Received 9 July 2023, accepted 15 August 2023, date of publication 21 August 2023, date of current version 29 August 2023.

Digital Object Identifier 10.1109/ACCESS.2023.3307308

Sentiment Analysis in E-Commerce Platforms:

A Review of Current Techniques and
Future Directions
HUANG HUANG , ADELEH ASEMI ZAVAREH , (Member, IEEE),
AND MUMTAZ BEGUM MUSTAFA , (Member, IEEE)
Department of Software Engineering, Faculty of Computer Science and Information Technology, University of Malaya, Kuala Lumpur 50603, Malaysia
Corresponding author: Mumtaz Begum Mustafa ([email protected])
This work was supported by the Ministry of Higher Education under the Fundamental Research Grant Scheme under
Grant FRGS/1/2020/ICT09/UM/02/1.

ABSTRACT Sentiment analysis (SA), also referred to as opinion mining, has become a widely used
real-world application of natural language processing in recent times. Its main goal is to identify the hidden
emotions behind the plain text. SA is especially useful in e-commerce fields, where comments and reviews
often contain a wealth of valuable business information that has great research value. The objective of
this study is to examine the techniques used for SA in current e-commerce platforms as well as the future
directions for SA in e-commerce. After examining the existing systematic review papers, it was found that
there is a lack of a single comprehensive review paper that addresses research questions. The findings of this
study can provide researchers in the field of SA with a comprehensive understanding of the current techniques
and platforms utilized, as well as provide insights into the future directions. Through the utilization of specific
keywords, we have identified 271 papers and have chosen 54 experimental papers for review. Among these,
26 papers (representing 48.%) have exclusively employed machine Learning techniques, while 24 (44.%)
have looked into addressing SA through deep learning techniques, and 4 (7.%) have employed a hybrid
approach using both machine learning and deep learning techniques. Additionally, our review revealed that
Amazon and Twitter emerged as the two most favored data sources among researchers. Looking ahead,
promising research avenues to include the development of more universal language models, aspect-based
SA, implicit aspect recognition and extraction, sarcasm detection, and fine-grained sentiment analysis.

INDEX TERMS Sentiment analysis(SA), E-commerce, natural language processing, machine learning, deep
learning, opinion mining.

I. INTRODUCTION Sentiment analysis (SA) is becoming a necessity as it

With the rapid development of smartphones, internet users enables businesses to gain valuable insights from the cus-
are changing from traditional information receivers to infor- tomer’s reviews about their products. Product reviews are
mation publishers. It has been revealed that e-commerce a rich source of customer feedback and opinions, and ana-
(EC) websites offer a significant volume of valuable infor- lyzing the sentiments expressed in these reviews provides
mation that surpasses the cognitive processing abilities of valuable information about how customers perceive and react
humans [3]. Zhang et al. mentioned that micro-blogs com- to products.
prise complex and copious sentiments that depict the user’s Sentiment analysis is widely used for product reviews
perspectives or viewpoints regarding a particular subject [10]. as it provides a comprehensive understanding of customer
feedback. By analyzing sentiments expressed in reviews,
businesses can assess product quality, gain insights into cus-
The associate editor coordinating the review of this manuscript and
tomer preferences, and identify areas for improvement. Sen-
approving it for publication was Ikramullah Lali. timent analysis also enables competitive analysis, supporting
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.
For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
VOLUME 11, 2023 90367
 H. Huang et al.: SA in E-Commerce Platforms: A Review of Current Techniques and Future Directions

businesses in understanding their position in the market. techniques. Sagnika et al. [55] conducted a review study
Additionally, it aids in decision-making processes, guiding focusing on machine-learning techniques for multilingual
product development, marketing strategies, and customer ser- SA. Mehta and Pandya [53] summarized various papers uti-
vice improvements. Effective SA further assists in brand lizing machine learning and lexicon analysis approaches.
reputation management by promptly addressing negative Shathik and Prasad [57] introduced the prevalent tech-
feedback and maintaining a positive brand image. niques used in SA from a machine learning perspective.
Existing SA techniques encompass rule-based, lexicon- Umar et al. [60] investigated the sentiment classification
based, machine learning, and deep learning techniques. level or data source on which supervised machine learning
Rule-based techniques utilize predefined rules or dictio- techniques like SVM, NB, Maximum Entropy, and other
naries, while lexicon-based techniques rely on sentiment technique such as lexicon-based which deliver optimal results
lexicons or dictionaries with annotated scores. Machine in Shathik and Prasad [57] and Umar et al. [60] solely encom-
learning algorithms like Support Vector Machines(SVM), passed a review focusing on machine learning techniques. It is
Naive Bayes(NB), and Random Forest to learn sentiment widely acknowledged that in addition to machine learning
patterns from labeled data. Deep learning techniques like techniques, deep learning techniques remain effective for
Recurrent Neural Network(RNN) and Convolutional Neu- addressing SA problems. Thus, there is a need for a more
ral Networks(CNN), excel at capturing complex patterns comprehensive review that covers both machine learning and
and context. These techniques provide valuable customer deep learning techniques.
insights, support reputation management, aid in market Other works offer a more thorough and organized exam-
research, enhance customer service, and optimize brand mar- nation of machine learning techniques and deep learning
keting efforts. approaches utilized in SA. Catelli et al. [30] compare
This study reviews papers published within the last five lexicon-based approaches and machine/deep learning tech-
years (from 2018 to 2022), exclusively narrowing the dis- niques in the context of SA, specifically focusing on the
cussion to machine learning and deep learning techniques. Italian market. Shayaa et al. [59] provided an overview
Despite the growing popularity of deep learning, existing of the publication trends in opinion mining and SA
research has yet to compare the efficiency of machine learn- from 2000 to 2016. Given the swift progress of techno-
ing and deep learning techniques for e-commerce sentiment logical advancements, it is important to conduct a study
analysis. This study aims to address this gap by offering that reviews papers published between 2018 and 2022,
an overview of different techniques utilized in e-commerce offering a more current viewpoint than earlier review
sentiment analysis. Besides, this study aims to fill the gap in studies.
research by summarizing the most popular e-commerce plat- Some of the review papers looks at the SA applications.
forms and identifying future directions of sentiment analysis Drus and Khalid [56] presented a comprehensive review
in e-commerce, as no previous study has done so. report on SA in social media, which investigates the tech-
In this study, section II of the paper delves into a review niques employed, social media platforms utilized, and their
of prior research on the application of sentiment analysis applications. Baragash and Aldowah [61] conducted a sys-
in e-commerce. Section III of this research focuses on the tematic review to explore the recent application of SA
methodology employed to review the techniques utilized in in higher education. The review aimed to categorize and
sentiment analysis for e-commerce, while Section IV con- identify the SA techniques and techniques commonly and
tains the results and pertinent discussions for each research effectively employed within the higher education domains.
question. Lastly, Section V provides a summary of the con- Obiedat et al. [54] focused on an Arabic SA review of the
clusions drawn from this review study. application of SA in social media, higher education domains,
and the Arabic texts approach. In addition, our study aims to
II. RESEARCH BACKGROUND investigate the utilization of SA, specifically in e-commerce.
A. EXISTING REVIEWS STUDIES ON SA IN E-COMMERCE The next group of review papers offers innovative ideas
Recent studies primarily focus on machine learning tech- from a technical perspective. Yue et al. [62] focus on typ-
niques for multi-lingual SA [55], SA in Arabic [54], recent ical techniques in the social media field of SA from three
approaches of implicit aspect extraction for SA [52], SA in distinct perspectives: task-oriented, granularity-oriented, and
social media and its application [56], application of SA methodology-oriented. Aspect extraction, is a component of
using machine learning techniques [57] and comparison aspect-based SA, involves identifying explicit and implicit
of lexicon-based and Bidirectional Encoder Representations aspects of a text. While explicit aspects are directly men-
from Transformers (BERT) based SA in Italian [30]. Despite tioned, implicit aspects need to be inferred. Detecting implicit
numerous studies on SA applied in e-commerce, there is cur- aspects poses challenges, yet it is crucial. Unfortunately,
rently no comprehensive work that summarizes the different there is limited research dedicated to the extraction of
techniques used, provides an overview of e-commerce data implicit aspects. Ganganwar and Rajalakshmi [52] surveyed
platforms, and suggests potential directions for future. recently proposed techniques for detecting implicit aspects.
Some existing review papers focus on consolidating Qazi et al. [65] examined the evidence presented in differ-
various machine-learning approaches to summarize SA ent reviews and outlined the challenges encountered by
90368 VOLUME 11, 2023
H. Huang et al.: SA in E-Commerce Platforms: A Review of Current Techniques and Future Directions

