Proceedings of the Fourth International Conference on Sentiment Analysis and Deep Learning (ICSADL-2025)

IEEE Xplore Part Number: CFP25UU5-ART; ISBN: 979-8-3315-2392-3

Deep Learning for Sentiment Analysis

Gayatri Rokade1 Radhe Ughade1 Palash Gaurshettiwar1
Computer Science and Medical Computer Science and Medical Computer Science and Medical
Engineering, Engineering, Engineering,
Faculty of Engineering and Faculty of Engineering and Faculty of Engineering and
Technology, Technology, Technology,
Sawangi(M), Wardha. Maharashtra. Sawangi(M), Wardha. Maharashtra. Sawangi(M), Wardha. Maharashtra.
India. India. India.
[email protected] [email protected] [email protected]
2025 4th International Conference on Sentiment Analysis and Deep Learning (ICSADL) | 979-8-3315-2392-3/25/$31.00 ©2025 IEEE | DOI: 10.1109/ICSADL65848.2025.10933099

Abstract— This paper represents that one of the critical and expanded items where characterizes item, sentiment
subfields of NLP, SA applies DL techniques to analyze the polarity, author and time of expressing. The fact that the
feelings expressed in text, image, and voice context. As new analysis is performed for a specific purpose post the division
technology of Web 2.0 emerges, there are many opportunities of the sentiments into several categories[3]. As the volume of
and challenges because of the large amount of content generated sentiment analysis on social media such as Facebook and
from blogs, forums, social networking sites, etc. This work Twitter increases, there are things that need to be addressed
explores the development of DL-based SA and notes that such as model reliability, reduction of process time and
enhancements have been made primarily on the document,
methods of varying type of data and domain specific. CNNs
sentence, phrase, as well as aspect levels. Ambiguity, sarcasm,
and RNNs are two deep learning models that have established
and negation are also discussed, as well as specific domain and
multimodal aspects. The work shows that it might be possible to
great promise in determining these problems, especially when
transition to semi-supervised, and unsupervised learning to it comes to processing short-term texts. Current research
overcome real-world data limitations and offers emergent shows how well these models extract sentiment division from
prospects in the real-time analysis, domain fine-tuning, and data social media data, with an importance on domain-specific
enhancement. The inspiration for future research is offered by applications such as marketing, finance, and healthcare[4].
the multimodal sentiment analysis and the enhancement of the
DL architectures. II. FACIAL EMOTION RECOGNITION TECHNIQUES
A. Pre-Processing
Keywords—Sentiment Analysis, Deep Learning,
Multimodal Analysis, Natural Language Processing, and Before the data is normalized the data is somewhat
Contextual Understanding. compressed and all forms of noise are removed from the data
set. The following are important steps:
I. INTRODUCTION (i) Facial Detection—It aims at realising faces in pictures,
Web 2.0 allows people share thoughts on social networks, frames or even in frames, which is highly equivalent to object
blogs, forums, providing useful information for opinion identification;
mining and recommender systems. Negation, domain
dependency and complications that come with NLP are some (ii) Dimension Reduction – reducing variables in dataset
of the challenges faced by sentiment analysis commonly for analysis and visualization work; and
known as SA because of the large amount of data from social (iii) Normalization—this is done to standardize ranges of
media. New research findings suggest that working with input data by scaling down or up the feature to other values
complex difficult changing data implies having to learn how and make training faster and more stable, techniques include
to improve accuracy for life[1]. Sentiment analysis focuses on Z Norma lization, Min-Max Normalization or the Unit Vector
the attitudes, beliefs, and feelings about particular ‘objects’ on Normalization. Together all these actions enhance the
the internet and social media, thanks to a wealth of available efficiency and effectiveness of the model[5].
information (Liu, 2010). In a phrase like “the laptop’s screen
is damaged, and its battery life is terrible” “screen” and B. Feature Extraction
“battery” are attributes of the object which is, in this case, “the The process of extracting characteristics that are
laptop.” Both text, audio, or Video or even images can be used significant for FER is known as feature extraction. It
to express feelings which may be neutral, positive or negative produces richer and smaller sets of characteristics that include
feelings[2]. elements like the corners, diagonal, and edges of the face as
In this regard it is noteworthy, that in connection with the well as other crucial details like the separation between two
growth of popularity of such Web 2.0 technologies as social eyes or the distance between the lips and eyes. This aids in
networks or e-shops, using of user-generated content rises. the quick learning of training data[6].
Natural language processing (NLP) has elevated sentiment C. Emotion Classficication
analysis to become one of the required methods of analyzing
this kind of data. It is a favorite for researchers and business Based on the characteristics that have been retrieved,
because it feeds wisdom from praising thoughts Sectioned on- algorithms are used to categorize the emotions. Images are
line. Thus, in sentiment analysis, the primary focuses are the categorized into different classifications using a variety of
identification and assessment of views and may envelop basic categorization techniques. Following completion of the

