This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.3011802, IEEE Access

Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000.
Digital Object Identifier 10.1109/ACCESS.2017.DOI

Aspect-based sentiment analysis using

a hybridized approach based on CNN
and GA
ADNAN ISHAQ1 , SOHAIL ASGHAR2 , AND SAIRA ANDLEEB GILLANI.3
1
Department of Computer Science, COMSATS University Islamabad, Islamabad PK 44000 (e-mail: [email protected])
2
Department of Computer Science, COMSATS University Islamabad, Islamabad PK 44000(e-mail: [email protected])
3
Department of Computer Science, Bahira Univeristy Karachi Campus, Karachi, PK (e-mail: [email protected])
Corresponding author: Adnan Ishaq (e-mail: [email protected]).

ABSTRACT Sentiment analysis is a computational analysis of unstructured textual data, used to assess
the person’s attitude from a piece of text. Aspect-based sentimental analysis defines the relationship among
opinion targets of a document and the polarity values corresponding to them. Since aspects are often implicit,
it is an extremely challenging task to spot them and calculate their respective polarity. In recent years, several
methods, strategies and improvements have been suggested to address these problems at various levels,
including corpus or lexicon-based approaches, term frequency and reverse document frequency approaches.
These strategies are quite effective when aspects are correlated with predefined groups and may struggle
when low-frequency aspects are involved. In terms of accuracy, heuristic approaches are stronger than
frequency and lexicon based approaches, however, they consume time due to different combinations of
features. This article presents an effective method to analyze the sentiments by integrating three operations:
(a) Mining semantic features (b) Transformation of extracted corpus using Word2vec (c) Implementation
of CNN for the mining of opinion. The hyperparameters of CNN are tuned with Genetic Algorithm
(GA). Experimental results revealed that the proposed technique gave better results than the state-of-the-
art techniques with 95.5% accuracy rate, 94.3% precision rate, 91.1% recall and 96.0% f-measure rate.

INDEX TERMS Aspect-based Sentiment Analysis; Convolutional Neural Network; Genetic Algorithm;

I. INTRODUCTION supervised learning, we train the machine by the mean of

The Natural Language Processing (NLP) always has a great data that is well labeled i.e. already tagged with the correct
importance in Sentiment Analysis (SA). SA is a subfield of answer. Then provide the machine with a new set of extended
NLP, could also be referred to as opinion mining. Opinion data, thus a supervised learning algorithm to analyze the
mining aims to identify and extract emotions attached to training data and give the correct result. In unsupervised
the piece of text. Nowadays, SA become much popular in learning, the machine is trained through the information that
online communities for processing social media data (blogs, is neither labeled nor classified, allows the algorithm to work
tweets, reviews, and forum discussion). SA evaluates the text, without any guidance.
either positive, neutral, or negative. It can acquire the level The development of network technology provides a new
of polarity (high, low, medium) of the text as well [1]. It way to communicate over user-created content (blogs, fo-
explores the emotions, mindset and attitude from gathering rums, social networks, websites reviews, e-commerce web-
feedback as reviews on numerous websites. sites, etc.). With this exponential growth, individuals and
Opinion mining could be carried out using two methods; organizations have a keen interest in data mining technology
the first is a machine learning method in which SA centered that uses this subjective information resource. SA is one of
on the rate of co-occurrence of words while the second the most frequently used research areas in computer science,
approach is lexicon-based, which comprises of ontologies, aimed at identifying and extracting user opinions [2].
lexicons, and semantic networks to find out the polarity The SA is interested in extracting the emotions conveyed
attendant by the text. The technique of machine learning is in the paragraphs of the text. Pioneering work on Aspect-
expanded over two parts, supervised and unsupervised. In Based Sentiment Analysis (ABSA) believes that SA research

VOLUME 4, 2016 1

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.3011802, IEEE Access

teristics of different categories i.e. (affirmative, negative, and

fair), and the final job is to identify words equivalent to the
characteristics.
Over the last few decades, the amount of interactive
content on the Internet has grown substantially worldwide,
and these interactive content websites allow users to publish
almost any content. The content you post contains your
judgments and emotional state about something. The content
is available on the web in the shape of reviews, blogs, and
literature. This data is present in the shape of the structure
and unstructured data. The structural data is one that is
comprised of transactional data, numeric data, and data that
is well organized. Such data is important for organizations to
take some serious decisions while unstructured data exist in
the shape of a text document, email, pdf file and reviews.
Mining this data to find people and their views who are
interested in a particular item or service is very significant
to many stakeholders, including governments and businesses.
Many of these tasks, such as big data and SA, pose many
challenges, like the complexity of processing data and the
complexity of processing unstructured text written in natural
language. The language that is used on the Internet is dif-
ferent for each individual, therefore a systematic approach
is needed for opinion containing data preprocessing and text
FIGURE 1: Proposed model classification.
ABSA take the reviews and study the aspect associated
with it. Every single aspect has a score associated with it
based on this score the polarity of the whole sentence could
can be performed at three levels: document, sentence, word,
be measured. After that, the reviews were classified as trained
or aspect. In document level, SA assumes how each docu-
and tested, and the model is now trained. The trained model
ment provides an opinion on a single entity. It is assumed
is able to find polarity in upcoming reviews. The working
that the document represents only one topic, but often this
design of the methodology is presented in Fig.1, which shows
is not the case. In a sentencing level analysis, each sentence
the training and testing units of the reviews that are passed
is treated as a different entity and it is considered that the
to the machine learning algorithm, and sentiment analysis is
sentence should hold merely one supposition. Subjective
carried out. The content available on the web in the shape of
characterization and grouping estimation are two undertak-
reviews, blogs and literature is growing on a daily basis. This
ings of the sentence-level analysis. Though, this is not very
data is present in the shape of structure and unstructured data.
handy in all circumstances. For example, a comment about
Structure data is one which is comprised transactional data,
a cell phone can have contradictory emotions in different
numeric data and data which is well organized. Such a data
parts of the cell phone. The reviewers praise the focus and
is important for organizations to take some serious decisions
image quality of the camera and sometimes criticize its
while unstructured data available in the form of text docu-
power consumption in the same sentence. To convey different
ment, email, pdf file and reviews. Aspect based sentimental
opinions about different aspects of the same product, you
analysis take the reviews and study the aspect associated
need to act on the ABSA method. In comparison with tra-
within it. Each single aspect has a score associated with it on
ditional SA, ABSA presents some challenges: (a) Link every
the basis of this score the polarity of whole sentence could be
portion of the text to a side (b) Determine the part of the
measured. After that, the reviews were classified as trained
text that describes the same side (e.g., a reviewer may be
and tested, and the model is now trained. The trained model
concerned about a battery and another reviewer discusses
is able to find polarity in upcoming reviews. The working
power consumption life, both are pointing to the identical
design of the methodology is presented in Fig. 1, Showing the
side) and (c) comparative sentences, etc. Though, most of
training and testing units of the study that are passed to the
the present work on ABSA is concentrated in English and
machine learning algorithm, sentimental analysis is carried
few in other languages [3]. The main function of ABSA is to
out. The language that is used on the internet is different for
enable investigations with component-based aspects, It may
each individual, therefore a systematic approach is needed for
have three operations, the first is to remove the questions that
opinion containing data preprocessing and text classification.
have been identified by the conclusion holder, The second
is the extremeness of the views that determine the charac-
2 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.3011802, IEEE Access