classification techniques in SA. They aim to comprehend study, we can get an overview of the most suitable tech-
how implementing enhanced techniques can address the con- nologies used in SA of e-commerce. Different approaches
ventional issues associated with classification techniques. are available, like machine learning techniques (eg. SVM,
These three related papers examine various aspects of SA Naive Bayes) require less data but more human intervention.
techniques, including different orientations, aspect extraction Machine learning is currently the most convenient technique
techniques, and classification techniques, all from a technical to train neural networks in the era of artificial intelligence.
standpoint. The primary limitation of these three papers is the Deep learning techniques such as Long Short-Term Memory
absence of a comprehensive summary regarding data sources (LSTM), and BERT do not require feature engineering but
and future directions in SA. rely on extensive data [16]. Deep learning techniques have
Others have also conducted similar research papers explor- demonstrated significant advancements by overcoming the
ing SA in e-commerce. Elzeheiry et al. [64] proposed an sequential limitations of prior models and introducing oper-
overview of the initial design for e-commerce is presented. ational parallelism, resulting in several advantages such as
Moreover, they discuss deep learning in e-commerce and context analysis, making the constituent embeddings more
SA. Then various versions of the commercial dataset are dynamic.
presented. Lastly, the challenges encountered by recom- This paper aims to provide an overview of the current SA
mendation systems and directions for future research are techniques in e-commerce, which will facilitate researchers
explained. Elzeheiry highlighted the challenges faced and the in obtaining a comprehensive understanding of the employed
future directions of recommendation systems. In contrast to technologies [30].
their study, our research aims to enumerate specific future In addition, this study summarizes the top e-commerce
research directions in SA. Marong et al. [66] provided a platforms that offer raw data, which can be analyzed to deter-
comprehensive overview of SA and relevant techniques in the mine the most popular and promising industries. By identify-
e-commerce sector, constantly seeking to under-stand con- ing the primary sources of e-commerce data, such as Amazon,
sumers’ opinions of their products and services. This paper Amazon is among the largest e-commerce platforms, with
primarily offers a technical perspective overview of SA in many reviews available for viewing. These findings can also
e-commerce. In contrast, our research aims to delve into the be applied to SA in similar platforms or industries, including
discussion of data sources and explore future directions in this social media platforms such as Twitter and Instagram, and
context. even anti-terrorism work. Moreover, this research provides a
Despite numerous studies on SA applied in e-commerce, valuable resource for future researchers to locate the appro-
no comprehensive work currently summarizes the different priate data sources quickly.
techniques(including machine learning and deep learning In our review, we identified various research gaps and
techniques) used over the past five years, provides an challenges that researchers would encounter in future work,
overview of e-commerce data platforms, and suggests poten- such as applying aspect-wise SA in e-commerce. With the
tial directions for future research. significant rise in user comments and reviews on social media
and e-commerce websites, there has been a corresponding
increase in the need to identify sentiment at the sentence and
B. IMPORTANCE OF THIS REVIEW STUDY aspect levels.On review boards, it is common to encounter
We cannot overstate the commercial value of SA in comments that are challenging to categorize as positive or
e-commerce. With the explosion of e-commerce in recent negative because of the presence of multiple aspects [54].
years, businesses need to be able to analyze and understand For example, comments such as ‘‘The restaurant has a nice
customer feedback on a massive scale. Through SA, busi- environment, but the food is terrible’’ has contrasting polar-
nesses can identify patterns and trends in customer sentiment, ities for different aspects. The utilization of technology for
such as which products or services are popular, which are aspect-wise SA in e-commerce remains a complex problem
not, and why. This information can help businesses make that requires further attention.
informed decisions about product development, marketing In practice, there are several promising research direc-
strategies, and customer service. Additionally, SA can help tions for the future, including more universal models in new
e-commerce businesses to stay ahead of their competition by domains and languages, Aspect-level SA models, implicit
identifying emerging trends and customer needs. The most Aspect Recognition and Extraction, Sarcasm detection, Lex-
prevalent form of feedback provided by customers on cur- icon based models introducing external knowledge, fine-
rent e-commerce platforms is through comments. Extracting grained SA. These topics will be discussed in section IV later.
insights from text analysis can serve as a useful reference
for other consumers and enable businesses on e-commerce III. REVIEW METHOD
platforms to enhance service quality and boost customer This study aims to thoroughly examine the literature on SA in
satisfaction [12]. the context of e-commerce. The primary aim of this review is
E-commerce platforms primarily consist of natural lan- to assess the present state of research in this area, focusing
guage sentences in their text or comments, which means on multiple aspects such as research objectives, the online
that SA is ubiquitous wherever text is present. From this shopping platforms analyzed, state-of-the-art SA techniques

VOLUME 11, 2023 90369

 H. Huang et al.: SA in E-Commerce Platforms: A Review of Current Techniques and Future Directions

employed, and diverse evaluation metrics employed to mea- • Paper type: Full text;
sure the effectiveness of the methodologies. In addition to • Language: English.
highlighting the current issues and challenges in SA for e-
Here are the exclusion criteria were utilized to filter out
commerce, this study also endeavors to suggest practical
irrelevant papers.
solutions and future directions for researchers in this field.
By leveraging the vast array of resources and data available on • Papers that do not specifically focus on SA and
online shopping platforms, this study aims to identify trends, e-commerce.
patterns, and key insights that can inform future research and • Papers that cover the topic of e-commerce with SA as a
guide the development of more effective and accurate SA supplementary subject.
models. This study aims to contribute to the advancement • Paper without detailed experiments
of SA in e-commerce and help researchers and practitioners A collection of 54 papers published between 2018 to
to better understand the challenges and opportunities ahead. 2022 have been sourced from various databases. The
Through a comprehensive review of existing works and a databases that were manually searched to procure these
critical analysis of the current state of research, this study papers include Science Direct, IEEE, and Google Scholar.
aims to provide a valuable resource for those seeking to make The papers have been handpicked to ensure that they meet the
significant contributions to this exciting and rapidly evolving relevance to the aims of the review. The selected papers pro-
field. vide the depth in SA topics, including cutting-edge technolo-
There are three stages in the review process: Research gies, innovative techniques, and groundbreaking research
questions raised in this review, Search methodology, Search findings, making it an invaluable resource for researchers,
results and analysis. scholars, and practitioners in the computer industry.

A. RESEARCH QUESTIONS RAISED IN THIS REVIEW STUDY

C. SEARCH RESULTS AND ANALYSIS
Here are the research questions to be addressed in this
Here, we will provide the results of the search and selection
research:
process, along with an examination of the chosen experimen-
Research Question 1: What are the current techniques used
tal papers concerning the publication databases utilized and
for sentiment analysis in e-commerce?
the explored techniques for conducting SA.
Research Question 2: What are the current e-commerce
After conducting a keyword search, we limited the papers
platforms that apply sentiment analysis?
to only those published between 2018 and 2022, identifying
Research Question 3: What are the future directions for
271 papers. We manually analyzed the titles and abstracts of
sentiment analysis in e-commerce?
the shortlisted papers to remove irrelevant ones, resulting in
the removal of 139 papers and leaving us with 132 papers.
B. SEARCH METHODOLOGY
We then read the full papers and applied the exclusion criteria,
Here, we employed a comprehensive search methodology
which resulted in the rejection of 78 papers and left us with
that encompassed various online databases and a meticulous
54 relevant papers. The final number of papers selected for
scrutiny of the chosen literature. Our search spanned across a
the review study was 54.
broad range of prominent databases such as:
As to the source databases, 16 articles (30.%), were
• Science Direct; obtained from the IEEE database. 34 papers (63.%) were
• IEEE; retrieved from Google Scholar, while 2(4.%) others from
• Google Scholar. Science Direct database.
Furthermore, we conducted a manual assessment in which we From Figure 1 below, it is clear that out of the 54 selected
examined the title and abstract of the papers. Subsequently, papers, 26 (48.%) employed machine learning techniques,
we chose the relevant papers while disregarding any that were while 24 (44.%) exclusively utilized deep learning tech-
deemed irrelevant. niques. Additionally, 4 papers (7.%) use both machine
The key words are as follows: and deep learning techniques. These findings indicate that
• SA; machine learning and deep learning techniques hold equal
• e-commerce; weight in the realm of SA.
• opinion mining;
• machine learning; IV. RESULTS AND DISCUSSION
• deep learning. A. RESEARCH QUESTION: WHAT ARE THE CURRENT
We have utilized the subsequent inclusion criteria in order to TECHNIQUES USED FOR SENTIMENT ANALYSIS IN
filter our initial search results and ensure that only pertinent E-COMMERCE?
studies are considered in our analysis. SA, also referred to as opinion mining, is a practical appli-
• Search domain: Artificial Intelligence, Computer Sci- cation of natural language processing (NLP) aims to identify
ence or Information Technology the hidden emotion behind the plain text. Chandio et al. [46]
• Publication type: Journals; stated that SA falls under the umbrella of text classification,

90370 VOLUME 11, 2023

H. Huang et al.: SA in E-Commerce Platforms: A Review of Current Techniques and Future Directions

FIGURE 1. Proportion of machine learning and deep learning techniques.