979-8-3315-2392-3/25/$31.00 ©2025 IEEE 240

Authorized licensed use limited to: Somaiya University. Downloaded on June 18,2025 at 06:11:12 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the Fourth International Conference on Sentiment Analysis and Deep Learning (ICSADL-2025)
IEEE Xplore Part Number: CFP25UU5-ART; ISBN: 979-8-3315-2392-3

preparatory stages of face identification and feature

extraction, a FER picture is classified. Later in the suggested
survey, other categorization methods are covered[7].
III. LEVEL OF SENTIMET ANALYSIS
The levels of document, sentence, phrase, and aspect have
all been studied in relation to sentiment analysis. Sentiment
analysis for documents, sentences, and phrases at every level.
A. Document level sentiment analysis
Document-level sentiment analysis assigns a single
polarity to the whole document following sentiment analysis.
Seldom is this kind of sentiment analysis employed. It is
useful for categorizing book chapters or pages as neutral,
FIG:- Sentiment analysis level[12].
negative, or positive. Both supervised and unsupervised
learning techniques can be used to the document's IV. CHALLENGES IN DEEP LEARNING FOR SENTIMENT
classification at this level. Document-level sentiment analysis ANALYSIS
faces two major challenges: cross-domain sentiment analysis
and cross-language sentiment analysis. (Saunders, 2021) Sentiment analysis has greatly benefited from deep
Sentiment analysis that is domain-specific has demonstrated learning, which makes it possible to extract emotional insights
exceptional accuracy while remaining extremely domain- from textual data. Nevertheless, a number of issues still hinder
sensitive. A constrained and domain-specific set of words these models' efficacy and precision. Some of the most
serves as the feature vector in these tasks[8]. important issues in deep learning for sentiment analysis are
listed below.
B. Sentence level sentiment analysis
A. Contextual Understanding
At this level of analysis, every sentence is examined and a
matching polarity is determined. When there are many Achieving a sophisticated contextual comprehension of
different and mixed emotions attached to a paper, this is quite language is one of the main issues in sentiment analysis.
helpful. Subjective categorization is linked to this Depending on their context, words may express a variety of
classification level. The same techniques employed at the emotions. As an example, the term "kill" may convey a good
document level will be used to independently identify the feeling in "the brand’s spicy wings are killing it with their
polarity of each sentence, but with more training data and flavor" yet a bad sentiment in "your customer support is killing
processing power. The polarity of each phrase can be used me". To prevent misclassification, this heterogeneity calls for
separately or together to determine the document's attitude. models that can understand words in their particular contexts.
Sentiment analysis at the document level is sometimes We present the contextual sentiment neural network (CSNN),
inadequate for particular applications. The goal of earlier a unique neural network model that can summarize the
sentence-level analytic work has been to identify subjective contents and explain the sentiment analysis prediction process
sentences. However, more challenging assignments, including in a way that people can understand[13].
dealing with ambiguous assertions or conditional clauses. B. Sarcasm and Irony Detection
Sentence-level sentiment analysis plays an important role in
these situations[9]. Irony and sarcasm detection is another major obstacle.
Although sayings like "Oh great, another delay!" seem upbeat,
C. Phrase level sentiment analysis they actually express disappointment or annoyance. Existing
An advanced technique within the larger subject of methods frequently miss these nuances, which results in
sentiment analysis, phase-level sentiment analysis focuses on inaccurate sentiment classifications. Inaccurate interpretations
comprehending sentiments expressed at various levels of of sarcasm can lead to serious mistakes in sentiment analysis
results. The inability to distinguish between ironic and
granularity, especially in respect to certain phases or
sarcastic interpretations has recently sparked scientific interest
characteristics of entities. For applications like product
in their identification from their literal counterparts. Detecting
evaluations or customer feedback, where it's critical to irony and sarcasm in languages other than English, such
comprehend the emotion toward particular features, this Italian, Japanese, Spanish, and Greek, has also been
approach is vital[10]. extensively studied. In the review analysis that follows, we
D. Aspect level sentiment analysis classify similar techniques based on the fundamental ideas
they have taken to deal with FL[14].
Sentiment analysis is carried out at the aspect level.
Multiple aspects may be present in a single sentence; C. Negation Handlling
therefore, sentiment analysis at the aspect level. An overall In natural language, negation is an element of language
sentiment has been generated for the entire sentence after that flips the meaning of statements. Due of their ability to
giving each element in the sentence primary attention and reverse a statement's emotion, negatives make sentiment
assigning polarity to each one[11].
analysis more difficult. For instance, "not bad" is often read
as positive, but models may mistakenly categorize it as
negative 3 if negation handling is not done appropriately.
Improving accuracy requires creating algorithms that can
recognize and understand negations. Negation handling is
regarded as one of the most challenging problems in natural

