A Comprehensive Review of Sentiment Analysis

Using Machine Learning Techniques

Yashaswini C R Ananda Babu J
Department of Information Science & Engg Department of Information Science & Engg
Malnad College of Engineering Malnad College of Engineering
Hassan,India Hassan,India

Abstract—Sentiment analysis, a subfield of natural language content recommendations [9]. Finance and healthcare sectors
processing (NLP), is crucial for understanding opinions in text use sentiment analysis for investor sentiment assessment and
data across domains like e-commerce, social media, and finance. patient feedback analysis [10].
Traditional lexicon-based methods often struggle with contextual
ambiguity, necessitating the use of machine learning (ML) tech- Despite progress, challenges such as sarcasm detection,
niques. This review explores supervised, unsupervised, and deep multilingual text handling, class imbalance, and real-time
learning approaches, including Naı̈ve Bayes, Support Vector Ma- scalability persist [11]. Hybrid approaches combining lexicon-
chines, Long Short-Term Memory networks, and Transformer- based and ML techniques improve accuracy and generalizabil-
based models. Challenges such as sarcasm detection, multilingual
ity [12]. Ensemble learning methods like Random Forest and
processing, and bias mitigation are discussed. Future directions
include multimodal sentiment analysis, domain adaptation, and Soft Voting Classifiers enhance classification by integrating
explainable AI to enhance sentiment classification accuracy and multiple models [13].
reliability. A comprehensive review of sentiment analysis using ML,
Keywords— Sentiment Analysis, Machine Learning, Natural covering methodologies, applications, challenges, and future
Language Processing, Deep Learning, Sentiment Classification.
research directions. It explores machine learning models, real-
I. I NTRODUCTION world applications, and emerging trends such as deep learning
integration, domain adaptation, and ethical considerations [14].
The exponential growth of digital communication and user- By presenting an in-depth review, this work serves as a
generated content has led to a rising demand for sentiment valuable resource for researchers and practitioners seeking to
analysis, an NLP task that extracts subjective opinions from advance sentiment classification techniques.
text. Sentiment analysis is vital in e-commerce, social media,
finance, healthcare, and entertainment, aiding businesses and II. A DVANCES IN S ENTIMENT A NALYSIS : M ACHINE
policymakers in decision-making [1]. L EARNING METHODS AND A PPLICATIONS
Traditional sentiment analysis methods, such as lexicon-
based approaches using Sent WordNet, provided an initial The field of sentiment analysis has evolved significantly
framework but struggled with contextual ambiguity, sarcasm, with the adoption of machine learning (ML) techniques, which
and evolving language use [2]. To overcome these limitations, have enhanced the ability to automatically classify opinions
machine learning (ML) techniques have been widely adopted. and emotions from textual data. Traditional lexicon-based ap-
Supervised learning models like Naı̈ve Bayes, SVM, Decision proaches, which rely on predefined sentiment word lists, have
Trees, and Logistic Regression perform well but require large proven to be limited in their ability to handle context, sarcasm,
labeled datasets [3]. Unsupervised approaches, such as clus- domain-specific expressions, and evolving linguistic trends.
tering and topic modeling, extract sentiment patterns without Machine learning, on the other hand, allows for adaptive
labeled data [4]. learning from data, improving the accuracy and robustness
Deep learning models, including RNNs, LSTMs, CNNs, and of sentiment classification [1]. Machine learning-based sen-
Transformer-based architectures like BERT and GPT-3, have timent analysis can be categorized into supervised learning,
significantly improved sentiment classification by capturing unsupervised learning, deep learning-based approaches, hybrid
contextual dependencies [5]. Sentiment analysis applications models, and ensemble techniques. This section explores these
span customer feedback analysis, brand monitoring, political methods, their advantages, challenges, and their applications
opinion mining, and market trend prediction [6]. in real-world sentiment classification tasks.
In e-commerce, analyzing product reviews on Amazon and
A. Supervised Learning-Based Sentiment Analysis
Flipkart helps businesses understand consumer preferences [7].
Social media sentiment analysis on Twitter and Facebook Supervised learning methods require labeled datasets, where
enables organizations to track public opinion and mitigate each text sample is tagged with a corresponding sentiment
reputational risks [8]. In entertainment, sentiment analysis of label (e.g., positive, negative, or neutral). These models are
movie and music reviews supports box office predictions and trained on labeled text data and subsequently used to classify
new, unseen data. The most widely used supervised learning typically lower than supervised models due to their reliance
algorithms for sentiment analysis include: on unstructured data representations [10].
• Naı̈ve Bayes (NB):
C. Deep Learning-Based Sentiment Analysis
The probabilistic classifier is based on Bayes’ theorem
and assumes independence between features. Despite its Deep learning models have revolutionized sentiment analy-
simplifying assumption, NB is computationally efficient sis by enabling context-aware and hierarchical feature extrac-
and performs well for text classification tasks, making it tion. Unlike traditional ML models, which rely on manually
a popular choice for sentiment analysis [2]. engineered features, deep learning models automatically learn
• Support Vector Machines (SVM):