A. RESEARCH CONTRIBUTION emotions, but these idioms are not considered in the proposed
This paper puts forward an efficient procedure sentimental method of analysis. Considering the document metaphor,
analysis. satirical and ironic challenges the proposed method. In [9], a
method based on deep learning has been proposed to catego-
• A structural technique for extraction of features from
rize user opinions expressed in comments (called RNSA). It
unsaturated data of hotel, automobiles and movie re-
performs well with the provided dataset, however, execution
views has been defined
failed while processing a large dataset. In [10], author used
• Preprocessing of data.
a deep feedforward neural network with a global mean pool
• Semantic features have been extracted.
and a long-term short-term memory model with dense layers
• Convolutional neural network (CNN) is being used to
to identify subjective information from text documents and
extract opinions
solve emotional prediction problems at the sentence level. In
• The tuning of CNN parameters is done with a multi-
[11], an RNN-based method is proposed that uses a word
objective function using Genetic Algorithm (GA).
vector as the input acquired from the GloVe method. The
The order of the paper is as follow Section: 2 include model they suggest took into account the impact of market
literature review, Section: 3 contains Problem statement, Sec- trends, and other factors on sentiment predictions.
tion: 4 and Section: 5 presents the proposed methodology and Rezaeinia et al. [12], an improved word vector model are
proposed techniques, Section: 6 contain experimental design proposed by combining dictionary-based methods, part-of-
while the study of the results in Section: 7 and eventually speech tagging methods, word localization techniques, and
Section: 8 concludes the paper. word2vec / GloVe methods. To learn to embed emotion-
specific words, [13] describes how to integrate emotional
II. LITERATURE REVIEW information in the text. They also propose how to develop
This section reflects a comprehensive literature review of a neural network model to handle fine SA. SA of Hindi
the research work carried out in the SA field. Optimiz- comments, [14], learning model centered on CNN. They
ing data structures, optimizing query technology and opti- tried distinctive settings to modify input size and regulariza-
mizing parallelism can improves system’s efficiency. As a tion technology to determine CNN parameters Output size,
built model, it performs well enough for users competing dropout rate, period, activation function, etc. To improve the
in power, response rate, and expansibility. Al-molsmi et al. accuracy of SA of Arabic data CNN and LSTM is proposed
[4] submitted a commentary on cross-domain SA, as SA in [15]. This model uses soft voting in it, the predicted
has gained much consideration in current years. Therefore, category probabilities of the data are averaged over the two
technology, methods, detailed overview of methods can be CNNs. Then select the LSTM model and the category with
used for cross-domain SA by providing a comprehensive the highest average as the final selection Integrated model
literature introduction. Jianqiang et al. [5] also performed a prediction.
comparative study to analyze the preprocessing methods used Lee et al. [16], assessed the impact of text quality Com-
in the analysis of twitter. The accuracy and F1-measurement ments established on comment length, word count, and
of the classifier for twitter classification can be enhanced with readability. Emotion analysis task is performed on movies
the use of a suitable preprocessing method. dataset. Three models of deep learning family (simple CNN,
Wei et al. [6] has developed an ontological system for the LSTM, and RNN). The authors claim the dataset is short
analysis of the product review. A sentiment ontology tree is and easy to read higher accuracy compared to long and
designed to reflect the information of product functionality short length data sets readability. The CNN-based method
and the emotions associated with it in a hierarchical relation- proposed in [17], user behavior (personal characteristics
ship. Human-labeled data analysis ensures greater accuracy. and social activity) for emotion analysis. In [18], authors
This model can extract features automatically in future. It proposed a new model of neural network with two hidden
may reduce manual effort, while it could compromise a bit layers. The first layer shows sentence vectors that indicate
of accuracy Karagoz, P., Kama, [7], presented a framework sentences in short-term and long-term memory networks
for SA of Turkish informal texts using Frequency Based and the second layer encodes sentence relationships into a
Aspect Extraction with Sentiment Word Support (FBAE- document representation. It also suggests an improved way
SWS), and Web Search Based Aspect Extraction (WSBAE). to first clean the dataset and remove the emotionally less
This article emphasized on improving aspect extraction as polar sentences in the dataset. In [19], SA used a rule-based
an unsupervised method and detect the polarity depending method with the help of SentiWordNet and SVM for feature
on the aspect of the emotional word. It also provides a tool, extraction along with term frequency and inverse document
including a Graphical User Interface (GUI) for implementing frequency.
the proposed algorithm and visualizing the analysis results. In [20], author suggested a way to determine the sentiment
In [8], a new SA integration method based on POS and of reviews based on the hotel dataset. Hotel reviews are
n-gram has been proposed. It also considered semantics, preprocessed into a list of terms. First, the potential Dirichlet
emotional cues, and the order between words called EnSWF. Allocation (LDA) was determined Glossary; Semantic sim-
Some comments or opinions include idioms that help express ilarity then sorts the term list according to the topics gen-
VOLUME 4, 2016 3

