convolutional neural networks (CNN) and Long short-term
memory (LSTM ). For feature ext raction, the CNN model
was used, and the LSTM model was used for long-term
dependency preservation. The output provided by LSTM is
fed to ML classifiers. By combin ing CNN and LSTM ,
A. Sentiment Analysis using Lexical Analysis
the accuracy and efficiency of sentiment analysis are
Lexical analysis is a technique that includes dissecting increased.
natural language texts into their individual words and phrases
to identify the meaning. Lexical analysis divides lengthy Gen Li et al. [12] proposed a model to learn the
texts into manageable segments that are more easily knowledge of sentiment in Chinese text. A hybrid task
understood by both computers and people [7]. Lexical learning method is designed so that valuable emotional
analysis looks for linguistic patterns that can be used to sentiments are learned and, hence, sentiment tendencies are
categorise materials based on their sentiment or emotional predicted. Experiments showed that the model performs

979-8-3503-9720-8/23/$31.00 ©2023 IEEE 139

Authorized licensed use limited to: Zhejiang University. Downloaded on March 05,2024 at 04:41:45 UTC from IEEE Xplore. Restrictions apply.
 International Conference on Innovative Data Communication Technologies and Application (ICIDCA-2023)
IEEE Xplore Part Number: CFP23CR5-ART; ISBN: 979-8-3503-9720-8

better compared to other models and has better generalisation for two classes. The reannotated tweets provided better
ability than other methods. results than the default sentiments, with an accuracy of 99%.
Naqvi et al. [13] proposed a framework by discovering H. Sadr et al. [21] proposed a model that finalised mixing
deep learning techniques in addition to different up the features derived from heterogeneous Artificial neural
representations of word vectors. For SA, the performance of network (ANN) by using multiple view classifiers to
some popularly known DL methods is calculated. An improve the combined performance of SA at the document
application of stacked layers is performed with a single layer level while also taking into account their co-relationship. The
of convolution; different filters are used in CNN. To more proposed network uses intermediate features derived from
accurately identify sentiment in longer texts, CNNs can be CNN and RNN so that classification can be performed.
used to capture long-term dependencies in text. The roles
A. El-Affendi et al. [22] proposed a deep learning dash-
performed by pre-trained models and unsupervised self-
trained models were also examined. Other deep learning based multilevel parallel attention neural model (MPAN). It
uses the concept of a simp le positioning binary embedding
models were outperformed by one with an accuracy of
77.9%. scheme, known as PBES. Therefore, the contextualized
embedding of character, word, and sentence levels is
Wang et al. [14] proposed a model in wh ich a computed simultaneously. Using the public Internet movie
differentiation in the sentiment informat ion of various words database, the movie dataset model achieved an accuracy of
was only achieved by using sentiment representation 96.13%.
methods. The paper uses the concept of sentiment notions for
solving problems. After comparing the word embedding Hasan M. et al. [23] proposed research on the normal
classification method for the measurement of the possibility
technique based on the sentiment idea with multip le
embeddings, the word embedding method's validity based on of allocating the data for every sentiment class. They
developed and computed the supervised learning method for
the sentiment concept is confirmed.
the automatic classification of sentiments in textual stream
Meylan Wongkar et al. [15] proposed a Naive Bayes data. Additionally, they planned an online technique to
(NB)-based SA model. The NB algorithm is used to classify measure public sentiment and identify sentiment rupture
social groups or levels of emotion. A comparison was tested emotions in lively streams.
between NB, support vector machines (SVM), and K-
Kotsilieris, T. et al. [24] developed a comparative survey
Nearest Neighbors (KNN) that produced an accuracy value
of 75% for NB, 64% for SVM, and 73% for KNN. NB on research for the calculation of a specific kind of
nervousness disorder and research for the calculation of
algorithm is used to classify social groups or levels of
emotion. suicide ability using ML methods. In addition, they
examined and compared the ML methods of treating anxiety
U. Sehar et al. [16] proposed a framework that integrates disorder.
elements such as graphic, acoustic, and text responses so that
context-conscious sentiments are detected. To uncover useful Yang et al. [25] proposed a new sentiment analysis
model, SLCA BG, based on a sentiment lexicon and a
patterns, a unique dataset that contains 1372 exp ressions is
used. Both decision and feature dash level fusion methods combination of a convolutional neural network with a bi-
directional gated recurrent unit based on attention (BiGRU).
are used to improve the sentiment clarity prediction. The
polarity detection ability was improved from 84% to 95%. The SLCABG model has the disadvantages of sentiment
lexicons, but it aids in overcoming the drawbacks of the
Tam et al. [17] proposed an integrating structure of bi- existing sentiment analysis model of product evaluations.
LSTM and CNN with word embeddings. Co mpared to
Anu J. Nair et al. [26] proposed a sentiment analysis
conventional approaches, BiLSTM is more reliable at
model by using the Twitter dataset for COVID-19. Sentiment
predicting sentiment. This is due to its improved ability to
grasp linguistic subtleties and comprehend sentence context. analysis classifies tweets as positive, negative, or neutral.
The author used logistic regression (LR), Vadar sentiment
In this model, Glo Ve and Word2Vec word embeddings were
used. The model outperformed other models with 91.13% analysis, and BERT sentiment analysis to perform sentiment
analysis. These techniques are more sensitive to social media
accuracy.
situations.
Al A min et al. [18] proposed a model for Bengali NLP.
Tripathy et al. [27] classified a social media dataset with
The author modified the VA DER model to support the
the help of supervised machine learning methods into
identification of Bengali sentiment polarity. English lexicon
polarity is included to improve the polarity of the Bengali sentiment categories. This paper classified film reviews by
using a supervised machine learning algorithm and later, by
lexicon. They used stemming, a list of Bengali boosting
words, to achieve a superior outcome. using an n-gram approach, applied these to the IMDB
dataset. Authors concluded that as the value of n increases
Davcheva E et al. [19]: The author's research goal was to classification accuracy decreases and combination of TF-IDE
understand better situations and changes in mental state. The & count vectorizer technique helps in improving accuracy.
outcomes of the research showed that sentiment scoring can
be implemented positively or negatively based on the
situation.
Gaye, B. et al. [20] proposed a stacked ensemble model.
The model used three LSTMs stacked one after another. The
output of the stacked model is fed to an LR classifier. The
authors reannotated the default sentiments by using Text Blob