sentiment representations from raw text data. Key deep learn-
SVM constructs an optimal hyperplane to separate sen- ing architectures for sentiment analysis include:
timent classes in high-dimensional space. It is highly • Recurrent Neural Networks (RNNs):
effective for binary sentiment classification and works Designed to process sequential data, RNNs are capable
well with term frequency-inverse document frequency of learning temporal dependencies in sentiment-laden
(TF-IDF) feature representations [3]. text. However, they suffer from the vanishing gradient
• Decision Trees (DT) and Random Forest (RF): problem, limiting their effectiveness for long sequences
These tree-based classifiers create rule-based splits in [11].
data, with RF combining multiple trees to enhance pre- • Long Short-Term Memory Networks (LSTMs):
diction stability and reduce overfitting. These methods are A specialized form of RNNs that introduces gating mech-
highly interpretable, making them useful for explainable anisms to retain long-range dependencies. LSTMs have
sentiment classification [4]. proven effective for document-level sentiment classifica-
• Logistic Regression (LR): tion [12].
A statistical model that predicts the probability of senti- • Convolutional Neural Networks (CNNs):
ment classes based on input features. It is often used as a While traditionally used for image processing, CNNs
baseline model for sentiment analysis due to its simplicity have also been adapted for sentiment analysis, where they
and interpretability [5]. extract n-gram features from text to detect sentiment cues
Supervised learning methods are widely used in industry due [13].
• Transformer-Based Models (BERT, GPT-3):
to their high accuracy and ability to learn domain-specific
sentiment patterns. However, they require large, annotated These state-of-the-art NLP models leverage self-attention
datasets, which can be expensive and time-consuming to create mechanisms to capture long-range dependencies in text,
[6]. achieving superior performance in context-aware senti-
ment analysis [14].
B. Unsupervised Learning in Sentiment Analysis Despite their high accuracy, deep learning models require
Unlike supervised learning, unsupervised learning tech- large-scale training data and high computational resources,
niques do not require labeled datasets. These methods rely on posing challenges for real-time sentiment analysis [15].
clustering and statistical learning to group similar sentiment D. Hybrid Approaches and Ensemble Models
expressions and extract hidden patterns from unstructured text
Hybrid approaches combine lexicon-based methods with
data. Some widely used unsupervised approaches in sentiment
machine learning models to enhance sentiment classification
analysis include:
accuracy. These methods leverage domain-specific sentiment
• K-Means Clustering:
dictionaries alongside ML algorithms to improve contextual
This algorithm partitions text samples into K clusters understanding [16]. Additionally, ensemble learning tech-
based on their feature similarity. It is often used for niques, such as Soft Voting and Stacking Classifiers, integrate
exploratory sentiment analysis when labeled data is un- multiple models to enhance sentiment prediction robustness.
available [7]. These approaches have been particularly effective in cross-
• Latent Dirichlet Allocation (LDA):
domain sentiment classification [17].
A probabilistic topic modeling technique that extracts
latent topics from a corpus. LDA can identify sentiment- E. Dataset
related themes in large text collections, making it useful A sentiment analysis dataset is a collection of labeled text
for aspect-based sentiment analysis [8]. data used to train and evaluate machine learning models. It
• Word Embeddings (Word2Vec, GloVe): includes text from sources like social media, product reviews,
These models learn vector representations of words based and news articles, categorized as positive, negative, or neutral.
on their contextual usage in text corpora. Word2Vec The dataset size varies, with some containing millions of
and GloVe capture semantic relationships, allowing for entries, while others are more specific. Metadata such as
improved sentiment classification in sparse datasets [9]. date, time, and location may be included to provide deeper
Although unsupervised approaches can be useful for domain sentiment insights. These datasets serve as essential resources
adaptation and low-resource settings, their performance is for natural language processing and sentiment classification.
They help train machine learning algorithms to recognize and F. Performance analysis of sentimental analysis
interpret emotions in text. By using labeled datasets, sentiment All four classifiers, including the voting classifier, had
classification accuracy improves significantly. Researchers use their accuracy and f1-measure compared. The compar-
them to evaluate and compare different models. High-quality ison’s findings are displayed in Table I, and the SVM
datasets enhance text analysis and NLP advancements. They classifier performs better than the other classifier. To
contribute to developing accurate predictive models for indus- create numerous models, the bagging classifier is
tries like marketing, finance, and customer service. Sentiment
analysis helps businesses understand consumer opinions and TABLE I
trends. It also plays a role in monitoring public perception C OMPARATIVE A NALYSIS OF FOUR CLASSIFIERS
and brand reputation. With the growth of machine learning,
sentiment analysis is becoming more refined. The availability Models Accuracy Score Precision Score Recall Score F1-Score
of structured datasets is key to improving AI-driven text
analysis. Ultimately, sentiment analysis datasets are crucial for Multinomial Naı̈ve Bayes 92.52 95.56 96.9 90.56
advancing NLP applications.
Support Vector Machine 96.46 98.39 93.57 90.39