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.3011802, IEEE Access

erated by potential Dirichlet assignments. (LDA) integrated dex utilized is a piece of SemEval 2015 and the outcome
into the five sides of the hotel. Next, when computing the underpins the Twitter SemEval 2015 benchmark and no
resemblance, the Frequency anti-clustering frequency (TF- high-quality highlights have been utilized in this structure.
ICF) method. Finally, classify consumer sentiment (satis- The F1score surveyed was 64.85%, Future scope of this
fied or not) with word embedding and long-term memory structure is that the entire framework would be recreated
(LSTM). Kontopoulos et al. [21] presented the execution through utilizing twitter-based organizations that could make
of a real ontology-based deployment methodology for more improvement in effectiveness of the model under exchange.
effectual twitter’s tweet SA. The suggested model utilizes Ouyang et al. In [34], an approach was proposed that in-
the ontology methodology to analyze the twitter in a more cluded Word2Vec and CNN, and three sets of convolutional
effective way not just by giving a complete tweet score, but layer and max pooling layer were used in this project. This
also by analyzing any hidden dimension of tweets, instead, is the first hypothesis that word2vec and CNN use statements
provide scores for a certain tweet which makes the process using a seven-layer structure model, a linear parameter recti-
of feature engineering more reliable to achieve a higher fier with standardized functionality, and crimping techniques.
degree of accuracy. Ciric et al. [22] suggested a model for Publicly accessible movie review data uses five different tags.
evaluating the tweet’s sentiment. Several machine learning Negative, some negative, neutral, some positive, and positive.
models are used in the proposed work to conduct opinion System test accuracy is 45.4
mining, and their assembly is carried out in order to achieve Jindal et al. [35] built the image system, emotion recog-
better outcomes. The gathered outcomes are equated with nition framework works with deep and complex neural net-
various approaches for thorough analysis. The tabular view work. This methodology has been pulled together with huge
of literature review is shown in Table 1. data for the identification and transition of objects. The data
Bahraini et al. [23] suggested a hybrid model for the study set utilized was the image called Flicker. In [36] differen-
of twitter tweets by target-based sentiment. A hybrid solution tial neural system model has been proposed. It combines a
is shown to outperform other strategies as well as demon- convolutional neural network with treebank data to guide
strated better results with different features and functions. the conclusion checking task. The model provides various
Freitas et al. [24] recommended a model for extraction of elements such as grammatical information and structural
features through ontology technique, and data sources are information data, and its effect is superior to any other single-
data sets for a movie review and hotel reviews. The analysis aspect model. Authors used Static word embedding , and
of sentiment is remarkably effective with higher efficiency static word embedding can be designed by showing which
and accuracy in their experiments. Zhao et al [25] the au- terms work to express the rating in a better way. We can also
thors proposed a Chinese-language approach that achieved provide prefix and suffix data for sensory extraction. Bouazizi
better results than previous models. It has acquired greater et al. [37] built up a framework for a feeling examination of
accuracy. In [26], author introduced a model in which the twitter tweets with a specific capacity to identify mocking
Iris General Electronics tweets were classified in 3-class explanations that improve the effectiveness of conclusion
sentiment. Using a supervised learning method and lexicon- mining. The capacity to identify wry tweets results at a more
based subjectivity score, 61.6 percent accuracy was achieved. significant level of exactness. Abbasi et al. [38] recommends
Sam et al. [27] simplified models has been developed to a model using multilingual SA to share data over the internet.
evaluate customer comments about electronic products on so- SA in the English and Arabic dialects is achieved by using
cial networking sites. The keyword extraction and electronic style and syntax tools to capture specific element parts. The
product ontology creation is about understanding consumer execution results show the accuracy is above 91%.
behavior online. Ravi et al. [28] and Yadollahi et al. [29] Valdivia et al. [39] introduced a model for the presentation
tended to an investigation of the systems utilized for opinion of a TripAdvisor SA matching between user emotions and
mining. Clavel et al. [30] additionally utilized SA in human- automated emotion detection algorithms. Che et al. [40]
agent interaction since conclusion digging was once in a propose an ABSA method that automatically compresses
while utilized for human-agent interaction prior. Dragoni et emotional sentences through a sentence compression random
al. [31] creating a new ontology of common sense for an field. This significantly improves ABSA performance. Bui
analysis of sentiments. Kim [32] used CNN to distribute et al. [41] performed a cancer survivor’s network study of
exam progress. This course was trained on sentence-level temporal connection interpretation of changes in emotions.
classification tasks based on pretrained word vectors. A ba- It is analyzed by presenting a new structure for the cancer
sic CNN technique with few hyperparameters adjustments survivors machine learning has been used to train sentiment
and static vectors delivers unimaginable results on various classifiers across positions that are manually labeled with
benchmarks. This technique is applied using a single layer sentiment labels to identify sentiment post such as negative
convolutional design. or positive. Wu et al. [42] have implemented a SA approach
Stojanvaski et al. [33] recommends a model having a to decision-making in the online stock platform, which pro-
profound convolutionary neural system. Pre-prepared word motes investor decision-making and the understanding of eq-
vector inserting is accomplished in this work by actualizing uity companies. An analysis of the results shows that investor
unaided learning in huge companies. The informational in- sentiment has a greater impact on value stocks. Comparative
4 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.3011802, IEEE Access

TABLE 1: Sentiment analysis comparison

Article Authors Approach Accuracy
[7] Karagoz, P., Kama, B., Oz- FBAE-SWS,WSBAE 85.0%
turk, M., Toroslu,
[8] Khan, J., Alam, A., Lee, Y. EnSWF 87.62% .
K.., & Hussain, (2019).
[9] Abdi, A., Shamsuddin, S. RNSA 86.75% .
M., Hasan,
[11] Spyromitros, E., k-nearest neighbor (kNN), BRkNN 68.8%
Tsoumakas, G., and
Vlahavas
[14] . Akhtar M., Gupta, Conditional Random Field (CRF), Maxi- 77.33 %
D., Ekbal, C .,and mum Entropy (ME) and Support Vector Ma-
Bhattacharyya, P chine (SVM)
[16] M., Deokar, A. V., Janze, Term frequency – Inverse document fre- 80.80%
C quency (TF-IDF) matrix
[18] Madjarov, G., Džeroski, Predictive Clustering Trees (PCT) with Ran- NA
S., Gjorgjevikj, D., & dom forests (RF), and Hierarchy of multi-
Džeroski, S and label classifiers (HOMER)
[21] Kontopoulos E., Dergiades Ontology based approach NA
T,. Berberidis C
[6] Wei W., Gulla JA HL SOT (sentiment ontology tree) approach, Hi- NA
erarchical Learning,
[44] Mouthami K., Nirmala D., Fuzzy context classification algorithm that NA
& Murali B includes parts of speech tag.
[34] Ouyang X., Pan Z., Hua C. For the evaluation of sentences with 45.5%
L., & Liu L word2Vec a 7-level CNN model has been
used
[35] Stuti Jindal and Sanjay A predictive system of image sentiments 53.5%
Singh is generated with a pre-trained CNN based
on large scaling data and object recognition
with a pre-trained CNN based on large scal-
ing data
[36] Yang T., Li Y., Pan Q.,& CNN using tree bank information Binary classification accuracy
Lantian G 94.7% accuracy), Fine grained
Accuracy 49.9%
[33] Stojanovski D., Strezoski CNN having multiples filters, two fully inte- F1 score 64.85%
G., Madjarov I., &Dim- grated layers as well as a softmax layer with
itrovski I different window sizes.

studies with direct evaluation approaches show the reliability of the most important tasks in ABSA [21].
of the method. In Table 2 shows a comparison of the latest
technologies. IV. PROPOSED MODEL
The steps of our proposed work are data collection, prepro-
III. PROBLEM STATEMENT cessing, semantic feature extraction, word2vec representa-
ABSA aims to predict the polarity of a specific document tion, and CNN implementation stages, as shown in Fig. 2.
for a particular aspect of an entity. The neural network
architecture successfully predicted the overall polarity of the
sentence, however, SA of certain aspects is an open question A. DATA COLLECTION
[14]. Although typical SA focus on predicting positive and Web scraping technique is used for review collection; hotel
negative polarities of a particular sentence. This task works reviews collected from “https://webhose.io/ ”, automobiles
when the specified text has only one aspect and polarity. A reviews are collected from “https://www.cvedia.com/” while
more common and complex task is to predict the aspects movie reviews are fetched from ” https://seedmelab.org/”.
mentioned in a sentence and the emotions associated with The raw view of dataset is shown in Fig. 3 The reviews are
each. In ABSA, this is similar to aspect and document, where splinted into positive and negative categories. This simplifies
the relationship between each word in the document is com- the task and makes it easier to distinguish various aspects
pared to the aspect vector. The relationship between aspect of positive and negative views. In the proposed model, three
and language lacks the ability to learn expressions [18]. types of data set (hotel, automobiles, and movies) are taken
Most existing opinion mining methods are based on text- using web scraping and then preprocessing is performed to
level analysis and can only detect well-expressed opinions. get valuable data from the data set. Data is provided to
The goal of ABSA is to identify aspects of an entity and the classifier, semantic features are obtained from specific
sentiments stated on each aspect. Extracting the aspect terms domains, and the total score is determined based on the
and presenting opinions from user-generated content is one initial features. Word2vec is designed for corpora processed
VOLUME 4, 2016 5