979-8-3503-9720-8/23/$31.00 ©2023 IEEE 140

Authorized licensed use limited to: Zhejiang University. Downloaded on March 05,2024 at 04:41:45 UTC from IEEE Xplore. Restrictions apply.
 International Conference on Innovative Data Communication Technologies and Application (ICIDCA-2023)
IEEE Xplore Part Number: CFP23CR5-ART; ISBN: 979-8-3503-9720-8

TABLE I. ANALYSIS OF VARIOUS SENTIMENT ANALYSIS TECHNIQUES REFERENCES

Classifier With Conclusion [1] Sharma, Gajendra, and Li Baoku. "Customer satisfaction in Web 2.0
S. No Author
Highest Accuracy and information technology development." Information technology &
Lal Khan Combined CNN and LST M people (2013).
1 CNN-LST M
et. al. and Achieved 84% accuracy
[2] González García, Cristian, Edward Rolando Núñez Valdéz, Vicente
Chinese language García Díaz, Begoña Cristina Pelayo García-Bustelo, and Juan
Gen Li et.
2 SINM classification and Achieved Manuel Cueva Lovelle. "A review of artificial intelligence in the
al
89% accuracy internet of things." International Journal Of Interactive Multimedia
Worked with different And Artificial Intelligence, 5 (2019).
U. Naqvi
3 Bi-LST M-AT T number of filters in CNN.
et. al. [3] He, Wu, Shenghua Zha, and Ling Li. "Social media competitive
Achieved 77% accuracy
analysis and text mining: A case study in the pizza industry."
Uses sentiment notion and
Y. Wang et. International journal of information management 33, no. 3 (2013):
4 Bi-GRU word embedding. Achieved
al. 464-472.
91.3% accuracy
Meylan Used Naïve bayes algorithm [4] Parveen, Farzana, Noor Ismawati Jaafar, and Sulaiman Ainin. "Social
5 Wongkar Naïve Bayes to achieve 75.58% accuracy media usage and organizational performance: Reflections of
et. al. Malaysian social media managers." T elematics and informatics 32,
Integrated multiple elements no. 1 (2015): 67-78.
U. Sehar et. to detect content conscious [5] Godara, N. and Kumar, S., 2020. T witter sentiment classification
6 MultiModal using ML techniques. Waffen-Und Kostumkunde Journal, 11(8),
al. emotions. Achieved 95.35%
accuracy pp.10-20.
Used multiple word [6] Hamed, A.R., Qiu, R. and Li, D., 2016. The importance of neutral
S. T am et.
7 CNN-BiLST M embedding techniques. class in sentiment analysis of Arabic tweets. Int. J. Comput. Sci.
al.
Achieved 91.13% accuracy Inform. Technol, 8, pp.17-31.
Using vader on Bengali [7] ] Hanks, Patrick. Lexical analysis: Norms and exploitations. Mit
Al Amin et.
8 B-VADER language, achieved 70% Press, 2013.
al.
accuracy
[8] MacAvaney, Sean, Hao-Ren Yao, Eugene Yang, Katina Russell,
Sentiment scoring of mental Nazli Goharian, and Ophir Frieder. "Hate speech detection:
Davcheva
9 Linear Regression situations and for better Challenges and solutions." PloS one 14, no. 8 (2019): e0221152.
E et. al.
understating of situations
Achieved 99% accuracy [9] ZAKARYAN, Arman. "Application of artificial intelligence (neural
Gaye,B. et. T hree Stacked networks) in education." Main Issues Of Pedagogy And Psychology
10 using stacked LST M
al. LST M 19, no. 1 (2021): 78-87.
technique
multi-view deep network [10] Arulmurugan, R., K. R. Sabarmathi, and H. J. C. C. Anandakumar.
H. Sadr et. Multi View Deep "Classification of sentence level sentiment analysis using cloud
11 outperforms single-view
al. Network machine learning techniques." Cluster Computing 22, no. 1 (2019):
deep neural networks
M. A. El- Using a custom deep learning 1199-1209.
Attention Based
