Volume 10, Issue 2, February – 2025 International Journal of Innovative Science and Research Technology

ISSN No:-2456-2165 https://doi.org/10.5281/zenodo.14959397

Exploring Sentiment Analysis through Deep

Learning: A Comprehensive Review
Rohit Ranvir1; Gopal Rathod2; Jay Gore3; Abhijit Zade4; Sachin Chavhan5
1,2,3,4
Student, Department of CSE, Babasaheb Naik College of Engineering, Pusad, Maharashtra
5
Assistant Professor, Department of CSE, Babasaheb Naik College of Engineering, Pusad, Maharashtra

Publication Date: 2025/03/03

Abstract: This survey paper provides an in-depth exploration of sentiment analysis on social media through the lens of
deep learning-based methods. It systematically examines various components essential to sentiment analysis, including
data classification, pre-processing techniques, text representations, and learning models, as well as their applications as
discussed in different research papers. The authors evaluate recent advancements in deep learning architectures,
highlighting their advantages and disadvantages in the context of sentiment analysis. Additionally, the paper addresses the
challenges faced in this field, such as the informal nature of social media language and the necessity for large and labelled
datasets, while also identifying factors that can enhance the accuracy of sentiment classification. A significant focus of the
article is on emotion recognition from text, which is recognized as a critical task within Natural Language Processing
(NLP) that can greatly benefit artificial intelligence and improve human interaction. To illustrate the effectiveness of deep
learning techniques, the authors propose a sentiment classification method utilizing Recurrent Neural Networks (RNN)
and Long Short-Term Memory (LSTM) networks, demonstrating high accuracy in emotion classification across three
different datasets. Overall, the paper offers valuable insights into the current state of sentiment analysis, emphasizing the
potential of deep learning to advance this important area of research. Recurrent Neural Network (RNN). Gated Recurrent
Unit (GRU) was reported as the best performer in terms of accuracy on benchmark datasets.

Keywords: Sentiment Analysis, Social Media, Deep Learning, Deep Learning, Pre- Processing.

How to Cite: Rohit Ranvir; Gopal Rathod; Jay Gore; Abhijit Zade; Sachin Chavhan (2025). Exploring Sentiment Analysis
through Deep Learning: A Comprehensive Review. International Journal of Innovative Science and Research
Technology, 10(2), 1193-1196. https://doi.org/10.5281/zenodo.14959397

I. INTRODUCTION consumer behavior. As a result, deep learning has become a

cornerstone in the development of advanced sentiment
Sentiment analysis, a subfield of Natural Language analysis systems, paving the way for more sophisticated
Processing (NLP), focuses on identifying and categorizing applications in various domains.
opinions expressed in text data, typically as positive,
negative, or neutral. With the exponential growth of online Businesses, researchers, and policymakers use
content, including social media, product reviews, and blogs, sentiment analysis to gain insights from user- generated
the demand for automated sentiment analysis has surged. content on platforms like Twitter, Facebook, and blogs. With
Traditional methods relied heavily on rule-based approaches the rise of big data, sentiment analysis has evolved from
and lexicon- based techniques, which often struggled to traditional machine learning techniques to advanced deep
capture the nuances of human emotions and contextual learning models. Deep learning methods, such as Recurrent
meanings. However, the advent of deep learning has Neural Networks (RNN) and Long Short-Term Memory
revolutionized sentiment analysis by enabling models to (LSTM), have significantly improved sentiment
learn complex patterns and representations from large classification accuracy. These models analyze text by
datasets. Techniques such as Convolutional Neural learning contextual relationships and patterns, making them
Networks (CNN) and Long Short-Term Memory (LSTM) more effective than conventional approaches. Sentiment
networks have proven particularly effective, as they can analysis can classify text into different categories, such as
process sequential data and retain contextual information positive, negative, and neutral sentiments. Advanced models
over long distances. further classify emotions into finer categories, including
happiness, sadness, anger, and excitement. Preprocessing
This capability allows deep learning models to achieve techniques, such as tokenization, stop word removal, and
higher accuracy and robustness in sentiment classification word embeddings, enhance the model's performance.
tasks, making them invaluable tools for businesses and
researchers seeking to understand public sentiment and