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.3011802, IEEE Access

classification is carried out and CNN hyperparameters were

tuned using GA to achieve better model simplification. The
proposed model for the sentiment analysis model is shown in
the Fig. 1

A. FEATURE EXTRACTION
Words are converted into feature vectors in the first layer
of the network to extract the semantics and morphological
information about words V words represent word vocabulary
Vchar represent character vocabulary. Given sentence contain
n words (w1, w2, w3 . . . . . . . . . ..wn ) And changing Wn
(every word) into vector, according to Eq. 1.
Vn = rword , rwchar

(1)
While rword words is word level embedding and rwchar is
character level embedding. To capture semantic and syn-
tactic information word-level embedding is intended, while
FIGURE 2: Phases of aspect based sentiment analysis character-level embedding is intended to capture shape and
morphological information.

through unsupervised neural networks. Conclusively, a vec- B. WORD LEVEL EMBEDDING

tor type of the preowned corpus is trained by the CNN Word level embedding is encrypted in an embedding ma-
alongside GA. trix Wword εRdword ∗ Vword by using vectors of column.
The word level embedding coincides to individual column
B. PREPROCESSING Wiword and Rdword through which ith vocabulary word is
In this step, raw data is transformed into structured informa- produced. Word W is word level embedded by the product of
tion data. Empty rows, empty cells may present in the data matrix vectors as shown in Eq. 2.
collected through scrappy, panda library is used to clean the
Rword = Vw Wword (2)
data and keep the useful data only as shown in Fig. 4. After
analyzing the reviews manually, the provided reviews are Where Vw represent the vector size of |v-word|. At index
used for labels during implementation of VADER. VADER W have 1 value 0 at o their positions Wword matrix is a
is an analysis tool specifically adapted to extract emotions parameter for learning and dword is embedding word level
expressed in text. VADER offers a set of lexical features (e.g. size. In this paper, word level embedding is performed by
words) that are generally classified as positive or negative word2vec.
subjective to their semantic orientation. VADER was found to
be quite effective when it comes to social media texts, movies C. WORD2VEC
and product reviews. VADER is not bound only to calculate Word2vec is part of open source tools that is provided by
the positive and negative score, however, it also tells us about Google under the Apache License 2.0 in 2013. This is a
the intensity, i.e., how positive or negative a sentiment is. It very effective tool that can extract the features of a particular
works remarkably well on text type of social media, but is domain without human intervention. In addition, it works
easily generalized to several domains. It does not include well whether for too small text size or a single word. By
training data, but it is designed on a generalizable, high- providing a huge corpus context and using word2vec, you
valence, human created standard lexicon. It is reasonably fast can create words with the right sense and run faster on large
that can be used online for streaming data and do not suffer datasets.
seriously from a speed-performance compromise. Word meaning is the ultimate perspective of deep learning,
and using word2vec to categorize larger entities can fully
V. PROPOSED TECHNIQUE satisfy the meaning of words. In the proposed method, the
The data used in our proposed model is collected through dataset is trained on vectors. This model contains 3 million
web scrapping using python pursed by its preprocessing. words mixed with 3 million words and phrases. By running
Thereafter, features are extracted from data and supplied to such a large corpus, it is possible to obtain accurate rela-
the classifier models. The semantics features are separated tionships. Fig. 5 shows a similar relationship between words
from the specified domain and total scored is determined using the word2vec transform and relationship between dis-
based on initial features extracted using VADER. After that similar words in the dataset. The width of edge indicates
Word2vec is design to process the given corpus by utilizing similarity, and nodes represent words. Words with the same
unsupervised neural network and at the end vector form of emotion label have the same vector, so you can easily specify
processed corpus is used to train through CNN classifier. The word similarity.
6 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.3011802, IEEE Access

FIGURE 3: Unprocessed reviews

vector operation on every character k char from the list of

characters. xm εrchar Rdchar as in Eq. 4. That is the embed-
ding character contain its left neighbor (rchar − 1)/2 and
right neighbors (rchar − 1)/2
  >
X n = kn−(rchar−1)/2
char char
/2 · · · kn+(rchar−1)/2 (4)

The vector k char have J th element calculated through con-

volutional layer in the CNN that is the w which is the
embedding character level, according to Eq. 5.
 wchar 
= max1<n<N W0 Xn + a0 j
 
K j
(5)

latterly, local features are extracted utilization same matrix

around every window character in the given word.

E. ANALYSIS AT SENTENCE LEVEL


A sentence y contains m words w1 , w2 , . . . wn which
is then changed over to the word level joints and embed-
ding level characters v 1 , v 2 , . . . v n Next stage consist of
FIGURE 4: Data preprocessing representation of sentence level extraction while extracting
features at sentence level we may have following issues.
1) Distinct size of sentence.
D. CHARACTER LEVEL EMBEDDING 2) The sentence may hold significant information = at any
All words employed important methods of extracting shape position.
and morphological information from the words and thought
These problems are solved by using CNN to calculate the
that these words selected important features of the classi-
remaining feature vectors of wide sentences. CNN always
fication stage. For example, sentiment analysis of Twitter
works along lines when used for character-level feature ex-
data uses various hashtags to display important information,
traction. Create local features for each word, use the max op-
such as # ihateit, and various adverb information, ending
eration to create a fixed-size feature vector for the sentence,
with a suffix, such as angry. Convolution generates local blur
and then link. According to Eq. 6,Matrix  vector operation
around the words around each character and connects them
performed on each word in the sequence u1 , u2 , . . . un
for character-level embedding using the max operation. The
specified word is a composition of m characters. Xn = {uc . . . ud } (6)
{k1 , k2 , . . . , km } every character km is converted in to nod nod
an embedded character Vm
|char|
, according to Eq. 3. Embed- where the c = n − u 2 −1 , d = n + u 2 −1 . convolutional
1/4
ding metrics have embedded character encoded in column layer computed the j th element whom vector rst ∈ Rc as
vectors. W char ε v char . Each character k in Eq. 7.
 st 
r j = 1 < m < M max wx1m + a1 j
 
Rchar = vk Wchar (3) (7)
1

V k represent the size of v char . v k have value 1 and index at
 char where w1 ∈ Rcu × dword + c0u k word . At last rxst vector
number 0 at other position. r1 , r2char . . . . . . . . . . . . rnchar that contain feature vector global comparative with x sen-
is convolutional layer input. The convolutional apply matrix tence is handled through two layers of neural network. To
VOLUME 4, 2016 7

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.3011802, IEEE Access

(a) Hotel features. (b) Movie features.