specifically focusing on classifying user sentiments into pos- 1) MACHINE LEARNING BASED TECHNIQUES
itive, negative, or neutral categories. Machine Learning techniques need human intervention to
In natural language processing tasks, SA is crucial, and improve the model performance. Before training, if an inac-
there are various ways to approach it. In previous studies, curate prediction occurs while training the model, an engineer
deep learning techniques would be considered a subset of has to step in and make adjustments. However, machine learn-
machine learning techniques. However, because of its dif- ing techniques require a smaller data size and less training
ferences with other machine learning techniques, we have time than deep learning models. Machine learning models
separated deep learning techniques and other machine learn- have lower computing complexity than deep learning tech-
ing techniques in this study. In this study, all other machine niques.
learning techniques except deep learning techniques are From Figure 2 and Table 1, it is clear that SVM
called machine learning techniques. [1,2,4,6,8,10,14,16,19,24-25,31,38-39,46,50,73,75] and NB
Machine learning techniques rely on statistical techniques [2,4,6,8,10,14-16,24-25,34,39-40,45,48,50] are the most
to learn from data and make predictions. These techeniques welcome machine learning techniques used in selected
use features extracted from the text, such as word frequen- papers.
cies, sentiment lexicons, and grammatical structures, to train Some researchers employ SVM models directly for their
a model that can classify the sentiment of a given text. studies. Huang et al. [1] use neural network (NN) model and
Jagdale et al. [4] indicated that NB and SVM are among the SVM model as their prediction models. In Hantoro et al. [73],
popular machine learning-based algorithms utilized in SA. SA was performed on Shopee application using SVM classi-
In contrast, deep learning based techniques leverage neural fication. The analysis utilized 990 training data samples and
networks to automatically learn representations of the text 110 test data samples. Out of the total test data of 110 assess-
and in recent time shown encouraging result in SA. These ments, 28 evaluations were classified as negative, while the
deep learning algorithms can capture complex relationships remaining 82 evaluations were classified as positive. The
between words and can handle large volumes of data without accuracy rate was 80.90%, indicating that 89 assessments
the need for manual feature [13]. Deep learning techniques were correctly classified in their respective sentiment classes.
used in SA includes CNN, RNN, LSTM, Transformer, BERT Pratama et al. [75] employed the SVM to classify reviews of
and Generative Pre-Trained (GPT) models. e-commerce beauty products. Their objective is to develop
From Figure 1, the majority of research papers rely y on a model capable of categorizing beauty product reviews
machine learning techniques, where 48% of the papers uti- and analyzing the accuracy of the classification. Elmurngi
lizes machine learning for SA. Furthermore, the trend towards and Gherbi [8] conducted a comparison of four supervised
deep learning methodologies is rising, with 44% of the papers machine learning techniques, namely NB, Decision Tree J48
employing this state-of-the-art technique. Notably, machine (DT-J48), Logistic Regression (LR), and SVM, for detect-
learning and deep learning techniques were applied in the ing unfair reviews. In Haque al. [14], the performance of
remaining 8% of the papers. Notably, both machine and deep various classification models, including Linear SVM, Multi-
learning techniques are commonly used in SA in e-commerce nomial NB, Stochastic Gradient Descent, Random Forest,
fields. LR, and DT, was evaluated on Amazon product reviews.

VOLUME 11, 2023 90371

 H. Huang et al.: SA in E-Commerce Platforms: A Review of Current Techniques and Future Directions

TABLE 1. Techniques comparison used in selected papers.

90372 VOLUME 11, 2023

H. Huang et al.: SA in E-Commerce Platforms: A Review of Current Techniques and Future Directions

TABLE 2. E-commerce platforms used in selected papers.

VOLUME 11, 2023 90373

 H. Huang et al.: SA in E-Commerce Platforms: A Review of Current Techniques and Future Directions

FIGURE 2. Techniques used in selected papers.

The findings revealed that SVM outperformed other tech- Amazon, specifically focusing on smart phones. By assessing
niques and achieved the most accurate classification results. the impact of the negation algorithm on SA tasks, they found
Dey et al. [16] conducted a comparison between two machine that the RNN, in combination with our negation marking
learning techniques to analyze the sentiment of customer processing, achieved the highest accuracy of 95.67%.
reviews on Amazon products. Experimental results have val- Many studies compared NB techniques and various
idated that SVM exhibits a higher accuracy rate in polarizing other classification techniques. Jagdale [4] utilized different
the feedback of Amazon products. machine learning techniques to classify Amazon reviews,
Linear Support Vector Classification (LinearSVC), which resulting in an accuracy of 98.17% for Naive Bayes and
is an alternative implementation of SVM algorithms, also 93.54% for SVM, specifically in the case of Camera Reviews.
demonstrates superior performance compared to other Basani et al. [6] applied NB and SVM techniques to Amazon
machine learning algorithms. Ahmed et al. [2] employed reviews datasets, and their accuracy and execution time were
applied linear SVC, Naive Bayes, and LR for SA. The results compared. The results demonstrate that the SVM outperforms
indicated that the linear SVM outperformed other classi- the NB in terms of accuracy, and the SVM also exhibits
fiers. A medicine recommendation system was developed shorter execution time than NB. In Bayhaqy’study [34], the
by Garg [24] utilizing patient reviews to predict sentiment. performance of DT, K-Nearest Neighbor (KNN), and NB was
Various machine learning techniques were employed, and the compared on Tweets datasets. The highest result obtained
results indicate that the Linear SVC classifier, combined with was from the NB techniques, achieving an accuracy of 77%,
TF-IDF vectorization, outperforms all other models with an precision of 88.50%, and recall of 64%. Lutfi et al. [39],
accuracy of 93%. performed SA on sales reviews of an Indonesian marketplace,
However, the performance could be better when com- utilizing SVM and NB. The result shows that SVM with a lin-
paring SVM with deep learning techniques like RNN. ear kernel has better than NB. Wongkar and Angdresey [45]
Dashtipour et al. [19] propose a novel hybrid framework for carried out a comparative study using NB, SVM, and K-NN
concept-level SA for Persian. The framework integrates lin- techniques in RapidMiner. The results revealed an accuracy
guistic rules and deep learning techniques to enhance polarity value of 75.58% for NB, 63.99% for SVM, and 73.34% for
detection and optimize analysis. Their proposed framework K-NN.
performs superior to state-of-the-art approaches, including SA was not limited to classifying sentiments as positive
SVMs and LR. Mukherjee et al. [25] incorporated a tailored or negative, but also incorporated for assessing the qual-
negation marking algorithm to detect explicit negation, and ity of online shopping platforms. Sari et al. [15], analyzed
they conducted experiments using various machine learning data from Tokopedia, one of Indonesian largest e-commerce
algorithms such as NB, SVMs, ANN, and RNN for SA of services, to evaluate the quality of its services. The NB

90374 VOLUME 11, 2023

H. Huang et al.: SA in E-Commerce Platforms: A Review of Current Techniques and Future Directions

classification technique was chosen due to its high accuracy and LSTM. LSTM is utilized for aspect term classifica-
and ability to handle large amounts of data. The findings indi- tion and aspect category classification tasks, while CNN is
cated that the dimensions of personalization and reliability employed for a polarity classification task. Murfi et al. [74]
required more attention, as they many negative sentiments. build upon previous work on hybrid deep learning approaches
On the other hand, the dimensions of trust and web design by incorporating BERT representation for Indonesian SA.
received high positive sentiments, indicating excellent service Their simulations demonstrate that utilizing BERT represen-
quality. The responsiveness dimension had a balanced mix tation enhances the accuracies of all hybrid architectures.
of positive and negative sentiments. Aagte et al. [48], found Particularly, the BERTLSTM-CNN model achieves slightly
that certain parties tried to post fake reviews on competi- higher accuracies than other hybrid architectures based on
tors’ websites. They suggest the integration of NB technique BERT.
to identify spam reviews. They cover the preprocessing of RNN imitates the process of human reading because of its
data prior to aspect identification, which involves classifying ability to capture sequential information and model depen-
reviews as positive, negative, or neutral. Then aspect ranking dencies over time. Just like humans read and comprehend text
is performed using NB. by considering the context and previous information, RNNs
To summarize, NB, DT, LR, and SVM are commonly used can process input data sequentially, preserving information
techniques in SA. Among them, SVM and NB techniques from previous steps and incorporating it into the current step.
are popular. Previous studies have compared the performance In SA, various types of RNN and their variations have been
of different techniques, and SVM and NB have consistently widely applied. Muhammad et al. [7] has demonstrated the
shown better results than others. NB can also be applied in utilization of Word2Vec and LSTM for sentiment classifi-
assessing the quality of online shopping platforms and detect- cation in hotel reviews. Agarap [11] utilized univariate and
ing spam. This highlights the versatility of SA. Additionally, multivariate analyses on various dataset features, excluding
some studies have compared SVM and NB with deep learning review titles and texts. They implemented a bidirectional
techniques like RNN, and the results have demonstrated that RNN with LSTM units for recommendation and sentiment
RNN outperforms other machine learning techniques. The classification tasks, and they observed that the bidirectional
subsequent sections will delve further into the discussion of LSTM achieved an F1-score of 0.88 for recommendation
deep learning techniques. classification and 0.93 for sentiment classification. A bidirec-
tional LSTM is a type of RNN that processes input sequences
2) DEEP LEARNING BASED TECHNIQUES in both forward and backward directions [11]. It uses two
In addition to machine learning techniques, deep learning are hidden states for each direction, allowing it to capture past
extensively employed in SA. A well-behaved deep learning and future context when making predictions or analyzing data
technique can identify whether its prediction is precise or not sequences. This enables the bidirectional LSTM to have a
based on its neural network (without human’s help). more comprehensive understanding of the input sequence and
In recent studies, BERT [3,21,26,30,36], CNN [3,5,12,19, potentially improve performance in tasks such as sequence.
23,35,74], RNN [2,3,11,13,22,25,27,74], LSTM [7,9,11,19, RNN imitates the process of human reading because of its
23,35,47,49,76] classification or prediction are the most fre- ability to capture sequential information and model depen-
quently used. As the model’s development, more complicated dencies over time. Just like humans read and comprehend text
models such as CNN-based bi-directional long short-term by considering the context and previous information, RNNs
memory (BiLSTM) (CNN-BiLSTM), attentionbased Bidi- can process input data sequentially, preserving information
rectional Gated Recurrent Unit (BiGRU) models have also from previous steps and incorporating it into the current step.
been developed to enhance the precision and efficiency of In SA, various types of recurrent neural networks (RNN)
deep learning models. and their variations have been widely applied. Muhammad,
Initially employed in computer vision domains, CNN [7] has demonstrated the utilization of Word2Vec and LSTM
exhibits exceptional performance by emulating the visual per- for sentiment classification in hotel reviews. Agarap [11] uti-
ception process of humans. Its application in SA represents lized univariate and multivariate analyses on various dataset
a significant innovation, leveraging the strengths of CNN to features, excluding review titles and texts. They implemented
capture and analyze textual information. Minaee et al. [23] a bidirectional recurrent neural network (RNN) with long-
introduces a model that combines LSTM and CNN. The short-term memory (LSTM) units for recommendation and
LSTM component is utilized to capture the temporal infor- sentiment classification tasks, and they observed that the
mation of the data, while the CNN component is employed bidirectional LSTM achieved an F1-score of 0.88 for rec-
to extract the local structure from the data. CNN can leverage ommendation classification and 0.93 for sentiment classifica-
the local correlations and patterns inherent in data through tion. A bidirectional LSTM is a type of RNN that processes
its learned feature maps. This distinctive characteristic of input sequences in both forward and backward directions.
CNN makes it particularly powerful, and it has been suc- It uses two hidden states for each direction, allowing it to
cessfully introduced to analyze text, enabling the extraction capture past and future context when making predictions
of meaningful features from textual data. Zuheros et al. [35] or analyzing data sequences. This enables the bidirectional
developed the ABSA model incorporated with both CNN LSTM to have a more comprehensive understanding of the