IJISRT25FEB1017 www.ijisrt.com 1193

Volume 10, Issue 2, February – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.5281/zenodo.14959397
 Classification of Sentiment Analysis
Sentiment analysis is a method used to analyze human
opinions or comments on products, services, or events
through various platforms. It is challenging due to the
multiple meanings in the text. Sentiment analysis is
performed using machine linguistics and natural language
processing tools, and it helps understand an individual's
psyche by analyzing emotions expressed in online text.
Opinion mining and emotion mining are the two most
common subcategories of sentiment analysis. Opinion
mining involves subjectivity detection and polarity
classification, while emotion mining involves emotion
detection, polarity classification, and emotion
categorization. sentiment analysis is divided into two types:
opinion mining and emotion mining. Opinion detection
includes opinion summarization, argument expression
detection, and opinion summarization. Emotion mining
includes emotion detection, emotion polarity classification,
and emotion cause detection. This review paper focuses on
deep learning-based opinion and emotion mining, which can
help in understanding an individual's psyche and
understanding their emotions expressed in online text.
Fig 1 Sentiment Analysis.

Fig 2 Taxonomy of Sentiment Analysis

II. LITERATURE REVIEW Additionally, LSTMs are particularly adept at handling

long sequences of text, making them suitable for
 Lei Zhang, etal ,The method discussed in the paper analyzing social media content where context is crucial.
involves utilizing deep learning techniques, particularly However, there are notable disadvantages, including the
Recurrent Neural Networks (RNN) and Long Short- need for large, labeled datasets for training, which can be
Term Memory (LSTM) networks, for sentiment analysis. difficult to obtain. Furthermore, the computational
This approach begins with preprocessing the text data to resources required for training deep learning models can
eliminate noise and irrelevant information, followed by be substantial, potentially limiting their use in smaller
feature extraction using techniques like word projects or organizations. Lastly, the informal and varied
embeddings. The processed data is then classified into nature of social media language can introduce challenges
various sentiment categories, including positive, in accurately interpreting sentiments [1].
negative, and neutral sentiments. One of the key
advantages of this method is its ability to achieve high  Shilpa C p.c, 2021 ,The method discussed in the paper
accuracy in sentiment classification, as deep learning focuses on sentiment analysis using deep learning
models can effectively capture complex patterns in data. techniques, specifically Recurrent Neural Networks