FIGURE 5: Extracted features from hotel and movie dataset

extricate one or more representation level and every senti-

mental level t ε T score is then calculated, according to Eq.
8.
S(x) = w3 h w2 rxst + a2 a3


(8)
Where learning parameter is given by w2 εRhu∗c1u and vec-
tors are given by w3 εR|T |∗Hu

F. NEGATIVE TRAINING
The system is trained over a training set A, to limit the
negative probability. A sentence is passed to the system con-
tain parameter set θ and each sentiment label is computing
τ ε T. These scores are converted into giving the label with
conditional probability distribution plus network parameter
set θ and then softmax operation is applied over τ ε T score,
as in Eq. 9.
eSθ (x)i
P (τ |x, θ) = P Sθ (x)i
(9)
∀i∈T e
The stochastic gradient descent is applied to limit the nega-
tive log gradient as in Eq. 10.
X FIGURE 6: Convolutional neural network
θ −→ − log P (v|x, θ) (10)
(u,v)∈A

(U, v) is the sentence compared to corpus training, and V degree of features utilizing filter with size 2. The fundamental
suppresses the corresponding label. aim is to provide sufficient range of features to grab all prob-
able descriptor of the text. Maxpool shall have the highest
G. CONVOLUTIONAL NEURAL NETWORK
output value of the vector on applying the filter. It picks a
The structure of CNN can be seen in Fig. 6, this consists strongest expression aspect in the extracted function from
of the separate layer stack which converts the volume of the output and nothing to do with the length of the word.
the input into target output by differential function. The Every sample is represented as n*1 where n represents the
constituent layers of CNN are length of dimension filter. This filter is practice as a drafting
1) Convolution layer window, for instants 3*1 filter on a sentence. I like the car
2) Max-policy layer very much! would yield (I like this, this car very, car very
3) Rely layer much, very much!) sentence splits up to equal length ahead of
4) Back propagation layer embedding hence all the filter does not result in the identical
The intention is to use sufficient filters (in this case, 128) to dimensions outcome. The area size (2, 3, 4 ) is identical to
catch enough features in a particular sentence. In classifica- 2, 3, 4 –G word and the first filter in this trigram will give
tion of images, different filters integrate different attributes different weight esteems to different words in trigrams. It
for example edges, color density at various spots, transforma- suggest that higher weight are allotted to the first index (0-
tion of colors etc. The problem of text classification stretches index) and lower to the second hence 128 filter will allocate
the similar idea to catch features for example "like" mean respective weight that will be trained to optional weight
positive rather than similarity, "very much" communicate the value after sometimes for precise forecasting. For a particular
8 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.3011802, IEEE Access

sentence, a sentimental label is formed by scoring a sentence,

by evaluating the order of words in a sentence as input and
by passing through layers that extricate its feature with a
high degree of complexity. Extraction of the features can be
executed on lexical and character level. The distinctiveness
of network design is to incorporate two convolutionary layers
that allow to handle any size of phrases and words.
FIGURE 8: Initial population
H. GENETIC ALGORITHM
GA is a technique to solve problems of optimization centered
on natural selection. GA is quite useful to incorporate when score of objective function is converted into a value within
you don’t know the best ranges and dependencies of different selected range as in Eq. 11 is happened by fitness scaling.
CNN parameters. As CNN defines a hyperplane between the Gleam + Gvalidation
data as shown in Fig. 6. This technique reflects the procedure F = (11)
2
of natural selection, where the best regenerative individuals
are chosen to create the next generation of offspring. The Selection: This phase is used to select the most suit-
natural selection process initiates with selecting the best able entities and let them forward their genes to the next
individual from the population. The offspring adopted the generation. It selects two sets of individuals (parents) based
characteristics from their parents and delivered to next gen- on their health score. The higher the physique, the more
eration. If the parent is in good shape then there are more opportunities for breeding. This selection uses the scaled
chances that offspring are better and can survive more than fitness values to select the next generation of parents, and
their parent. This process is continually repeated, eventually typically applies higher selection probabilities to individuals
finding the optimal generation for the individual. The overall with higher scaled values, as in Eq. 12.
steps are shown in Fig. 7. This concept can be used to find an f(x)
g(x) = P (12)
x∈p k

the number of copies of each feature t x0 in p(k) , that is

chosen number of copies per function tX0 in p(k) , It’s selected
(k)
for pn , is offered by the fractional part of the e(x) If the
number of features selected in such a way is smaller than m
(k)
(the usual case), then choose the leftover features for pn
FIGURE 7: Genetic algorithm steps using fractional parts of e(x) from the highest values down.
In general, this technique is designed to remove less adaptive
appropriate solution for problem. GA selects a best solution features and repeat more adaptive features.
from the set of solutions for specified problem. The genetic
CR(y) = n+
y /my (13)
algorithm considers five stages.
In Eq. 13, with n+ y denoting the number of patterns of class y
1) Initial population
that are classified as member of class y (correct) my denoting
2) Fitness function
the number of all patterns belonging to class y.
3) Selection
Crossover: Crossover is very important stage of genetic
4) Crossover
algorithm. A random exchange point is selected from the
5) Mutation
gene for each parental pair that is mated. The crossover
Initial Population: This process is initiated with the group helps the genetic algorithm to derive and recombine the most
of individuals called population. Everyone is the solution of suitable features from different individuals to a potentially
the problem that you need to eliminate. Human characteris- superior child, as shown in Eq. 14.
tics are consist of parameters (variables) called genes. Gene
child = parent2 + Rcros × (parent1 − parent2) (14)
interface with a string to frame a chromosome (solution). In
genetic algorithms, a person’s gene set is represented by a For example, consider an crossover point at 3, as shown in
string and a letter as shown in Fig. 8. Binary values (strings of Fig. 9. Offspring are created by exchanging genes between
1s and 0s) are typically used. It might be said that we encode parents until the crossover point is reached as in Fig. 10. New
a gene on a chromosome. offspring are added to the population as shown in Fig. 11.
Fitness Function: Fitness functions determine a person’s Mutation: In the formation of some new offspring, some
health (the capability of one to compete with others). It pro- of those genes can be mutated with a low probability. It
vides a fitness score to everyone. The probability is calculated means that few bits in the bit string can be flipped as shown
by fitness score to select person for breeding. The raw fitness in Fig. 12. In addition to mating offspring, genetic algorithms
VOLUME 4, 2016 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.3011802, IEEE Access

alization of the proposed model, 20 iterations are considered

in the CNN classification process. Besides, various models
are implemented and outcomes are compared to validate the
strength of the recommended model.
1) True positive (TP): the values of class that is correctly
categorized.
2) True negative (TN): correctly categorized as uninter-
esting.
3) False negative (FN): misclassified as a category of no
FIGURE 9: Selection point for crossover interest.
4) False positive (FP): class of interest is wrongly classi-
fied.
The formulas of performance evaluation metrics are shown
in Table 2.