12 Affendi et. model, produces an accuracy [11] Khan, L., Amjad, A., Afaq, K.M. and Chang, H.T., 2022. Deep
Neural Network sentiment analysis using CNN-LST M architecture of English and
al of 96.13%
Hasan M. Supervised Detecting emotions in live Roman Urdu text shared in social media. Applied Sciences, 12(5),
13 p.2694.
et. al. Learning text streams
Multiple Machine Hybrid machine learning and [12] Li, G., Zheng, Q., Zhang, L., Guo, S. and Niu, L., 2020, November.
Kotsilieris,
14 learning SVM produces best results Sentiment infomation based model for Chinese text sentiment
T . et. al.
techniques analysis. In 2020 IEEE 3rd International Conference on Automation,
Combining lexical analysis Electronics and Electrical Engineering (AUTEEE) (pp. 366 -371).
L. Yang et.
15 SLCABG and deep learning produces IEEE.
al.
better results [13] Naqvi, U., Majid, A. and Abbas, S.A., 2021. UT SA: Urdu text
BERT and Vader Useful for government and sentiment analysis using deep learning methods. IEEE Access, 9,
Anu J Nair
16 AND Logistic Health Officials pp.114085-114094.
et. al.
Regression [14] Wang, Y., Huang, G., Li, J., Li, H., Zhou, Y. and Jiang, H., 2021.
value of ‘n’ in n-gram Refined global word embeddings based on sentiment concept for
T ripathy et. n-gram based
17 increases the classification sentiment analysis. IEEE Access, 9, pp.37075-37085.
al. machine learning
accuracy
[15] Wongkar, M. and Angdresey, A., 2019, October. Sentiment analysis
using Naive Bayes Algorithm of the data crawler: T witter. In 2019
IV. CONCLUSION Fourth International Conference on Informatics and Computing
Multiple machine learn ing and deep learning techniques (ICIC) (pp. 1-5). IEEE.
were evaluated in this paper for their performance on [16] Sehar, U., Kanwal, S., Dashtipur, K., Mir, U., Abbasi, U. and Khan,
sentiment analysis. Mostly, data is gathered from social F., 2021. Urdu Sentiment Analysis via Multimodal Data Mining
Based on Deep Learning Algorithms. IEEE Access, 9, pp.153072 -
media sites such as Twitter, as nowadays most people voice 153082.
their opinions on social med ia. Deep learning has become an [17] Tam, S., Said, R.B. and T anriöver, Ö.Ö., 2021. A ConvBiLST M deep
increasingly popular choice for sentiment analysis due to the learning model-based approach for T witter sentiment classification.
advancements in hardware that have made it more IEEE Access, 9, pp.41283-41293.
accessible. Deep learning algorith ms are useful for sentiment [18] Amin, A., Hossain, I., Akther, A. and Alam, K.M., 2019, February.
analysis tasks because they can learn from large datasets and Bengali vader: A sentiment analysis approach using modified vader.
automatically identify complicated correlations between data In 2019 International Conference on Electrical, Computer and
Communication Engineering (ECCE) (pp. 1-6). IEEE.
points. As a result, deep learning models may recognise
textual patterns with greater accuracy than is practicable or [19] Davcheva, E., Adam, M. and Benlian, A., 2019. User dynamics in
mental health forums–a sentiment analysis perspective.
even viable using more conventional machine learning
[20] Gaye, B., Zhang, D. and Wulamu, A., 2021. A T weet sentiment
techniques. classification approach using a hybrid stacked ensemble technique.
Information, 12(9), p.374.
[21] Sadr, H., Pedram, M.M. and Teshnehlab, M., 2020. Multi-view deep
network: a deep model based on learning features from heterogeneous