IJISRT25FEB1017 www.ijisrt.com 1194

Volume 10, Issue 2, February – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.5281/zenodo.14959397
(RNN) and Long Short-Term Memory (LSTM) on the other hand, provide higher accuracy and better
networks. The approach involves preprocessing Twitter handling of complex data but require significant
data to remove noise and irrelevant information, computational resources, large labeled datasets, and are
followed by feature extraction using techniques like TF- often seen as "black boxes" due to their complexity.
IDF and Doc2Vec. The processed data is then classified Researchers must choose based on their needs, balancing
into positive and negative sentiments, with further accuracy and resource availability.[4]
subclassification into specific emotions such as
happiness, anger, and sadness. One of the primary  Fengli Zhang a 2024 , Proposes in his paper the study is
advantages of this method is its high accuracy in the PEW- MCAB model, which combines Positional
emotion classification, achieving up to 91.3% accuracy Embedding with Glove Word Vectorization , a
for specific subclasses of emotions. The use of LSTM Multichannel Convolutional Neural Network , and an
networks helps mitigate issues like vanishing gradients, Attention-based Bidirectional Long Short-Term Memory
allowing for better handling of long- term dependencies for sentiment analysis of online movie reviews. The
in text data. Additionally, the Model's ability to model enhances text representation by incorporating
incorporate semantic and emotional word vectors word order information through positional embedding
enhances its predictive capabilities. However, there are and leverages Glove embeddings for global co-
some disadvantages to this approach. The reliance on occurrence statistics. PEW-MCAB model demonstrates
large datasets for training can be a limitation, as high accuracy 90.3% in sentiment analysis across
obtaining labeled data for sentiment analysis can be multiple datasets, outperforming other baseline models.
challenging and time-consuming. Furthermore, the The PEW-MCAB model excels in handling complex
complexity of deep learning models may require textual data by integrating word order information and
significant computational resources, making them less global context, leading to superior performance in
accessible for smaller organizations or projects. Lastly, sentiment analysis. It achieves high accuracy, precision,
the informal nature of Twitter data can introduce recall, and F1 scores across multiple datasets,
variability that complicates the sentiment classification outperforming traditional and deep learning baselines.
process.[2] The model's ability to capture both local and global
semantic features makes it robust for document-level
 Aadil Gani Ganie 2022 ,The method employed in the sentiment analysis. The model's complexity, due to the
paper for sentiment analysis involves the use of deep integration of multiple deep learning components (CNN,
learning techniques, specifically Recurrent Neural BiLSTM, attention mechanisms), may lead to higher
Networks (RNN) and Long Short-Term Memory computational costs and longer training times.
(LSTM) networks. This approach begins with Additionally, the reliance on pre- trained embeddings
preprocessing the text data to remove noise and like Glove may limit its adaptability to domain-specific
irrelevant elements, followed by feature extraction using language nuances, requiring fine-tuning for optimal
advanced techniques like word embeddings. The performance in specialized contexts. [5]
processed data is then classified into sentiment
categories, such as positive, negative, and neutral. The  Amit Sharma 2023 , Proposes in his paper for sentiment
paper discusses various accuracy percentages ranging analysis employs a hybrid approach that combines
from approximately 45.71% to 91.3% across different Natural Language Processing (NLP) techniques with
algorithms and datasets. Deep learning models (LSTMs) deep learning, specifically using Long Short-Term
are highly accurate classification due to their ability to Memory (LSTM) networks. This process begins with
learn complex patterns and manage long-term text data preprocessing to clean and prepare textual reviews
dependencies. However, they require large, labeled by removing noise and irrelevant information. Following
datasets, significant computational resources, and the this, a hybrid feature extraction technique is utilized,
informal nature of social media language, potentially which integrates review-related features and aspect-
leading to inaccuracies.[3] related features to create a comprehensive feature vector.
The LSTM model is then trained to classify sentiments
 Qurat Tul Ain 2017 Proposes in his paper ,The article into categories such as positive, negative, or neutral. The
reviews sentiment analysis methods, comparing paper presents the 94.46% accuracy results for the
traditional machine learning (e.g., Naive Bayes, SVM) proposed model: Disadvantages, including the need for
with deep learning techniques (e.g., CNN, LSTM, large labeled datasets for training, which can be difficult
RNN). Traditional methods rely on feature extraction to gather. The computational resources required for
techniques like TF-IDF and unigrams, while deep training deep learning models can also be significant,
learning uses advanced methods like Word2Vec and making it less accessible for smaller projects.
Glove for better context understanding. The paper Furthermore, the variability and informal nature of
mentions that the proposed models often outperform language in consumer reviews can introduce challenges
traditional models like SVM and Naive Bayes, achieving in accurately interpreting sentiments, potentially leading
accuracies in the range of 80% to 94.9% across various to misclassifications. [6]
datasets. Traditional methods are less resource-intensive
and easier to interpret but often achieve lower accuracy  Nikhil Sanjay Suryawanshi 2024, Proposes in his paper
due to limited feature extraction. Deep learning models, for sentiment analysis include traditional machine