VII. RESULTS AND DISCUSSION

It is observed in experimental observations that the extraction
FIGURE 10: Exchange of genes between parents of semantic features produces exceptionally refined data.
Information gathered by examining the particular domain re-
views decreases the false negative and false positive rates that
produce better accuracy. Table 6 shows the comparison of
precision, accuracy, and recall of different classifiers such as
SVM, maximum entropy, random forest, stabilized discrim-
inant analysis, decision tree and generalized linear model.
FIGURE 11: New offspring
CNN has been found to produce significant improvement of
91.6% , 93.4% , 92.2% , 89.23% in precision, accuracy, recall
also generate mutant offspring by randomly mutating a single and f1 measure, respectively.
parent of the current generation. Because mutations con-
tribute to population diversity, the algorithm is more likely model accuracy

to generate fitness values that are better suited to individuals. 100

train

test

80
accuracy

60

40

FIGURE 12: Mutation 20

Mutations occur to maintain diversity within a population 0.0 2.5 5.0 7.5 10.0

Iterations
12.5 15.0 17.5 20.0

and prevent premature fusion.

Termination: If the population converges, the algorithm FIGURE 13: Accuracy vs iterations
terminates (no descendants are significantly different from
the previous generation). GA are said to provide a set of The suggested approach has been contrasted with the
solutions to our problem. SVM, leading to high accuracy rate and f-score value for
positive as well as negative labels as shown in Table 4. The
VI. EXPERIMENTAL DESIGN accuracy and loss rate of the suggested method are shown
Implementation of the proposed framework is done on in Fig. 13 and Fig. 14. The graph shows that as the number
the system having specifications; Intel Core i3-4005 CPU of training iterations and tests increases, the percentage of
@1.70GHZ and 4GB RAM.The simulation runs on software accuracy increases, and as the number of iterations decreases,
called Anaconda Spyder and a python language environment. the percentage of loss decreases.
Software packages used are Keras v2.3.1, Tensorflow v2.0, Fig. 15 graphically illustrates the performance of the inte-
pandas v1.0.3 And numpy v1. The various indicators for grated method on a hotel dataset. The accuracy of ensemble
model validation are Precision, Accuracy, Recall, and F1- approach for hotel datasets is higher than the comparison
Score as defined in Table 2. For the validation and gener- techniques. The CNN-GA hybrid model’s overall result is
10 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.3011802, IEEE Access

TABLE 2: Performance Metrics Formula’s

Metric Definition
Accuracy (TruePositve+TrueNegative)/(TruePositve+TrueNegative+FalsePositive+FalseNegative)
Precision TruePositve/(TruePositve+FalsePostive)
Recall TruePositve/(TruePositve+FalseNegative)
F1-Score (precision x 2 x recall ) / ( precision + recall )

TABLE 3: Detailed results on Automobiles Dataset

Parameters SVM DecTree LDA LinMod RF LogReg Proposed
Accuracy 85.5 67 65 78.5 83 80 93
95% CI (0.84, 0.87) (0.65, 0.68) (0.64, 0.66) (0.75, 0.80) (0.82, 0.85) (0.78, 0.82) (0.90, 0.95)
Information rate 0.6 0.62 0.6257 0.635 0.624 0.625 0.629
P0-value 2.20E-16 2.20E-16 2.20E-16 2.20E-16 2.20E-16 2.20E-16 2.20E-16
Kappa 0.7112 0.6828 0.4429 0.637 0.6712 0.7179 0.7321
Mcnemar’s 0.4107 0.002628 0.00992 0.7273 0.0953 0.0013 0.5203
Pos Pred value 82.8 70.6 84.8 74 80.8 85.6 88.4

TABLE 4: Performance Analysis on Hotel Dataset

model loss
Models Precision Accuracy Recall Fmeasure
train
Proposed 95.5 94.3 91 96.6
2.0 test
DT 72.71 67.0 74.52 75.44
SVM 91.5 92.3 90.2 92.34
1.5
LDA 78.53 65.0 62.43 76.75
LM 82.36 78.5 84.85 86.10
RF 86.3 83.0 84.8 85.69
loss

1.0
LG 82.36 80.5 82.85 81.10

0.5

100
0.0

0.0 2.5 5.0 7.5 10.0 12.5 15.0 17.5

Iterations
80

FIGURE 14: Loss vs iterations 60 Accuracy

Precision
Recall
40 F-Measure

20

0
RF CNN DT LM LDA SVM LG
Movies Dataset

FIGURE 16: Performance on movies dataset

Fig. 17 graphically illustrates the performance of the

integrated method on automobiles dataset. Table 6 clearly
demonstrates the performance of the suggested method for
automobile data. Evaluated with the prior art, the proposed
FIGURE 15: Performance on hotel dataset method has the highest precision, recall, accuracy and f-
metric. The performance of the integrated method is 6%
better than SVM, 21% better than decision trees, 15% better
stronger than other state-of-the-art techniques, can be seen than LDA, and 11% better than RF, and 10% better than LG.
in Table. 4. The movie review dataset is used as the standard bench-
Table 4 clearly shows the performance of the integrated mark dataset for IMDB. In Fig. 16, it can be observed that
approach for hotel data. Compared with the prior art, the the outcomes of the integrated method (GA with CNN) are
proposed method has the highest precision, recall, accuracy much better than the benchmark algorithm (RF, DT, LM,
and f-measure. The performance of integrated method is 4% LDA, SVM, and LG). GA based integrated methods provide
better than SVM, 23% better than decision trees, 17% better superior performance in terms of accuracy, precision, recall,
than LDA, and 10% better than RF, and 12% better than LG. and f measurements. The results are shown graphically in
VOLUME 4, 2016 11

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.3011802, IEEE Access

TABLE 5: Performance evaluation on Movies Dataset TABLE 7: Comparison of proposed model with SVM model
Models Precision Accuracy Recall Fmeasure Parameters Proposed Model SVM
Proposed 91.5 95.2 92.3 94.0 Labels 2 2
DT 75 79 85 82 Data points 200 200
SVM 89.4 92.5 88.8 90.3 Accuracy 88.4 84.8
LDA 89 80 75 85
Label NEG NEG
LM 87 86 88 90
RF 90.8 87.5 84.8 88.0 F-score 89.23% 87.69%
LG 87 88 86 85 TP: 185 FP: 6 TN: 9 TP: 168 FP: 4 TN: 23
FN: 0 FN: 5
Label POS POS
F-score 78% 64%
Fig. 16 and in tabular in Table 5. In Table 5, from an accuracy TP: 185 FP: 6 TN: 9 TP: 166 FP: 6 TN: 25
FN: 0 FN: 3