979-8-3503-9720-8/23/$31.00 ©2023 IEEE 141

Authorized licensed use limited to: Zhejiang University. Downloaded on March 05,2024 at 04:41:45 UTC from IEEE Xplore. Restrictions apply.
 International Conference on Innovative Data Communication Technologies and Application (ICIDCA-2023)
IEEE Xplore Part Number: CFP23CR5-ART; ISBN: 979-8-3503-9720-8

neural networks for sentiment analysis. IEEE access, 8, pp.86984- [25] L. Yang et al, “Sentiment Analysis for E-Commerce Product Reviews
86997.] in Chinese Based on Sentiment Lexicon and Deep Learning,” IEEE
[22] El-Affendi, M.A., Alrajhi, K. and Hussain, A., 2021. A novel deep Access vol 20, pp. 23522-23530, Feb. 2020.
learning-based multilevel parallel attention neural (MPAN) model for [26] Anu J Nair et al, “Comparative study of T witter Sentiment On
multidomain arabic sentiment analysis. IEEE Access, 9, pp.7508 - COVID - 19 T weets,” Proceedings of the Fifth International
7518. Conference on Computing Methodologies and Communication
[23] Hasan, M., Rundensteiner, E. and Agu, E., 2019. Automatic emotion (ICCMC 2021),Apr. 2021. Pp 1773-1778.
detection in text streams by analyzing twitter data. International [27] Tripathy, A. Agarwal, A.& Rath, S.K. “ Classification of Sentiment
Journal of Data Science and Analytics, 7(1), pp.35-51. Reviews Using n-gram Machine Learning Approach,” Expert System
[24] Theodore, K., Ioannis, L., Panagiotis, P. and Emmanouil, P., 2018. with Applications,57, pp 117-126,Sept. 2016.
Reviewing machine learning techniques for predicting anxiety
disorders.

979-8-3503-9720-8/23/$31.00 ©2023 IEEE 142

Authorized licensed use limited to: Zhejiang University. Downloaded on March 05,2024 at 04:41:45 UTC from IEEE Xplore. Restrictions apply.