IJISRT25FEB1017 www.ijisrt.com 1195

Volume 10, Issue 2, February – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.5281/zenodo.14959397
learning techniques like Naive Bayes, Support Vector approaches that integrate traditional and modern techniques,
Machines (SVM) ,In (SVM) Achieved accuracy of and developing strategies to improve data acquisition. By
83.31% on US airline Twitter data., and Logistic overcoming these challenges, the field of sentiment analysis
Regression, as well as deep learning approaches such as can continue to evolve, providing more sophisticated tools
Recurrent Neural Networks (RNNs), Long Short- Term for understanding public sentiment and consumer behavior
Memory (LSTM) networks, and Convolutional Neural in an increasingly digital world.
Networks (CNNs).In LATM Achieved accuracy of
81.20% on US airline sentiment dataset and In CNN REFERANCES
Achieved accuracy of 99.33% on IMDb dataset.
Traditional methods are advantageous due to their [1]. Lei Zhang, Lei Zhang Shuai, Wang, Bing Liu, “Deep
simplicity and effectiveness, especially when combined Learning for Sentiment Analysis, ”Jurnal
with robust feature selection techniques like n-grams and Administrasi Publik: arXiv preprint arXiv CS , vol .
TF-IDF. However, they often require extensive manual N/A , pp. N/A, January 2018.
feature engineering and struggle with capturing the [2]. Shilpa C p.c, Susmi Jacob, P. Vinod Cochin,
complexities of natural language. On the other hand, “Sentiment Analysis Using Deep Learning,”Jurnal
deep learning models excel in automatically learning Administrasi Publik: IEEE, vol . N/A , pp. N/A,
discriminative features from raw text, handling high- February 2021.
dimensional data, and capturing long-range [3]. Aadil Gani Ganie, Samad Dadvandipour,
dependencies in text. They also benefit from pre-trained “Traditional or deep learning for sentiment analysis
embeddings like word2vec and Glove, which enhance ,”Jurnal Administrasi Publik: Multidiszciplináris
semantic understanding. Despite their superior Tudományok , vol .12 , pp. 3-12, January 2022.
performance, deep learning models are computationally [4]. Qurat Tul Ain, Mubashir Ali, Amna Riaz, Amna
intensive, require large datasets, and can be less Noureen, Muhammad Kamran, Babar Hayat, A.
interpretable compared to traditional methods. Both Rehman, “Sentiment Analysis Using Deep Learning
approaches have their strengths and limitations, making Techniques: A Revie,” Jurnal Administrasi Publik:
the choice of method dependent on the specific context International Journal of Advanced Computer Science
and dataset characteristics.[7] and Applications (IJACSA), vol . 8 , pp . 6 , 2017.
[5]. Atandoh, Fengli Zhang, Mugahed A. Al-antari,
 Wael Etaiwi , 2021 , Proposes in his paper for sentiment Daniel Addo, Yeong Hyeon Gu, “Scalable deep
analysis employs deep learning techniques, particularly learning framework for sentiment analysis prediction
using Convolutional Neural Networks (CNN) and Long for online movie reviews, ” Jurnal Administrasi
Short-Term Memory (LSTM) networks. This approach Publik: Elsevier, Vol . 10, Issue 10, 30 May 2024.
begins with preprocessing the text data to clean it and [6]. Gagandeep Kaur & Amit Sharma, “A deep learning-
remove any irrelevant information. It then utilizes word based model using hybrid feature extraction approach
embeddings to convert words into numerical vectors that for consumer sentiment analysis, ” Jurnal
capture semantic meanings. In CNN Achieved accuracy Administrasi Publik: Springer, vol . 10 , Article
of 79% in multimedia sentiment analysis combining text number: 5 (2023).
and images. The CNN is used to extract local features [7]. Nikhil Sanjay Suryawanshi “Sentiment analysis with
from the text, while the LSTM captures long-range machine learning and deep learning: A survey of
dependencies, allowing for a comprehensive techniques and applications” Jurnal Administrasi
understanding of sentiment in the data. [8] Publik: N/A, vol . N/A, pp . N/A June 2024.
[8]. Wael Etaiwi, Dima Suleiman, Arafat Awajan, “ Deep
III. CONCLUSION Learning Based Techniques for Sentiment Analysis:
A Survey ” Jurnal Administrasi Publik: Informatica,
In conclusion, this survey highlights the significant vol . 45, pp.89-95 , 2021.
advancements in sentiment analysis through the application
of deep learning techniques, particularly in the context of
social media. Traditional methods have often struggled to
capture the complexities of human emotions, but deep
learning models, such as Recurrent Neural Networks (RNN)
and Long Short-Term Memory (LSTM) networks, have
demonstrated remarkable accuracy, achieving rates between
79% and 99.33%. Despite these successes, challenges
remain, including the need for large, labeled datasets and
substantial computational resources, which can hinder
accessibility for smaller projects. Additionally, the informal
nature of social media language can complicate sentiment
interpretation. Effective preprocessing and advanced feature
extraction methods, such as word embeddings, are crucial
for enhancing model performance. Future research should
focus on addressing these limitations, exploring hybrid

IJISRT25FEB1017 www.ijisrt.com 1196