100
TABLE 8: Performance Evaluation w.r.t iterations on Movie
80 and Hotel Datasets
Iteration loss Accuracy Time Elapsed Base Learning Rate
60 Accuracy 0 9.5881 38.91% 1s 1ms 1.00E-04
Precision 1 0.4045 80.78% 0s 703us 1.00E-04
Recall 2 0.0755 92.34% 0s 586us 1.00E-04
40 F-Measure
3 0.0833 92.34% 0s 622us 1.00E-04
4 0.0851 91.56% 0s 563us 1.00E-04
20 5 0.0692 92.19% 0s 547us 1.00E-04
6 0.0702 92.03% 0s 531us 1.00E-04
7 0.0662 92.34% 0s 531us 1.00E-04
0
RF CNN DT LM LDA SVM LG 8 0.0654 92.03% 0s 531us 1.00E-04
Automobiles Dataset 9 0.0657 92.19% 0s 539us 1.00E-04
10 0.0649 92.19% 0s 531us 1.00E-04
FIGURE 17: Performance on Automobiles Dataset

perspective, the performance of the integrated method is 2% nearly all the provided classifiers for all test parameters. It
better than SVM and 6% better than RF. The accuracy of the can therefore be concluded that the recommended approach
integrated method is 91.5%, while the accuracy of RF and is 93.0% more accurate than all six comparator classifiers.
SVM is 90.8% and 89.4%. In addition, the performance of The result comparison table supports the Ensemble meth-
integrated approach is significantly better to other existing ods given. The comparison results clearly indicate that the
technologies in recall and f-measure. suggested approach produces better results in terms of accu-
racy, precision and f-measure. For certain datasets based on
TABLE 6: Performance Evaluation on Automobiles Dataset positive and negative ratings based on experimental settings,
Models Precision Accuracy Recall Fmeasure SVM has the highest accuracy. .
Proposed 91.6% 93.4% 92.2% 89.23%
DT 70% 70.6% 81.20% 77.59% VIII. CONCLUSION
SVM 85.8% 84.8% 89.9% 87.33%
This article presents an effective classification method for
LDA 76.75% 74.0% 74.15% 81.90%
LM 81.8% 80.8% 88% 87.33% sentiment analysis using Convolutional neural network and
RF 82.3% 85.6% 84.8% 85.33% Genetic Algorithm. Semantic features are mined, and then,
LG 81.8% 82.8% 86% 82.33% several models are trained together with proposed CNN
based ensembler, SVM, maximum entropy, random forest,
Fig. 17 shows the performance outcomes of the latest tech- stabilized discriminant analysis, decision tree, generalized
nologies and integration methods for automobiles datasets. linear model. Using data collected through CNN execution
The proposed method has excellent performance on automo- and analysis, domain-specific reviews reduce false positives
biles dataset in terms of accuracy, precision, recall, and f- and false positives and improve accuracy. we adjust the CNN
measure. The results are described in detail in Table 6. Table hyperparameters using a genetic algorithm to get optimal
7 shows the performance index values in percentage format, values. The experimental results show that the recommended
showing that the proposed method performs well with all method outperforms all other recent methods, with 95.5%,
metrics as compared to SVM. Table 8 shows the detailed 94.3%, 91.1%, and 96.6% for accuracy, precision, recall, and
results for the different iteration levels and the different f-measurement, respectively. Future strategies are to integrate
iterations for accuracy, loss, progress, and Base Learning parallel computing to speed up computation and explore
Rate. For different parameters, we compare the outcomes of metaheuristic-related features. A preferred work is to have
the recommended classifier with all six other classifiers in a web-based ontology framework automation to incorporate
Table 3 and conclude that the proposed approach outperforms sentiment analysis on social sites.
12 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.3011802, IEEE Access