VOLUME 11, 2023 90375

 H. Huang et al.: SA in E-Commerce Platforms: A Review of Current Techniques and Future Directions

input sequence and potentially improve performance in tasks commonly utilized technique, being employed in 10 out of
such as sequence classification or prediction. the 54 papers. Both LSTM and GRU are derived from RNNs,
Other deep learning algorithms or attention mechanisms and their recurrence mechanisms for resolving seq2seq issues
are often combined with RNN or its variants for vari- have demonstrated exceptional performance.
ous applications. Meng et al. [9] presented a SA approach In summary, the dominant deep learning techniques in
called Feature Enhanced Attention CNN-BiLSTM that oper- NLP consist of RNN and its variations, such as LSTM and
ates at the aspect level. Their technique utilizes CNN to GRU. While CNN was originally developed for computer
extract higher-level phrase representations from the embed- vision, it has also found applications in text analysis. RNN
ding layer, enhancing subsequent coding tasks. They employ algorithms can take different forms, including Bidirectional
BiLSTM to capture local phrase features and global and RNN and Bidirectional LSTM, expanding their capabilities.
temporal sentence semantics to improve context encoding Additionally, RNN can be effectively combined with other
quality and preserve semantic information. An attention deep learning algorithms or integrated with attention mecha-
mechanism is also incorporated to model the interaction nisms to concentrate on specific aspects of textual data.
between aspect words and sentences, enabling an ade-
quate context representation. Xu et al. [47] introduced an B. RESEARCH QUESTION: WHAT ARE THE CURRENT
enhanced word representation approach incorporating sen- E-COMMERCE PLATFORMS THAT APPLY SENTIMENT
timent information into the traditional TF-IDF technique, ANALYSIS?
resulting in weighted word vectors. These weighted word The data source can significantly influence the accuracy and
vectors are then fed into a BiLSTM model to effectively cap- usefulness of collected data during the data collection. The
ture contextual information and improve the representation of SA task aims to understand and interpret human opinions on
comment vectors. The sentiment tendency of the comments various objects, topics, and events. In this regard, comments
is determined using a feedforward neural network classifier. and feedback from online shopping platforms have emerged
Iqbal et al. [77] implemented deep learning-LSTM and RNN as one of the most abundant and informative data source.
for sentiment classification and analysis. Three models were By analyzing the data from various sources, we can gain an
introduced, each utilizing different architectural variations overview of the different fields where SA is applied. This
and parameter tuning based on these deep-learning tech- study has divided the data sources into four categories: online
niques. Chandio et al. [46] proposed RU-BiLSTM, a deep shopping platforms, social media, travel agencies, and other
recurrent architecture for SA of Roman Urdu. The model sources. In other words, SA is commonly utilized in popular
combines BiLSTM with word embedding and an atten- areas: online shopping, social media, travel industries, and
tion mechanism. By utilizing BiLSTM, the model captures other fields.
context information in both directions, while the attention Based on Figure 3 and Table 2, it is evident that Ama-
mechanism focuses on essential features. Finally, the last zon [24,6,8,13-14,16-18,21-22,25,30-32,75] is the primary
dense softmax output layer is employed to obtain binary and platform from which researchers collect data, with 17 out
ternary classification results. Huang, Lin, et al. [76] proposed of 50 papers utilizing it. Amazon, the largest online retailer
a ERNIE-BiLSTM-Att (EBLA). The technique incorporates in the world, is a treasure trove of customer feedback data.
the Enhanced Representation through Knowledge Integra- Through its platform, customers from all over the globe post
tion (ERNIE) word embedding model to generate dynamic their thoughts and opinions on various products and services,
word vectors. These vectors are fed into a BiLSTM network creating a massive repository of comments invaluable to data
to extract text features. The Attention Mechanism (Att) is scientists and market researchers.
applied to refine the weights of the hidden layer. Lastly, soft- Social media platforms such as IMDB [23,28,36,49] and
max is utilized as the output layer for sentiment classification. Twitter [9,33-34,42,44-45] are also crucial data sources, with
Shrestha and Nasoz [13] organized and arranged review 11 out of 50 papers using them. Given the widespread use of
embeddings to create a product sequence to enhance the sen- social media and other online platforms for purchasing goods
timent classification process. This sequence was inputted into and services, analyzing customer feedback and comments has
a recurrent gated unit (GRU) to learn the product embedding. become integral to SA. These comments provide valuable
The review embeddings generated from paragraph vectors insights into the customer’s perspective on a product or ser-
and product embeddings from GRU were used to train a SVM vice and help businesses identify areas for improvement and
for sentiment classification. GRU [13] is another type of RNN enhance customer satisfaction. In addition, applying natural
architecture used for sequential data analysis, such as natural language processing techniques can aid in evaluating the sen-
language processing tasks. GRU has gating mechanisms that timent conveyed in these comments, facilitating the automatic
enable it to selectively update and forget information from categorization of positive, negative, or neutral sentiments.
previous time steps allowing it to capture long-term depen- This data can then be utilized to enhance marketing strategies,
dencies in the data efficiently. product development, and customer service.
Figure 2 shows that the employment of various deep learn- Besides the above websites, there are other online plat-
ing techniques is relatively evenly distributed among the forms where users can express their opinions and share their
groups. Among the selected papers, LSTM was the most experiences. Websites such as Ctrip [3,47], TripAdvisor [35]