979-8-3315-2392-3/25/$31.00 ©2025 IEEE 241

Authorized licensed use limited to: Somaiya University. Downloaded on June 18,2025 at 06:11:12 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the Fourth International Conference on Sentiment Analysis and Deep Learning (ICSADL-2025)
IEEE Xplore Part Number: CFP25UU5-ART; ISBN: 979-8-3315-2392-3

language processing (NLP) for opinion mining and is a the capacity to capture complex emotions, such as sarcasm
significant subtask in sentiment analysis. The automatic and emotional nuances. high-performance real-time fine-
identification of polarity shifts in opinions presented in grain object identification framework that fixes a number of
natural language text format is known as "negative handling" issues with plant disease detection that make existing
in natural language processing[15]. approaches less effective, including textural similarity, multi-
scale object classes, irregular morphology, and dense
D. Multi-Class Sentiment Analysis
distribution[19].
Multi-class sentiment analysis adds complexity by going
beyond binary classifications (positive/negative). Because B. Real-Time Analysis
humans express their emotions in a variety of subtle ways, it Real-time sentiment research is now more important than
can be difficult for models to distinguish between distinct ever thanks to social media and instant messaging apps. Due
emotional states, such as neutral or mixed sentiments 12. This to the high rate with which deep learning models can work on
calls for complex algorithms that can recognize minute big amounts of data, organizations might quickly and almost
changes in sentiment. Sentiment analysis of operator in the blink of an eye identify the sentiments of the public
understanding and opinion data is done to content user needs concerning their products or services. Smack communication
and improve the value of the product or service. Moreover, it ensures that businesses can quickly address consumer
can contributing in influential whether or not customers are comments or embrace new trends enhancing their operation’s
involved in their goods. This might improve a product's competitive nature. Real time picture and video data are
presentation in the marketplace[16]. quantified using three You-Only-Look-Once (YOLO)-based
E. Data Quality and Bias deep learning (DL) models to identify if a worker is wearing
a hard helmet, vest or both[20].
A very important fact is how good the training data is, this
is because it determines the effectiveness of a sentiment
C. Multimodal Sentiment Analysis
analysis model. That is, the probability estimates of some
Deep learning enables the integration of textual, pictorial,
attitudes may be overestimated or underestimated when
and audio data all for sentiment analysis. Better
outcomes are skewed due to bias in the data set. This paper understanding of the public emotion may be achieved with
therefore concludes that appropriating and utilizing the help of messages that contain both the text and graphics
diversified training data set is key towards building strong
to consider the societal posts. These insights may be of deeper
models that operate optimally under different conditions.
data acquisition due to this multiple mode analysis thus
Applications of IoT are built on the quality of sensor data,
making it applicable in different fields such as brand building
and inadequate quality data make them ineffective. The
and advertising. Since sentiment analysis is increasing
objective of this systematic review is to provide an rapidly, the possibilities of multimodal sentiment analysis are
introduction and guide for academics with an interest in continuing to grow because the complementary data streams
quality-related issues of physical sensor data[17].
make it possible to achieve a more profound and precise
F. Domain Specificity identifying of sentiment based on data that go beyond
Domain based issues are generally a big problem with text[21].
most sentiment analysis models and this one is no exception.
Different languages and expressions in the context of social D. Domain-Specific Applications
network comments may result in low efficiency of a model Applications of deep learning models might be inclined to
trained for product evaluations. This implies the need to specific fields like entertainment, health or finance sectors.
design specific models for specific uses possibly going an Companies may enhance the performance accuracy of
extra extent of being time and resource consuming. A trained sentiment classification relevant to its industry by training
supervised learning classifier, the RF, if employed, can be models employing domain-specific data sets. Such as
utilised to extract hurricane-relevant tweets with domain analyzing the mood among the consumer in the financial
specificity. A domain-adversarial neural network (DANN) markets may help in decision making regarding the risks or
can then be used to identify the sentiment of hurricane- even when choosing investments. Moreover, we indicate
relevant tweets. The data-rich human and built environment qualitative and quantitative limitations in existing approaches
systems benefit from an intelligent and flexible disaster and propose a novel deep learning method named \sysname
information system that incorporates domain-specific (Re vivifying Function Sign at ure Inference using DEEP
sentiment analysis[18]. LEARNING) to incorporate domain knowledge relevant to
V. OPPORTUNITIES IN DEEP LEARNING FOR SENTIMENT the specific compiler optimization into the process[22].
ANALYSIS
E. Handling Large Datasets
A. Enhanced Accuracy and Precision In sentiment analysis, one of the strengths seen from deep
Deep learning models have demonstrated better learning algorithms is the ability to handle large data sets.
performance than conventional machine learning methods. Deep learning models are able to effectively manage the
Examples of these models include Convolutional Neural massive volumes of textual data that organizations produce
Networks (CNNs) and Recurrent Neural Networks (RNNs). via input from customers, social media interactions, and
These models have the ability to automatically extract reviews. Organizations may get insights from large data
intricate representations and patterns from huge datasets, thanks to its scalability, which eliminates the need for
which improves sentiment categorization accuracy. Market conventional computing constraints[23].
research and consumer feedback analysis might benefit from

979-8-3315-2392-3/25/$31.00 ©2025 IEEE 242

Authorized licensed use limited to: Somaiya University. Downloaded on June 18,2025 at 06:11:12 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the Fourth International Conference on Sentiment Analysis and Deep Learning (ICSADL-2025)
IEEE Xplore Part Number: CFP25UU5-ART; ISBN: 979-8-3315-2392-3