COMPLIANCE WITH ETHICAL STANDARDS [23] Bahrainian S-A, Dengel A (2013) Sentiment analysis and summarization
of twitter data. In: 2013 IEEE 16th international conference on computational
Conflict of interest: Their is no conflict of interest involved
science and engineering (CSE). IEEE, pp 227–234.
in this research. [24] Freitas LA, Vieira R (2013) Ontology based feature level opinion mining
for Portuguese reviews. In: Proceedings of the 22nd international conference
on World Wide Web. ACM, pp 367–370.
REFERENCES [25] Zhao Y, Niu K, He Z, Lin J, Wang X (2013) Text sentiment analysis
[1] Kumar, Ravindra, Husanbir Singh Pannu, and Avleen Kaur Malhi. "Aspect- algorithm optimization and platform development in social network. In: 2013
based sentiment analysis using deep networks and stochastic optimization." sixth international symposium on computational intelligence and design
Neural Computing and Applications (2019): 1-15 (ISCID), vol 1. IEEE, pp 410–413.
[2] Cambria, Erik, et al. "Sentiment analysis is a big suitcase." IEEE Intelligent [26] Bakliwal A, Foster J, van der Puil J, O’Brien R, Tounsi L, Hughes M
Systems 32.6 (2017): 74-80 (2013) Sentiment analysis of political tweets: towards an accurate classifier.
[3] Al-Smadi, Mohammad, et al. "Human annotated arabic dataset of book Association for Computational Linguistics, pp 49–58.
reviews for aspect based sentiment analysis." 2015 3rd International Con- [27] Sam KM, Chatwin C (2013) Ontology-based sentiment analysis model
ference on Future Internet of Things and Cloud. IEEE, 2015. of customer reviews for electronic products. Int J e-Educ e-Bus e-Manag
[4] Al-Moslmi T, Omar N, Abdullah S, Albared M (2017) Approaches to e-Learn 3(6):477.
cross-domain sentiment analysis: a systematic literature review. IEEE Access [28] Ravi K, Ravi V (2015) A survey on opinion mining and sentiment analysis:
5:16173–16192 tasks, approaches and applications. Knowl Based Syst 89:14–46.
[5] Jianqiang Z, Xiaolin G (2017) Comparison research on text preprocessing [29] Yadollahi A, Shahraki AG, Zaiane OR (2017) Current state of text senti-
methods on twitter sentiment analysis. IEEE Access 5:2870–2879. ment analysis from opinion to emotion mining. ACM Comput Surv: CSUR
[6] Wei W, Gulla JA (2010) Sentiment learning on product reviews via sen- 50(2):25
timent ontology tree. In: Proceedings of the 48th annual meeting of the [30] Clavel C, Callejas Z (2016) Sentiment analysis: from opinion mining to
association for computational linguistics. Association for Computational human-agent interaction. IEEE Trans Affect Comput 7(1):74–93.
Linguistics, pp 404–413. [31] Dragoni M, Poria S, Cambria E (2018) Ontosenticnet: a commonsense
ontology for sentiment analysis. IEEE Intell Syst 33(3):77–85.
[7] Karagoz, P., Kama, B., Ozturk, M., Toroslu, I. H., & Canturk, D. (2019).
[32] Kim Y (2014) Convolutional neural networks for sentence classification.
A framework for aspect based sentiment analysis on turkish informal texts.
arXiv preprint arXiv:1408.5882
Journal of Intelligent Information Systems, 53(3), 431-451.
[33] Stojanovski D, Strezoski G, Madjarov G, Dimitrovski I (2015) Twitter
[8] Khan, J., Alam, A., Hussain, J., & Lee, Y. K. (2019). EnSWF: effective
sentiment analysis using deep convolutional neural network. In: International
features extraction and selection in conjunction with ensemble learning
conference on hybrid artificial intelligence systems. Springer, pp 726–737.
methods for document sentiment classification. Applied Intelligence, 49(8),
[34] Ouyang X, Zhou P, Li CH, Liu L (2015) Sentiment analysis using convo-
3123-3145.
lutional neural network. In: 2015 IEEE international conference on computer
[9] Abdi, A., Shamsuddin, S. M., Hasan, S., & Piran, J. (2019). Deep learning-
and information technology; ubiquitous computing and communications;
based sentiment classification of evaluative text based on Multi-feature
dependable, autonomic and secure computing; pervasive intelligence and
fusion. Information Processing and Management, 56(4), 1245-1259.
computing (CIT/ IUCC/DASC/PICOM). IEEE, pp 2359–2364.
[10] S. Rezaeinia, R. Rahmani, A. Ghodsi, and H. Veisi, "Sentiment analysis [35] Jindal S, Singh S (2015) Image sentiment analysis using deep convolu-
based on improved pre-trained word embeddings", Expert Systems with tional neural networks with domain specific fine tuning. In: 2015 interna-
Applications, Vol.117, No.44, pp.139-147, 2019. tional conference on information processing (ICIP). IEEE, pp 447–451.
[11] Souma, W., Vodenska, I., & Aoyama, H. Enhanced news sentiment anal- [36] ) Yang T, Li Y, Pan Q, Guo L (2016) Tb-CNN: joint tree-bank information
ysis using deep learning methods. Journal of Computational Social Science, for sentiment analysis using CNN. In: 2016 35th Chinese control conference
1–14 (2019). (CCC). IEEE, pp 7042–7044.
[12] Rezaeinia, S. M., Rahmani, R., Ghodsi, A., & Veisi, H. Sentiment analysis [37] Bouazizi M, Ohtsuki T (2016) Sentiment analysis: from binary to multi-
based on improved pre-trained word embeddings. Expert Systems with class classification: a pattern-based approach for multiclass sentiment anal-
Applications, 117, 139–147 (2019). ysis in twitter. In: 2016 IEEE international conference on communications
[13] Tang, D., & Zhang, M. Deep Learning in Sentiment Analysis. In Deep (ICC). IEEE, pp 1–6.
Learning in Natural Language Processing (pp. 219–253). Springer (2018). [38] Abbasi A, Chen H, Salem A (2008) Sentiment analysis in multiple
[14] Rani, S., & Kumar, P. Deep Learning Based Sentiment Analysis Using languages: Feature selection for opinion classification in web forums. ACM
Convolution Neural Network. Arabian Journal for Science and Engineering, Trans Inf Syst: TOIS 26(3):12
1–10 (2018). [39] Valdivia A, Luzo´n MV, Herrera F (2017) Sentiment analysis in tripadvi-
[15] Heikal, M., Torki, M., & El-Makky, N. Sentiment Analysis of Arabic sor. IEEE Intell Syst 32(4):72–77
Tweets using Deep Learning. Procedia Computer Science, 142, 114– 122 [40] Che W, Zhao Y, Guo H, Su Z, Liu T (2015) Sentence compression
(2018). for aspect-based sentiment analysis. IEEE/ACM Trans Audio Speech Lang
[16] Li, L., Goh, T. -T., & Jin, D. How textual quality of online reviews affect Process 23(12):2111–2124.
classification performance: a case of deep learning sentiment analysis. Neural [41] Bui N, Yen J, Honavar V (2016) Temporal causality analysis of senti-
Computing and Applications, 1–29 (2018). ment change in a cancer survivor network. IEEE Trans Comput Soc Syst
[17] Alharbi, A. S. M., & de Doncker, E. Twitter sentiment analysis with adeep 3(2):75–87.
neural network: An enhanced approach using user behavioral information. [42] Wu DD, Zheng L, Olson DL (2014) A decision support approach for
Cognitive Systems Research, 54, 50–61 (2019). online stock forum sentiment analysis. IEEE Trans Syst Man Cybern Syst
[18] W. Huang, G. Rao, Z. Feng, and Q. Cong, "LSTM with sentence repre- 44(8):1077–1087.
sentations for Document-level Sentiment Classification", Neurocomputing, [43] Rani S, Kumar P (2017) A sentiment analysis system to improve teaching
Vol.308, No.45, pp.49-57, 2018. and learning. Computer 50(5):36–43.
[19] M. Fikri and R. Sarno, “A Comparative Study of Sentiment Analysis using [44] Mouthami K, Devi K.N, Bhaskaran VM (2013) Sentiment analysis and
SVM and SentiWordNet”, Indonesian Journal of Electrical Engineering and classification based on textual reviews. In: 2013 international conference
Computer Science, Vol.13, No.3, pp.902-909, 2019]. on information communication and embedded systems (ICICES). IEEE, pp
[20] Priyantina, R. A., & Sarno, R. (2019). Sentiment Analysis of Hotel 271–276
Reviews Using Latent Dirichlet Allocation, Semantic Similarity and LSTM. [45] Zhang Y, Wallace B (2015) A sensitivity analysis of (and practitioners’
International Journal of Intelligent Engineering and Systems, 12(4), 142-155. guide to) convolutional neural networks for sentence classification. arXiv
[21] Kontopoulos E, Berberidis C, Dergiades T, Bassiliades N (2013) preprint arXiv:1510.03820
Ontology-based sentiment analysis of twitter posts. Expert Syst Appl
40(10):4065–4074.
[22] Ciric M, Stanimirovic A, Petrovic N, Stoimenov L (2013) Comparison of
different algorithms for sentiment classification. In: 2013 11th international
conference on telecommunication in modern satellite, cable and broadcasting
services (TELSIKS), vol 2. IEEE, pp 567–570 .

VOLUME 4, 2016 13

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.3011802, IEEE Access

ADNAN ISHAQ is student of MS Software

Engineering at COMSATS University Islamabad,
Islamabad. Currently, he is working on NLP and
aspect base sentiment analysis.

SOHAIL ASGHAR is working as Professor

& Chairman of Computer Science at COMSATS
University Islamabad. In 2011 he joined as a Di-
rector of the University Institute of Information
Technology, PMAS-Arid Agriculture University,
Rawalpindi. Previously, In 1994, he graduated
with honors in Computer Science from the Univer-
sity of Wales, United Kingdom. He then recieved
his PhD from Faculty of Information Technol-
ogy at Monash University, Melbourne Australia in
2006. Professor Sohail has taught and researched in Data Mining (including
structural Learning, Classification, and Privacy Presevation in Data Ming,
Text and Web Mining), Big Data Analytics, Data Science and Information
Technology areas, and he has published extensively (More than 150 pub-
lications) in international journals as well as conference proceedings. He
has also consulted widely on information Technology matters, especially in
the framework of Data Mining and Data Science. In 2004 he acquired the
Australian Postgraduate Award for Industry. Professor Sohail is a member
of, the Australian Computer Society (ACS), and IEEE and also Higher
Education Commission Approved Supervisor. Professor Sohail is in the
Editorial Team of well reputed Scientific Journals.He has also served as
Program Committee member of numerous International Conferences and
regularly speaks at international conferences, seminars and workshops.

SAIRA ANDLEEB GILLANI is Senior Assistant

Professor at Department of Computer Science,
Bahria University Karachi Campus. She has re-
ceived her PhD Degree from Corvinus University
of Budapest, Hungary in 2016. She is working in
NLP, text mining and data mining domain.

14 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.