90376 VOLUME 11, 2023

H. Huang et al.: SA in E-Commerce Platforms: A Review of Current Techniques and Future Directions

and Traveloka [7,40] are just a few examples of popular learning or deep learning techniques. Even in state-ofthe-
platforms where customers can leave reviews and ratings for art models like GPT-4, the support for minority languages
hotels, restaurants, and tourist attractions. These customer remains limited. A universal model which can be applied
sentiments are invaluable to other travelers looking for infor- across domains and languages still has a long way to go.
mation to help plan their trips. By reading reviews, travelers For one specific machine learning or deep learning model,
can understand what to expect and make more informed deci- in most situations, it can only perform well in one specific
sions about where to stay, what to eat, and what to see during domain [3]. Hong [5] states that in the subsequent study,
their travels In other industries, data comes from different online feedback on various types of fresh agricultural prod-
platforms but all of them are in service industries platforms ucts will be considered as the research subject to enhance
including Women’s Clothing reviews, comments from dan- the generalizability of the experimental outcomes. In Puspita
gdang.com, restaurant and laptop reviews from SemEval Kencana Sari’s study, they clarified a thorough investiga-
2014 twitter dataset, The Alibaba Group and JD, Flipkart tion of the data collection process is necessary, particularly
eBay, Walmart, BestBuy, gross box-office revenue data for regarding the specific duration during which data is col-
the movies. lected. To obtain more representative results, additional data
source such as social media platforms like Twitter and Face-
C. RESEARCH QUESTION: WHAT ARE THE FUTURE book should be utilized to evaluate service quality [15]. For
DIRECTIONS FOR SENTIMENT ANALYSIS IN instance, if a model is fed and trained by digital products
E-COMMERCE? review, it cannot perform well in restaurant service fields.
According to the reviewed papers, many future direction Alhassan Mabrouk states that in the current system, they only
suggestions are given by related authors. concentrated on comparing a group of products belonging
to the same category and summarizing their aspect-based
1) MORE UNIVERSAL MODELS IN NEW DOMAINS AND opinions. However, they also aim to extend its application to
LANGUAGES other domains such as movies or restaurants [3]. In Wongkar
According to researchers, a particular machine or deep learn- and Angdresey they intend to examine the sentiment of public
ing model may only suit a single language. For example, satisfaction with the performance of the elected president of
if English was used to train a language model, its performance the Republic of Indonesia by analyzing data from other social
with Chinese or Arabic could be poor, suggesting that each media platforms, such as Facebook and Instagram. A compre-
language requires its unique language model. hensive language model that can analyze all languages in all
In most implemented models, the data source is mainly in domains should be proposed [45].
English. Therefore, those models can only perform better at
predicting English inputs. Conducting cross-cultural studies 2) ASPECT-LEVEL SENTIMENT ANALYSIS TECHNIQUES
in SA by obtaining sentiment reviews in languages other than Previous studies have shown a shift from coarse-grained
English is still a promising direction. SA to fine-grained SA. Meanwhile, the research target of
Researchers have conducted SA in various languages SA is changing from document-level SA to sentence-level
other than English. Dashtipour et al. [19] presented a novel SA. However, traditional coarse-grained SA cannot meet the
hybrid framework for concept-level SA for Persian lan- requirements for business analysis, more fine-grained SA is
guage. Their framework combines linguistic rules and deep a must for more advanced business analysis.
learning techniques to enhance the polarity detection and The emergence of aspect-based SA (ABSA) has garnered
achieve better results. Yang et al. [12] proposed a model significant attention from researchers in machine learning
called SLCABG, which stands for Sentiment Lexicon-based and natural language processing. Most research studies in
CNN with attention-based Bidirectional Gated Recurrent this area are centered around two main directions: neural
Unit (BiGRU). It combines the sentiment lexicon with networks and attention mechanisms. With the integration of
CNN and BiGRU to perform SA. The model is trained attention mechanisms into ABSA, there has been a significant
using Chinese comments data obtained from dangdang.com. increase in the utilization of hybrid methods that combine
Saleh et al. [67] proposed an optimized heterogeneous stack- neural networks and attention mechanisms. These hybrid
ing ensemble model to improve the performance of Arabic approaches have gained widespread popularity as they cap-
SA. The model is designed to effectively combine multiple italize on the respective strengths of both techniques.
techniques to achieve better accuracy and results in analyzing Han et al. [20] proposed the Pretraining and Multi-task
sentiments in Arabic text. Noor et al. [38] focuses on analyz- learning based on Double BiGRU (PM-DBiGRU). The
ing Urdu Roman reviews obtained from Daraz.pk, one of the PM-DBiGRU model leverages pre-trained weights obtained
most popular and widely accessed e-commerce websites in from a sentiment classification task on short text-level drug
Pakistan. reviews to initialize the relevant weights in our model.It
SA research focused on languages such as English, Chi- utilizes two Bidirectional Gated Recurrent Unit (BiGRU)
nese, Persian, Arabic, and Roman Urdu. However, there is a networks to generate bidirectional semantic representations
lack of extensive studies and training on other languages such of the target and drug review. Additionally, an attention mech-
as Japanese, Cantonese, and German text data using machine anism is employed to capture target-specific representations

VOLUME 11, 2023 90377

 H. Huang et al.: SA in E-Commerce Platforms: A Review of Current Techniques and Future Directions

FIGURE 3. E-commerce platforms.

for aspect-level drug reviews. To further enhance the perfor- aspect-based SA, another novel approach is the utilization
mance, they employ multi-task learning to transfer valuable of generative models. Hosseini-Asl et al. [68] focuses on
domain knowledge from the short text-level drug review few-shot settings and aim to transform the extraction and
corpus. Meng et al. [9] presented a neural network model prediction tasks into a sequence generation task. To achieve
for aspect-level SA known as Feature Enhanced Attention this, they employ a generative language model with unidi-
CNN-BiLSTM. To enhance the context encoding quality rectional attention, primarily using GPT2 unless specified
and preserve semantic information, they utilize BiLSTM to otherwise. By adopting this approach, the model learns to
capture not only the local features of phrases but also the perform the tasks through language generation, eliminating
global and temporal sentence semantics. Additionally, they the requirement for training task-specific layers. Up to this
incorporate an attention mechanism to model the interaction point, the performance of GPT-4 in this research has been
relationships between aspect words and sentences. This atten- exceptionally impressive, indicating a promising research
tion mechanism allows them to focus on the keywords related direction, especially for developing large language models.
to the targets, enabling them to learn more effective con-
text representations Feng et al. [77] included a customized 3) IMPLICIT ASPECT RECOGNITION AND EXTRACTION IN
masked attention mechanism designed for ABSA. They intro- ASPECT-BASED SENTIMENT ANALYSIS
duced two different approaches for generating the mask. The Over the past years, aspect extraction has emerged as a
first technique involves setting an attention weight threshold critical stage in SA for performing condensed sentiment clas-
based on the maximum weight value and retain only the sification. Nonetheless, prior research on SA has primarily
attention scores above that threshold. The second approach concentrated on extracting explicit aspects, with relatively
was to select the top words with the highest weights. Both little attention given to implicit aspects.
techniques aim to eliminate lower score components that are Implicit aspects can be inferred from the sentence’s context
deemed less relevant to the targeted aspect, thereby focusing and the world’s knowledge. For instance, if someone says,
on the most important aspects. ‘‘This phone is running fast,’’ we understand that they are
Despite recent advancements, the current state-of-the-art referring to the phone’s performance even though ‘‘perfor-
models for aspect-level SA are limited in number. Aspect- mance’’ is not explicitly mentioned. Similarly, when someone
based SA should be applicable across various types of says, ‘‘I cannot see anything in this room,’’ we know they are
datasets, ensuring its universality. Nguyen et al. [26] states referring to the lack of light in the room, even though the word
that they plan to expand the proposed technique for aspect- ‘‘lightness’’ is not explicitly mentioned.
based SA, and they plan to conduct more comprehensive Maitama et al. [69] have identified techniques for extract-
experiments on various datasets. In addition to employ- ing implicit aspects, explicit aspects, or both. Wang et al. [70]
ing neural networks and attention mechanisms to address have developed a model called ‘‘Hierarchical Knowledge

90378 VOLUME 11, 2023

H. Huang et al.: SA in E-Commerce Platforms: A Review of Current Techniques and Future Directions

Enhancement and Multi-pooling’’ (HKEM) that effectively incorporating knowledge graphs or other sources of external
integrates knowledge information from various levels within information. Sait et al. [72] developed a sarcasm classifica-
the text. This integration is achieved through hierarchical tion technique called Deep Learning with Natural Language
knowledge enhancement, addressing the issue of ‘‘weak fea- Processing (DLNLP-SA). The DLNLP-SA technique utilizes
tures’’ to enhance the overall performance. While the study deep learning and natural language processing to detect and
takes into account the overlapping of features in two distinct classify instances of sarcasm in the input data.
segments using the domain’s feature hierarchy, it has limited However, despite these efforts, sarcasm detection remains
effectiveness in aspect classification. Wei et al. [71] proposed a challenging problem in SA, and further research is needed to
an implicit SA model named BiLSTM with multi-polarity improve the accuracy of SA in identifying and understanding
orthogonal attention. By utilizing multi-polarity attention sarcastic language.
instead of the traditional single attention model, they can A different form of sarcasm known as ‘‘numerical sar-
distinguish the variance between words and sentiment orien- casm’’ was discussed by Kumar et al. [51], which is com-
tation. This disparity serves as a crucial feature in implicit SA. monly observed on social media platforms. This type of
To improve the performance of their model, they incorporates sarcasm is based on alterations in numerical values that
external knowledge bases, such as a common-sense knowl- ultimately impact the polarity of the text. An example of
edge base, into the model. Becasue external knowledge is numerical sarcasm would be:
commonly implicit, it is not explicitly mentioned in sentences ‘‘They drive so slowly—only 20 km/h.’’ (Non-sarcastic)
or directly indicated by the context. Therefore, it is still ‘‘They drive so slowly—only 160 km/h.’’ (Sarcastic)
essential for them to determine an appropriate technique to It can be challenging to identify sarcasm when it is con-
represent this external knowledge. cealed within numerical data or the context of a statement
Co-occurrence, semantic-based approaches, ontology, during SA. Detecting covert sarcasm in text requires con-
CRF, SVM, LSTM, Hierarchy, CNN, lexicon-based tech- textual understanding and can be addressed using algorithms
niques, matrix factorization, and topic modeling are com- such as RNN, LSTM, GRU, Transformers, or Bert models.
monly used in solving implicit aspect extraction problems. These models are either descended from RNN algorithms
Implicit aspect extraction is recognized as a relatively new or rely on attention mechanisms. The efficiency of sar-
and inherently ambiguous area, relying more on semantic casm detection has greatly improved due to the contextual
understanding rather than explicit indications. Consequently, understanding capabilities of RNN algorithms and the atten-
a considerable number of studies suggest that implicit aspect tion mechanism. The future of sarcasm detection models is
extraction holds promising potential as a future direction. expected to be dominated by large language models.