VI. FUTURE DIRECTIONS statements and they need advanced approaches for handling
poor quality textual data and preventing system bias. The
The quality and availability of data have a big influence
future development of sentiment analysis will require
on how well DL models function. Subpar outcomes might
advanced mathematical models to process complex linguistic
arise from inadequate data, highlighting the necessity of elements while showing interpretability features to users
effective preprocessing methods catered to the specifics of during application. These models also need to address ethical
the data. In DL, supervised learning is still prevalent, and dimensions and fairness aspects in their use to
MLP, CNN, and RNN architectures are frequently employed. promote/discourage sentiment analysis applicability.
Yet, accuracy and the relevance of the techniques to real life Sentiment analysis demonstrates revolutionary potential
applications could be enhanced through the development of across multiple domains through its ability to advance
new method or enhancement of existing ones. For example, business intelligence and customer service and social science
DL models require enormous input in data, especially labeled research while assisting political analysis because deep
data, and this can become an issue in domains such as health learning technologies can overcome previous obstacles.
care or in situations requiring real-time data processing. Digital data growth demonstrates that accurate sentiment
Possible solutions are offered through data generation analysis will become a more vital necessity.
techniques such as data augmentation and transfer learning
that make use of information from large datasets in training REFERENCES
on smaller datasets. In fact, to work on real data and not [1] Dang, Nhan Cach, María N. Moreno-García, and Fernando De la
Prieta. "Sentiment analysis based on deep learning: A comparative
necessarily with a large amount of labeled data, supervised study." Electronics 9, no. 3 (2020): 483.
learning has to give way to a relatively greater share of semi- [2] Yadav, Ashima, and Dinesh Kumar Vishwakarma. "Sentiment analysis
supervised and unsupervised learning. The relevance of using deep learning architectures: a review." Artificial Intelligence
DL[24]. Review 53, no. 6 (2020): 4335-4385.
It is noted that deep learning in sentiment analysis will [3] Habimana, O., Li, Y., Li, R. et al. Sentiment analysis using deep
have bigger breakthrough in the future as long as academics learning approaches: an overview. Sci. China Inf. Sci. 63, 111102
(2020). https://doi.org/10.1007/s11432-018-9941-6
continue dealing with current problem and exploring new
[4] Dang, Cach N., María N. Moreno-García, and Fernando De la Prieta.
methods. Other promising area of research is the possibility "Hybrid deep learning models for sentiment analysis." Complexity
to use several sources of information, i.e., text and voice data, 2021, no. 1 (2021): 9986920.
as well as videos for sentiment analysis. Unlike the method [5] Patel, Keyur, Dev Mehta, Chinmay Mistry, Rajesh Gupta, Sudeep
that just offers the textual data to analyze, this one also Tanwar, Neeraj Kumar, and Mamoun Alazab. "Facial sentiment
analysis using AI techniques: state-of-the-art, taxonomies, and
includes such extras as voice pitch and face expressions to challenges." IEEE access 8 (2020): 90495-90519.
better understand the mood. The ability to identify and [6] R. Gupta, S. Tanwar, S. Tyagi, and N. Kumar, “Tactile internet and its