Random Forest 96.73 96.63 97.53 97.41

Decision Tree 96.23 95.89 95.91 96.49

Voting Classifier 96.47 96.88 96.91 90.73

Fig. 1. Dataset

• Precision (p): Precision score is a measure of how accu-

rately the model identifies positive or negative sentiment
in a text. The ratio of true positive (TP) to the total of
both true positive and false positive (FP) predictions is
used to determine precision score.
True Positive
Precision =
True Positive + False Positive
• Accuracy: The number of correctly identified examples
(true positive and true negative) to the total number of
instances in the dataset is used to calculate accuracy.
(TP + TN)
Accuracy =
(TP + FP + TN + FN)
• F1-Score: This is known as the mean which is a
combination measure of p-r.
Fig. 2. Dataset
2×p×r
F1-Score =
p+r
C ONCLUSION
• Recall(r): Recall score is a measure of how well a The implementation of bagging and voting classifier
model correctly identifies positive or negative sentiment techniques led to a significant improvement in the per-
in a text. Recall is independent of the number of negative formance of four sentiment analysis algorithms applied
sample classifications. Further, if the model classifies all to tweets. The voting classifier played a crucial role
positive samples as positive, then Recall will be 1. in combining predictions from multiple models, thereby
True Positive enhancing the overall accuracy of sentiment classifica-
Recall = tion. Simultaneously, bagging helped reduce variance
True Positive + False Negative
and prevented overfitting, ensuring that the models per- [8] “Sentiment Analysis of Machine Learning Algorithms: A
formed consistently across different datasets. The results Transformer-Based Approach”, Sudabathula, Vijay Sai Kumar
and Varma, Nadimpalli Madana Kailash and Mattaparty, Sri Harsh
indicated that using both bagging and voting classifiers and Naik, Banoth Krishna Mohan and Ahmeduddin, Syed and
increased the model’s accuracy from 96.27% to 97.21%, Aryan, Adla, 2024 4th International Conference on Advancement
demonstrating the effectiveness of ensemble learning in Electronics & Communication Engineering (AECE) pp.909–
914, 2024, IEEE
techniques. [9] “Sentiment Analysis Using Machine Learning and Deep Learning
By leveraging these approaches, sentiment analysis mod- Models”, Vu, Hoang-Dieu and Pham, Quang-Tu and Solanki,
els become more robust and reliable in analyzing large Vijender Kumar and Hoang, Trong-Minh and Tran, Due-Tan,
2024 IEEE International Conference on Machine Learning and
datasets, including social media posts, product reviews, Applied Network Technologies (ICMLANT), pp.68–73, 2024,
and customer feedback. The integration of ensemble IEEE
learning methods allows for better generalization, reduc- [10] “Sentiment Analysis using NLP Libraries and Machine Learning”,
Maria, Meherun Nessa and Kabir, Tahira and Akter, Sakiba and
ing errors that arise from individual models. Additionally, Khan, Riasat, 2024 8th International Conference on I-SMAC (IoT
these techniques can be extended beyond sentiment analy- in Social, Mobile, Analytics and Cloud)(I-SMAC), pp.911–916,
sis to other domains within machine learning and natural 2024, IEEE
[11] “Fine grained sentiment analysis using machine learning and deep
language processing (NLP), such as text classification, learning”, Chauhan, Rahul and Gusain, Aman and Kumar, Prabhat
spam detection, and recommendation systems. and Bhatt, Chandradeep and Uniyal, Ishita, 2023 International
The scalability and adaptability of bagging and voting Conference on Sustainable Emerging Innovations in Engineering
and Technology (ICSEIET), pp.423–427, 2023, IEEE