5) FINE-GRAINED SENTIMENT ANALYSIS

4) SARCASM DETECTION In traditional SA classes, researchers tend to classify senti-
Using positive words to express negative sentiments is a ments into only two categories: Positive or negative, alterna-
common feature of sarcastic text, which poses a challenge tively, the polarity can be classified exclusively as positive,
for SA models that are not explicitly designed to detect it. negative, or neutral [49]. However, sentiment should not be
In real-life situations, individuals may express their opinions limited to two or three groups. In the future, there will be a
sarcastically, which can be challenging to discern. Therefore, greater emphasis on performing More fine-grained SA.
in future work, it is necessary to enhance the technique to In Chandio’s [49] they plan to create a benchmark dataset
classify sentiments in such cases [29]. in Roman Urdu to facilitate fine-grained SA and question
Sarcasm is frequently found in user-generated content, classification. This fine-grained SA will encompass five dis-
such as Facebook comments and tweets, and detecting it tinct categories, namely strongly negative, weakly negative,
accurately in SA requires a thorough understanding of the neutral, weakly positive, and strongly positive. In Yang’s
context, topic, and environment. study, the approach presented is limited to classifying sen-
Sarcasm can be a challenge for SA because it involves timent into only positive and negative categories, which may
using language that is intended to convey the opposite mean- not be adequate in domains that demand more fine-grain SA.
ing of what is being said. This can be difficult for machines Hence, the subsequent step involves exploring the classifica-
to detect because the words used may appear positive or neg- tion of sentiment granularity in text [12].
ative in isolation, but their intended meaning is the opposite.
For example, ‘‘Great, now I have to walk home in V. CONCLUSION
the rain’’ may appear positive because ‘‘great’’ is usually This paper concluded that recent papers applied SA in e-
associated with positive sentiment. However, the intended commerce. It categorized the models used based on the
meaning is negative because the person is expressing frus- approaches and the platform from which the source data was
tration about walking home in the rain. To address the obtained. This paper also discussed the limitations and future
sarcasm problem in SA, researchers have explored various works as suggested by the researchers. Despite the challenges
techniques such as using contextual information, analyz- faced in applying SA in e-commerce, its importance cannot
ing the speaker’s tone of voice or facial expressions, and be overlooked, and further research in this area is warranted.

VOLUME 11, 2023 90379

 H. Huang et al.: SA in E-Commerce Platforms: A Review of Current Techniques and Future Directions

SA applied in e-commerce relies on machine learning and [19] K. Dashtipour, M. Gogate, J. Li, F. Jiang, B. Kong, and A. Hussain,
deep learning techniques, each with their own advantages ‘‘A hybrid Persian sentiment analysis framework: Integrating depen-
dency grammar based rules and deep neural networks,’’ Neurocomputing,
and disadvantages. The primary data sources for SA, such as vol. 380, pp. 1–10, Mar. 2020.
Amazon, Twitter, and IMDB, remain crucial. In the future, [20] Y. Han, M. Liu, and W. Jing, ‘‘Aspect-level drug reviews sentiment analysis
researchers can explore applying SA to other e-commerce based on double BiGRU and knowledge transfer,’’ IEEE Access, vol. 8,
pp. 21314–21325, 2020.
platforms, such as stock exchange platforms. Future studies
[21] H. Xu, B. Liu, L. Shu, and P. S. Yu, ‘‘BERT post-training for review
could focus on developing more universal models for new reading comprehension and aspect-based sentiment analysis,’’ 2019,
domains and languages, aspect-level SA models, implicit arXiv:1904.02232.
aspect recognition and extraction, sarcasm detection, and [22] H. T. Ismet, T. Mustaqim, and D. Purwitasari, ‘‘Aspect based sentiment
analysis of product review using memory network,’’ Sci. J. Informat.,
fine-grained SA to increase the usage of SA in e-commerce. vol. 9, no. 1, pp. 73–83, May 2022.
It is hoped that more attention will be given to these areas. [23] S. Minaee, E. Azimi, and A. Abdolrashidi, ‘‘Deep-sentiment: Senti-
ment analysis using ensemble of CNN and bi-LSTM models,’’ 2019,
arXiv:1904.04206.
REFERENCES [24] S. Garg, ‘‘Drug recommendation system based on sentiment analysis of
[1] L. Huang, Z. Dou, Y. Hu, and R. Huang, ‘‘Textual analysis for online drug reviews using machine learning,’’ in Proc. 11th Int. Conf. Cloud
reviews: A polymerization topic sentiment model,’’ IEEE Access, vol. 7, Comput., Data Sci. Eng. (Confluence), Jan. 2021, pp. 175–181.
pp. 91940–91945, 2019. [25] P. Mukherjee, Y. Badr, S. Doppalapudi, S. M. Srinivasan, R. S. Sangwan,
[2] H. M. Ahmed, M. J. Awan, N. S. Khan, A. Yasin, and H. M. F. Shehzad, and R. Sharma, ‘‘Effect of negation in sentences on sentiment analysis and
‘‘Sentiment analysis of online food reviews using big data analytics,’’ polarity detection,’’ Proc. Comput. Sci., vol. 185, pp. 370–379, Jan. 2021.
Elementary Educ. Online, vol. 20, no. 2, pp. 827–836, 2021. [26] Q. T. Nguyen, T. L. Nguyen, N. H. Luong, and Q. H. Ngo, ‘‘Fine-
[3] A. Mabrouk, R. P. D. Redondo, and M. Kayed, ‘‘SEOpinion: Summa- tuning BERT for sentiment analysis of Vietnamese reviews,’’ in Proc. 7th
rization and exploration of opinion from E-commerce websites,’’ Sensors, NAFOSTED Conf. Inf. Comput. Sci. (NICS), Nov. 2020, pp. 302–307.
vol. 21, no. 2, p. 636, Jan. 2021. [27] N. Majumder, S. Poria, A. Gelbukh, M. S. Akhtar, E. Cambria, and
[4] R. S. Jagdale, V. S. Shirsat, and S. N. Deshmukh, ‘‘Sentiment analysis A. Ekbal, ‘‘IARM: Inter-aspect relation modeling with memory networks
on product reviews using machine learning techniques,’’ in Proc. Cogn. in aspect-based sentiment analysis,’’ in Proc. Conf. Empirical Methods
Inform. Soft Comput. (CISC). Singapore: Springer, 2019, pp. 639–647. Natural Lang. Process., 2018, pp. 3402–3411.
[5] W. Hong, C. Zheng, L. Wu, and X. Pu, ‘‘Analyzing the relationship between [28] A. Dubey, A. Gupta, N. Raturi, and P. Saxena, ‘‘Item-based collaborative
consumer satisfaction and fresh E-commerce logistics service using text filtering using sentiment analysis of user reviews,’’ in Proc. 1st Int. Conf.
mining techniques,’’ Sustainability, vol. 11, no. 13, p. 3570, Jun. 2019. Appl. Comput. Commun. Technol. (ICACCT), Delhi, India. Singapore:
[6] Y. Basani, H. V. Sibuea, S. I. P. Sianipar, and J. P. Samosir, ‘‘Application of Springer, Mar. 2018, pp. 77–87.
sentiment analysis on product review E-commerce,’’ J. Phys., Conf. Ser., [29] H. Nguyen, T. Huynh, S. Hoang, V. Pham, and I. Zelinka, ‘‘Language-
vol. 1175, no. 1, Mar. 2019, Art. no. 012103. oriented sentiment analysis based on the grammar structure and improved
[7] P. F. Muhammad, R. Kusumaningrum, and A. Wibowo, ‘‘Sentiment self-attention network,’’ in Proc. 15th Int. Conf. Eval. Novel Approaches
analysis using word2vec and long short-term memory (LSTM) for Softw. Eng. (ENASE), 2020, pp. 339–346.
Indonesian hotel reviews,’’ Proc. Comput. Sci., vol. 179, pp. 728–735, [30] R. Catelli, S. Pelosi, and M. Esposito, ‘‘Lexicon-based vs. BERT-based
Jan. 2021. sentiment analysis: A comparative study in Italian,’’ Electronics, vol. 11,
[8] E. I. Elmurngi and A. Gherbi, ‘‘Unfair reviews detection on Amazon no. 3, p. 374, Jan. 2022.
reviews using sentiment analysis with supervised learning techniques,’’ [31] N. Nandal, R. Tanwar, and J. Pruthi, ‘‘Machine learning based aspect level
J. Comput. Sci., vol. 14, no. 5, pp. 714–726, May 2018. sentiment analysis for Amazon products,’’ Spatial Inf. Res., vol. 28, no. 5,
[9] W. Meng, Y. Wei, P. Liu, Z. Zhu, and H. Yin, ‘‘Aspect based sentiment pp. 601–607, Oct. 2020.
analysis with feature enhanced attention CNN-BiLSTM,’’ IEEE Access, [32] S. Zhang and H. Zhong, ‘‘Mining users trust from E-commerce
vol. 7, pp. 167240–167249, 2019. reviews based on sentiment similarity analysis,’’ IEEE Access, vol. 7,
[10] S. Zhang, Z. Wei, Y. Wang, and T. Liao, ‘‘Sentiment analysis of Chinese pp. 13523–13535, 2019.
micro-blog text based on extended sentiment dictionary,’’ Future Gener. [33] P.-F. Pai and C.-H. Liu, ‘‘Predicting vehicle sales by sentiment anal-
Comput. Syst., vol. 81, pp. 395–403, Apr. 2018. ysis of Twitter data and stock market values,’’ IEEE Access, vol. 6,
[11] A. F. Agarap, ‘‘Statistical analysis on E-commerce reviews, with sentiment pp. 57655–57662, 2018.
classification using bidirectional recurrent neural network (RNN),’’ 2018, [34] A. Bayhaqy, S. Sfenrianto, K. Nainggolan, and E. R. Kaburuan, ‘‘Sen-
arXiv:1805.03687. timent analysis about E-commerce from tweets using decision tree,
[12] L. Yang, Y. Li, J. Wang, and R. S. Sherratt, ‘‘Sentiment analysis for E- K-nearest neighbor, and Naïve Bayes,’’ in Proc. Int. Conf. Orange Technol.
commerce product reviews in Chinese based on sentiment lexicon and deep (ICOT), Oct. 2018, pp. 1–6.
learning,’’ IEEE Access, vol. 8, pp. 23522–23530, 2020. [35] C. Zuheros, E. Martínez-Cámara, E. Herrera-Viedma, and F. Herrera,
[13] N. Shrestha and F. Nasoz, ‘‘Deep learning sentiment analysis of Ama- ‘‘Sentiment analysis based multi-person multi-criteria decision making
zon.com reviews and ratings,’’ 2019, arXiv:12704.04096. methodology using natural language processing and deep learning for
[14] T. U. Haque, N. N. Saber, and F. M. Shah, ‘‘Sentiment analysis on large smarter decision aid. Case study of restaurant choice using TripAdvisor
scale Amazon product reviews,’’ in Proc. IEEE Int. Conf. Innov. Res. reviews,’’ Inf. Fusion, vol. 68, pp. 22–36, Apr. 2021.
Develop. (ICIRD), May 2018, pp. 1–6. [36] Y. Liu, J. Lu, J. Yang, and F. Mao, ‘‘Sentiment analysis for e-commerce
[15] P. K. Sari, A. Alamsyah, and S. Wibowo, ‘‘Measuring E-commerce service product reviews by deep learning model of BERT-BiGRU-softmax,’’ Math.
quality from online customer review using sentiment analysis,’’ J. Phys., Biosci. Eng., vol. 17, no. 6, pp. 7819–7837, 2020.
Conf. Ser., vol. 971, Mar. 2018, Art. no. 012053. [37] S. Riaz, M. Fatima, M. Kamran, and M. W. Nisar, ‘‘Opinion mining on
[16] S. Dey, S. Wasif, D. S. Tonmoy, S. Sultana, J. Sarkar, and M. Dey, large scale data using sentiment analysis and k-means clustering,’’ Cluster
‘‘A comparative study of support vector machine and naive Bayes classifier Comput., vol. 22, no. S3, pp. 7149–7164, May 2019.
for sentiment analysis on Amazon product reviews,’’ in Proc. Int. Conf. [38] F. Noor, M. Bakhtyar, and J. Baber, ‘‘Sentiment analysis in E-commerce
Contemp. Comput. Appl. (ICA), Feb. 2020, pp. 217–220. using SVM on Roman Urdu text,’’ in Proc. 2nd Int. Conf. Emerg.
[17] S. Zhang, D. Zhang, H. Zhong, and G. Wang, ‘‘A multiclassification Technol. Comput. (iCETiC), London, U.K. Cham, Switzerland: Springer,
model of sentiment for E-commerce reviews,’’ IEEE Access, vol. 8, Aug. 2019, pp. 213–222.
pp. 189513–189526, 2020. [39] A. A. Lutfi, A. E. Permanasari, and S. Fauziati, ‘‘Sentiment analysis
[18] S. Wassan, X. Chen, T. Shen, M. Waqar, and N. Z. Jhanjhi, ‘‘Amazon in the sales review of Indonesian marketplace by utilizing support vec-
product sentiment analysis using machine learning techniques,’’ Revista tor machine,’’ J. Inf. Syst. Eng. Bus. Intell., vol. 4, no. 1, pp. 57–64,
Argentina de Clínica Psicolgica, vol. 30, no. 1, p. 695, 2021. 2018.