classify the sentiment in the written part of the digital applications in 5g era: A comprehensive review,” International Journal
communication and social media will enhance the accuracy of Communication Systems, vol. 32, no. 14, p. e3981, 2019. e3981
and the coverage of the sentiment analysis systems as the dac.3981.
digital communication tools expand their functionality to [7] S. Tanwar, K. Parekh, and R. Evans, “Blockchain-based electronic
healthcare record system for healthcare 4.0 applications,” Journal of
multimedia space. Information Security and Applications, vol. 50, p. 102407, 2020.
[8] Behdenna, Salima, Fatiha Barigou, and Ghalem Belalem. "Document
ACKNOWLEDGMENT level sentiment analysis: a survey." EAI endorsed transactions on
The authors would like to express their gratitude to their context-aware systems and applications 4, no. 13 (2018): e2-e2.
colleagues and academic supervisors who contributed to the [9] Behdenna, Salima, Fatiha Barigou, and Ghalem Belalem. "Sentiment
discussion and whose feedback helped when preparing the analysis of Arabic tweets: opinion target extraction." Journal of
Digital Information Management 16, no. 6 (2018): 325.
review. We wish to express deep appreciation [to this
[10] Zhang, Yongfeng. "Incorporating phrase-level sentiment analysis on
organization/institution if any] for providing opportunity to textual reviews for personalized recommendation." In Proceedings of
use research infrastructures and tools that really helped to ease the eighth ACM international conference on web search and data
the task. In conclusion, the authors express the gratitude to the mining, pp. 435-440. 2015.
scientific community which has worked to make [11] Schouten, Kim, and Flavius Frasincar. "Survey on aspect-level
achievements in the sphere of deep learning and sentiment sentiment analysis." IEEE transactions on knowledge and data
analysis. engineering 28, no. 3 (2015): 813-830.
[12] Wankhade, M., Rao, A.C.S. & Kulkarni, C. A survey on sentiment
analysis methods, applications, and challenges. Artif Intell Rev 55,
5731–5780 (2022). https://doi.org/10.1007/s10462-022-10144-1
CONCLUSION [13] Ito, Tomoki, Kota Tsubouchi, Hiroki Sakaji, Tatsuo Yamashita, and
Kiyoshi Izumi. "Contextual sentiment neural network for document
Human emotions in various text media and audiovisual sentiment analysis." Data Science and Engineering 5, no. 2 (2020):
streams require vital understanding which sentiment analysis 180-192.
reveals as an essential interpretive technique. Social media [14] Potamias, R.A., Siolas, G. & Stafylopatis, A . A transformer-based
platforms producing user-generated content has created strong approach to irony and sarcasm detection. Neural Comput & Applic
demand for advanced methods which analyze public 32, 17309–17320 (2020). https://doi.org/10.1007/s00521-020-05102-
sentiment and market data along with customer feedback. 3.
Real-time datasets processing along with enhanced accuracy [15] Singh, Prakash Kumar, and Sanchita Paul. "Deep learning approach
becomes possible through CNNs and RNNs deep learning for negation handling in sentiment analysis." IEEE Access 9 (2021):
102579-102592.
techniques which demonstrate remarkable potential for
[16] Mithillesh, Kumar P., and M. Supriya. "Multi class sentiment analysis
sentiment analysis applications. Pixel arrays in deep learning- of 5g tweets." In 2022 3rd international conference for emerging
based models show promise but they still need better context technology (INCET), pp. 1-5. IEEE, 2022.
processing mechanisms to properly handle sarcastic or ironic