classifiers make them suitable for real-time applica- [12] “A review on sentiment analysis using machine learning”, Sindhu,
tions, ensuring high accuracy even with vast amounts Sumit and Kumar, Sanjeev and Noliya, Amandeep, 2023 Interna-
of unstructured text data. Businesses, researchers, and tional Conference on Innovative Data Communication Technolo-
gies and Application (ICIDCA), pp.138–142, 2023, IEEE
data analysts can benefit from these methods to extract [13] “Twitter sentiment analysis using supervised machine learning”,
meaningful insights and make data-driven decisions. In Yadav, Nikhil and Kudale, Omkar and Rao, Aditi and Gupta,
conclusion, ensemble learning techniques provide a pow- Srishti and Shitole, Ajitkumar, Intelligent data communication
technologies and internet of things: Proceedings of ICICI 2020,
erful framework for improving model performance and pp.631–642, 2021, Springer
accuracy, making machine learning applications more [14] “Sentiment analysis in czech social media using supervised ma-
effective, efficient, and reliable in various real-world chine learning, Habernal, Ivan and Ptáček, Tomáš and Stein-
berger, Josef, Proceedings of the 4th workshop on computational
scenarios. approaches to subjectivity, sentiment and social media analysis,
pp.65–74, 2013, Springer
R EFERENCES [15] “On multi-tier sentiment analysis using supervised machine learn-
[1] “Enhancing Accuracy in Social Media Sentiment Analysis ing, Moh, Melody and Gajjala, Abhiteja and Gangireddy, Siva
through Comparative Studies using Machine Learning Tech- Charan Reddy and Moh, Teng-Sheng, 2015 IEEE/WIC/ACM In-
niques”, Yogi, Kottala Sri and Gowda, V Dankan and Sindhu, D ternational Conference on Web Intelligence and Intelligent Agent
and Soni, Hariprasad and Mukherjee, Saptarshi and Madhu, GC, Technology (WI-IAT), pp.341–344, 2015, IEEE
2024 International Conference on Knowledge Engineering and
Communication Systems (ICKECS), vol.1, pp.1–6, 2024, IEEE
[2] “Logistic Regression based Sentiment Analysis System: Rectify”,
Singh, Harsh Pratap and Singh, Nagendra and Mishra, Anuprita
and Sen, Santosh Kumar and Swarnkar, Mamta and Pandey,
Deepak, 2024 IEEE International Conference on Big Data &
Machine Learning (ICBDML), pp.186–191, 2024, IEEE
[3] “Multilingual Sentiment Analysis for YouTube Educational
Videos Using NLP And Machine Learning Approaches”, Yo-
geshkannah, K and Prem, S and Pooranachandiran, G and Viyash,
V and others, 2024 Third International Conference on Smart Tech-
nologies and Systems for Next Generation Computing (ICSTSN),
pp.1–6, 2024, IEEE
[4] “Sentiment Analysis: A Machine Learning Perspective”, Varma,
Nadimpalli Madana Kailash and Mattaparty, Sri Harsh and Ismail,
Shifa and Thaduri, Joel and Arora, Gagan Deep and others,2024
First International Conference on Electronics, Communication and
Signal Processing (ICECSP), pp.1–6, 2024, IEEE
[5] “Sentiment Analysis: Analyzing Flipkart Product Reviews using
NLP and Machine Learning”, Sharma, Hemant and Kakran,
Vandana, 2024 International Conference on Computing, Sciences
and Communications (ICCSC), pp.1–7, 2024, IEEE
[6] “Sentiment Analysis and Classification of Product Reviews: A
Comprehensive Study Using NLP and Machine Learning Tech-
niques”, Godia, Adarsh and Tiwari, LK, 2024 10th International
Conference on Advanced Computing and Communication Sys-
tems (ICACCS), vol.1, pp.1247–1252, 2024, IEEE
[7] “Sentiment Analysis in Customer Reviews for Product Recom-
mendation in E-commerce Using Machine Learning”, Panduro-
Ramirez, Jeidy, 2024 International Conference on Advances in
Computing, Communication and Applied Informatics (ACCAI),
pp.1–5, 2024, IEEE