90380 VOLUME 11, 2023

H. Huang et al.: SA in E-Commerce Platforms: A Review of Current Techniques and Future Directions

[40] R. Watrianthos, S. Suryadi, D. Irmayani, M. Nasution, and [62] L. Yue, W. Chen, X. Li, W. Zuo, and M. Yin, ‘‘A survey of sentiment
E. F. Simanjorang, ‘‘Sentiment analysis of traveloka app using nave analysis in social media,’’ Knowl. Inf. Syst., vol. 60, pp. 617–663, Jul. 2019.
Bayes classifier method,’’ Int. J. Sci. Technol. Res., vol. 8, no. 7, [63] M. Hilario, D. Esenarro, I. Petrlik, and C. Rodriguez, ‘‘Systematic liter-
pp. 786–788, 2019. ature review of sentiment analysis techniques,’’ J. Contemp. Issues Bus.
[41] M. Umer, I. Ashraf, A. Mehmood, S. Kumari, S. Ullah, and G. Sang Government, vol. 27, no. 1, pp. 506–517, 2021.
Choi, ‘‘Sentiment analysis of tweets using a unified convolutional neu- [64] S. Elzeheiry, W. A. Gab-Allah, N. Mekky, and M. Elmogy, ‘‘Sentiment
ral network-long short-term memory network model,’’ Comput. Intell., analysis for e-commerce product reviews: Current trends and future
vol. 37, no. 1, pp. 409–434, Feb. 2021. directions,’’ Preprints, May 2023, Art. no. 2023051649. [Online].
[42] E. Prabhakar, M. Santhosh, A. H. Krishnan, T. Kumar, and R. Sudhakar, Available: https://www.preprints.org/manuscript/202305.1649/v1, doi:
‘‘Sentiment analysis of U.S. airline Twitter data using new Adaboost 10.20944/preprints202305.1649.v1.
approach,’’ Int. J. Eng. Res. Technol., vol. 7, no. 1, pp. 1–6, 2019. [65] A. Qazi, R. G. Raj, G. Hardaker, and C. Standing, ‘‘A systematic literature
[43] M. H. Munna, M. R. I. Rifat, and A. S. M. Badrudduza, ‘‘Sentiment review on opinion types and sentiment analysis techniques,’’ Internet Res.,
analysis and product review classification in E-commerce platform,’’ in vol. 27, no. 3, pp. 608–630, 2017.
Proc. 23rd Int. Conf. Comput. Inf. Technol. (ICCIT), Dec. 2020, pp. 1–6. [66] M. Marong, N. K. Batcha, and R. Mafas, ‘‘Sentiment analysis in E-
commerce: A review on the techniques and algorithms,’’ J. Appl. Technol.
[44] I. Awajan, M. Mohamad, and A. Al-Quran, ‘‘Sentiment analysis technique
Innov., vol. 4, no. 1, p. 6, 2020.
and neutrosophic set theory for mining and ranking big data from online
[67] H. Saleh, S. Mostafa, A. Alharbi, S. El-Sappagh, and T. Alkhalifah, ‘‘Het-
reviews,’’ IEEE Access, vol. 9, pp. 47338–47353, 2021.
erogeneous ensemble deep learning model for enhanced Arabic sentiment
[45] M. Wongkar and A. Angdresey, ‘‘Sentiment analysis using naive Bayes
analysis,’’ Sensors, vol. 22, no. 10, p. 3707, May 2022.
algorithm of the data crawler: Twitter,’’ in Proc. 4th Int. Conf. Informat.
[68] E. Hosseini-Asl, W. Liu, and C. Xiong, ‘‘A generative language model for
Comput. (ICIC), Oct. 2019, pp. 1–5.
few-shot aspect-based sentiment analysis,’’ 2022, arXiv:2204.05356.
[46] B. Chandio, A. Shaikh, M. Bakhtyar, M. Alrizq, J. Baber, A. Sulaiman, [69] J. Z. Maitama, N. Idris, A. Abdi, L. Shuib, and R. Fauzi, ‘‘A systematic
A. Rajab, and W. Noor, ‘‘Sentiment analysis of Roman Urdu on E- review on implicit and explicit aspect extraction in sentiment analysis,’’
commerce reviews using machine learning,’’ Comput. Model. Eng. Sci., IEEE Access, vol. 8, pp. 194166–194191, 2020.
vol. 131, no. 3, pp. 1263–1287, 2022. [70] H. Wang, M. Hou, F. Li, and Y. Zhang, ‘‘Chinese implicit sentiment anal-
[47] G. Xu, Y. Meng, X. Qiu, Z. Yu, and X. Wu, ‘‘Sentiment analysis of ysis based on hierarchical knowledge enhancement and multi-pooling,’’
comment texts based on BiLSTM,’’ IEEE Access, vol. 7, pp. 51522–51532, IEEE Access, vol. 8, pp. 126051–126065, 2020.
2019. [71] J. Wei, J. Liao, Z. Yang, S. Wang, and Q. Zhao, ‘‘BiLSTM with
[48] A. A. Aagte, W. Vrushali, P. Vishwakarma, and S. Kamble, ‘‘Spam detec- multi-polarity orthogonal attention for implicit sentiment analysis,’’ Neu-
tion using sentiment analysis,’’ in Proc. 10th Int. Conf. Comput. Knowl. rocomputing, vol. 383, pp. 165–173, Mar. 2020.
Eng. (ICCKE), 2019, pp. 209–214. [72] A. R. W. Sait and M. K. Ishak, ‘‘Deep learning with natural language pro-
[49] B. A. Chandio, A. S. Imran, M. Bakhtyar, S. M. Daudpota, and J. Baber, cessing enabled sentimental analysis on sarcasm classification,’’ Comput.
‘‘Attention-based RU-BiLSTM sentiment analysis model for Roman Syst. Sci. Eng., vol. 44, no. 3, pp. 2553–2567, 2023.
Urdu,’’ Appl. Sci., vol. 12, no. 7, p. 3641, Apr. 2022. [73] K. K. Hantoro, D. Handayani, and S. Setiawati, ‘‘A implementation of text
[50] F. A. Ramadhan, R. R. P. Ruslan, and A. Zahra, ‘‘Sentiment analysis of mining in sentiment analysis of Shopee Indonesia using SVM,’’ Bull. Inf.
E-commerce product reviews for content interaction using machine learn- Technol., vol. 3, no. 2, pp. 115–120, 2022.
ing,’’ Cakrawala Repositori IMWI, vol. 6, no. 1, pp. 207–220, Feb. 2023. [74] H. Murfi, Syamsyuriani, T. Gowandi, G. Ardaneswari, and S. Nurrohmah,
[51] L. Kumar, A. Somani, and P. Bhattacharyya, ‘‘‘Having 2 hours to write ‘‘BERT-based combination of convolutional and recurrent neural network
a paper is fun!’: Detecting sarcasm in numerical portions of text,’’ 2017, for Indonesian sentiment analysis,’’ 2022, arXiv:2211.05273.
arXiv:1709.01950. [75] M. R. Pratama, F. A. G. Soerawinata, R. R. Zhafari, Rendy, and
[52] V. Ganganwar and R. Rajalakshmi, ‘‘Implicit aspect extraction for sen- H. N. Imanda, ‘‘Sentiment analysis of beauty product E-commerce using
timent analysis: A survey of recent approaches,’’ Proc. Comput. Sci., support vector machine method,’’ Jurnal RESTI, Rekayasa Sistem dan
vol. 165, pp. 485–491, Jan. 2019. Teknologi Informasi, vol. 6, no. 2, pp. 269–274, Apr. 2022.
[53] P. Mehta and S. Pandya, ‘‘A review on sentiment analysis methodolo- [76] W. Huang, M. Lin, and Y. Wang, ‘‘Sentiment analysis of Chinese E-
gies, practices and applications,’’ Int. J. Sci. Technol. Res., vol. 9, no. 2, commerce product reviews using ERNIE word embedding and attention
pp. 601–609, 2020. mechanism,’’ Appl. Sci., vol. 12, no. 14, p. 7182, Jul. 2022.
[77] A. Feng, X. Zhang, and X. Song, ‘‘Unrestricted attention may not be all
[54] R. Obiedat, D. Al-Darras, E. Alzaghoul, and O. Harfoushi, ‘‘Arabic aspect-
you need–masked attention mechanism focuses better on relevant parts in
based sentiment analysis: A systematic literature review,’’ IEEE Access,