979-8-3315-2392-3/25/$31.00 ©2025 IEEE 243

Authorized licensed use limited to: Somaiya University. Downloaded on June 18,2025 at 06:11:12 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the Fourth International Conference on Sentiment Analysis and Deep Learning (ICSADL-2025)
IEEE Xplore Part Number: CFP25UU5-ART; ISBN: 979-8-3315-2392-3

[17] Teh, Hui Yie, Andreas W. Kempa-Liehr, and Kevin I-Kai Wang. [22] Lin, Yan, Debin Gao, and David Lo. "Resil: Revivifying function
"Sensor data quality: A systematic review." Journal of Big Data 7, signature inference using deep learning with domain-specific
no. 1 (2020): 11. knowledge." In Proceedings of the Twelfth ACM Conference on Data
[18] Yao, Fang, and Yan Wang. "Domain-specific sentiment analysis for and Application Security and Privacy, pp. 107-118. 2022.
tweets during hurricanes (DSSA-H): A domain-adversarial neural- [23] Li, Qian, Hao Peng, Jianxin Li, Congying Xia, Renyu Yang, Lichao
network-based approach." Computers, Environment and Urban Sun, Philip S. Yu, and Lifang He. "A survey on text classification:
Systems 83 (2020): 101522. From traditional to deep learning." ACM Transactions on Intelligent
[19] Roy, Arunabha M., Rikhi Bose, and Jayabrata Bhaduri. "A fast Systems and Technology (TIST) 13, no. 2 (2022): 1-41.
accurate fine-grain object detection model based on YOLOv4 deep [24] Taye, Mohammad Mustafa. "Understanding of machine learning with
neural network." Neural Computing and Applications 34, no. 5 deep learning: architectures, workflow, applications and future
(2022): 3895-3921. directions." Computers 12, no. 5 (2023): 91.
[20] Nath, Nipun D., Amir H. Behzadan, and Stephanie G. Paal. "Deep
learning for site safety: Real-time detection of personal protective
equipment." Automation in construction 112 (2020): 103085.
[21] Das, Ringki, and Thoudam Doren Singh. "Multimodal sentiment
analysis: a survey of methods, trends, and challenges." ACM
Computing Surveys 55, no. 13s (2023): 1-38.

979-8-3315-2392-3/25/$31.00 ©2025 IEEE 244

Authorized licensed use limited to: Somaiya University. Downloaded on June 18,2025 at 06:11:12 UTC from IEEE Xplore. Restrictions apply.