aspect-based sentiment analysis,’’ IEEE Access, vol. 10, pp. 8518–8528,
vol. 9, pp. 152628–152645, 2021.
2022.
[55] S. Sagnika, A. Pattanaik, B. S. P. Mishra, and S. K. Meher, ‘‘A review on
multi-lingual sentiment analysis by machine learning methods,’’ J. Eng.
Sci. Technol. Rev., vol. 13, no. 2, pp. 154–166, Apr. 2020.
[56] Z. Drus and H. Khalid, ‘‘Sentiment analysis in social media and its HUANG HUANG received the B.S. degree from
application: Systematic literature review,’’ Proc. Comput. Sci., vol. 161, Wuhan Textile University and the M.S. degree in
pp. 707–714, Jan. 2019. data science from the University of Malaya, Kuala
[57] J. A. Shathik and K. K. Prasad, ‘‘A literature review on application of Lumpur, in 2021.
sentiment analysis using machine learning techniques,’’ Int. J. Appl. Eng.
In college, his GPA is 3.24. From 2019 to 2021,
Manag. Lett., vol. 4, no. 2, pp. 41–67, 2020.
he was majoring in data science with the Uni-
[58] A. Ahmet and T. Abdullah, ‘‘Recent trends and advances in deep learning-
versity of Malaya. Since 2021, he has been a
based sentiment analysis,’’ in Deep Learning-Based Approaches for
Software Engineer with HSBC. He has been con-
Sentiment Analysis (Algorithms for Intelligent Systems), B. Agarwal,
R. Nayak, N. Mittal, and S. Patnaik, Eds. Singapore: Springer, 2020, doi: tributing to the advancement of state-of-the-art
10.1007/978-981-15-1216-2_2. models within the field of sentiment analysis, since
[59] S. Shayaa, N. I. Jaafar, S. Bahri, A. Sulaiman, P. S. Wai, Y. W. Chung, 2021. His research interests include natural language processing, with a
A. Z. Piprani, and M. A. Al-Garadi, ‘‘Sentiment analysis of big data: particular emphasis on sentiment analysis. He won the ‘‘Excellent Graduate’’
Methods, applications, and open challenges,’’ IEEE Access, vol. 6, Prize. He has participated in ‘‘Internet+’’ competition for two times. His
pp. 37807–37827, 2018. project ‘‘Online Monitoring System Based on Image Recognition and Multi-
[60] M. Umar, A. A. A. Bena, and B. Wadata, ‘‘Sentiment analysis techniques Sensors’’ won the Bronze Award from the third ‘‘Internet+’’ College Student
and application-survey and taxonomy,’’ J. Eng. Technol., vol. 4, no. 1, Innovation and Entrepreneurship Competition, from to October 2016 to
pp. 781–789, 2021. July 2017. Another project named ‘‘The system of Raspberry pi Sensor
[61] R. Baragash and H. Aldowah, ‘‘Sentiment analysis in higher education: Real-Time Motion Detection based on Cloud Computing’’ is a ‘‘National-
A systematic mapping review,’’ J. Phys., Conf. Ser., vol. 1860, no. 1, Level Project,’’ from December 2017 to September 2018, on college student
Mar. 2021, Art. no. 012002. innovation and entrepreneurship platform.

VOLUME 11, 2023 90381

 H. Huang et al.: SA in E-Commerce Platforms: A Review of Current Techniques and Future Directions

ADELEH ASEMI ZAVAREH (Member, IEEE) MUMTAZ BEGUM MUSTAFA (Member, IEEE)
received the B.Sc. degree from Ashrafi Esfahani received the B.Sc. degree in software engineering
University, Iran, in 2006, the M.Sc. degree from from Universiti Putra Malaysia (UPM), in 2002,
Pune University, India, in 2008, and the Ph.D. and the M.Sc. degree in software engineering and
degree from the Faculty of Computer Science and the Ph.D. degree in computer science from the
Information Technology, University of Malaya, University of Malaya (UM), in 2006 and 2012,
under the Prestigious MTCP Scholarship, in 2014. respectively. She is currently an Associate Profes-
She was conducting research in HCI and Human sor with UM. She has undertaken several Speech
Decision Making Simulation area, as a Postdoc- Synthesis research and holds grants from the Min-
toral Fellow with the Department of Software istry of Higher Education. She has established
Engineering, University of Malaya, in 2015. She has been a Lecturer and network with a number of the International Speech Research Laboratory in
an Assistant Professor with Esfahan University and the Safahan Institute Japan and Singapore. She supervises a group of Ph.D. and master’s students,
of Higher Education, Esfahan, Iran, respectively, since 2011. She received working on speech synthesis, speech recognition, and speech signal process-
training in modeling and simulation and attended workshops on big data and ing. She has published several papers in prestigious speech conferences and
R-programming with the University of Malaya. She supervised more than journals. She has published her work in many of prestigious international
30 postgraduate students and taught both undergraduate and postgraduate journals. Her research interests include emotional speech synthesis and
computer science subjects. She joined the Faculty of Computer Science speech assistive tools for disabled individuals. Her research and development
and Information Technology, University of Malaya, as a Visiting Research of the HMM-based Malay speech synthesis system has won the Most Presti-
Fellow, in 2019, and has been an Affiliated Senior Lecturer, since 2020. She gious Award (MPA) for Excellent Research 2012 from the MIMOS Berhad
is currently a Senior Lecturer with the Department of Software Engineering, the National Research and Development Centre in ICT and has won gold
University of Malaya. Her research interests include the intersection of medals for several national level competitions.
decision analysis including simulation/prediction/evaluation and artificial
intelligence. She has worked to improve the decision analysis by using
theories from disparate areas and formulate the evaluations of different
alternatives, such as engineering methods and systems to multi-criteria deci-
sion analysis. The breadth of her research agenda spans theoretical analysis
and systems development, and her research style is not only to develop
decision-making techniques and analyze their formal guarantees but also to
implement them for industries in the physical world. In addition, she has
forged interdisciplinary alliances with experts in agriculture, software engi-
neering, law, environmental science, and speech recognition. Since 2019, she
has been extended her area of research to the multi-criteria evaluation of big
data and the IoT techniques.

90382 VOLUME 11, 2023