International Journal of Computer Science Trends and Technology (IJCST) – Volume 11 Issue 1, Jan-Feb 2023

RESEARCH ARTICLE OPEN ACCESS

Sentiment Analysis for E-Commerce Product Reviews Using

Deep Learning
Rogaia Yousif [1], Ashraf Abdalla [2]
[1]
Department of Computer Science, University of Gezira, Wad Madani - Sudan
]2[
Department of Computer Science, Sudan University of Science and Technology, Khartoum - Sudan

ABSTRACT
Sentiment analysis (SA) is a process of extensive exploration of data stored on the web to identify and categorize the views
expressed in a part of the text. In the sentiment classification, the accuracy or speed are difficulty to remove semantic ambiguity
in sentiment analysis from the text ,but there are many machine learning that have a lot of similarities in features. Therefore, in
this paper, we have expanded the sentiment into sentence analysis. We used deep learning model in natural language processing
to analyse the sentence present in a given English language text, we used data on social media platforms such as Twitter. After
working on processing the text and putting it in an understandable way for the classification model to find more efficient pre-
processing techniques and more accurate and fast way to analyze sentiment from text, we used deep learning models by a
recurrent neural network (RNN) with a state memory and multilayer cell structure Long Short-Term Memory (LSTM). We
experimented and evaluated the method using Recurrent Neural Networks and Long short-term memory on dataset consists of a
nearly 3000 Amazon customer reviews as input text , star ratings, date of review, variant and feedback of various amazon Alexa
products for learning how to train learning for sentiment analysis, our use of this data to analyze Amazon Alexa product and
API data; Discover insights into consumer reviews that are aided by deep learning models to achieve high emotion
classification accuracy. A thorough evaluation shows that the system gains emotion prediction on LSTM model with 92.10 %
accuracy for positive/negative classification with train time 25m for dataset and 93.18 % accuracy for positive and negative
with train time 10s for API data.
Keywords: - Natural Language Processing (NLP); Sentiment Analysis; Twitter Platform ;Deep Learning Classifiers; RNN;
LSTM.

computers etc. Sentiment are physiologic thoughts engendered

I. INTRODUCTION in human reactions to the events [1] .Sentiment analysis (SA) is
Natural Language Processing is a field that belongs to a process of extensive survey of data stored on the Web to
artificial intelligence, sentiment analysis recognition from text identify and categorize the views expressed in a part of the ext.
is crucial natural language processing (NLP) task which can The intended outcome of this process is to assess the author
contribute enormous benefits to different areas artificial attitude toward a specific topic, movie, product, etc. The result
intelligence. Which acts as an interface between humans and is positive, negative, or neutral. These studies illustrated
computers for interaction using various computational different techniques in SA approach for extracting and
algorithms. Recent uses of NLP algorithms are spam mail sentiments analyses associated with the polarity of positive,
detection, meeting report analysis, auto question answering, negative, or neutral on the topic selected. Social networks SA
summarizing media reports, machine translation, and many can be a useful source of information and data. SA acquires
more. Deep Learning is an postponement of machine learning important in many areas of business, politics, and thought [2].
and artificial intelligence that teaches computers to learn from Analysis of these Sentiment without facial and voice inflection
experiences the same as humans do. Many deep learning are critical and requires a supervisory approach for proper
models are successfully deployed for numerous natural interpretation of sentiment. In spite of these challenges, it’s
language processing tasks for the last scarce years. Natural essential to acknowledge the human sentiment as they
Language Processing (NLP) is originated from Artificial progressively communicate using maltreatment text through
Intelligence It reads, deciphers, understands, and convert the social media applications such as Facebook, Twitter etc. Here,
human input in a valuable manner. It takes text data as input we use deep learning techniques to classify the sentiments of an
and converts it into more understandable structured data, easily expression into positive or negative sentiment. The positive
sentiment is further classified into eagerness, fun, happiness,
love, neutral, relief, surprise and negative sentiment are
analyzed and summarized. Using NLP algorithms, computers classified into anger, boredom, emptiness, hate, sadness, worry.
can read real-time data or listen from the audio source, We investigated and evaluated the method using Recurrent
understand it, and try to extract sentiment, feelings, and Neural Networks and Long short-term memory on three
meanings associated with it recent human interaction with different datasets to show how to achieve high sentiment

ISSN: 2347-8578 www.ijcstjournal.org Page 26

International Journal of Computer Science Trends and Technology (IJCST) – Volume 11 Issue 1, Jan-Feb 2023
classification accuracy. Sentiment analysis is widely utilized by deep learning models such as deep neural networks(RNN),
different parties such as shoppers or marketers to gain insights convolutional neural networks(CNN) and many more for
into merchandise or know the market trends [3]. Also, this is solving different problems of sentiment analysis such as
critical in predicting the exchange rate or the product rating of a sentiment classification, cross lingual problems, textual and
selected organization. This is done by analysing the sentiment visual analysis and product review analysis, etc. The authors
of the public towards the corporation with regard to time and showed full concept about SA techniques and its classification
positioning [4] .There is a need for a social science tool in order and methods used it. In [9], proposed a sentimental
to understand the importance of public sentiments and thereby classification of multitude of tweets. Here, they used deep
the market value of the enterprises. Firms can also estimate how learning techniques to classify the sentiments of an expression
well their product is moving in the market. Furthermore, it will into positive or negative sentiment. The positive sentiment are
help to analyze the positive and negative feedbacks about the further classified into enthusiasm, fun, happiness, love, neutral,
products. This information will assist the corporation in relief, surprise and negative sentiment are classified into anger,
analyzing different responses so that they can market their boredom, emptiness, hate, sadness, worry. They experimented
products in an optimized manner. The firms can also market and evaluated the method using Recurrent Neural Networks
their products using appropriate techniques, understand latest (RNN) and Long short-term memory (LSTM)on three different
trends and develop market challenging products. The datasets to show how to achieve high emotion classification
investigation of a numerical improvement made the application accuracy. A thorough evaluation shows that the system gains
to change to the circle of AI (Artificial Intelligence).Data emotion prediction on LSTM model with 88.47% accuracy for
processing has a major role in research areas and in the field of positive/negative classification and 89.13% and 91.3% accuracy
study among AI. While considering its application across for positive and negative subclass respectively. In [10] based on
various businesses and research problems, AI is also considered the deep learning model, an improved sentiment on social
as prognostic systematic [5]. During the last two decades, social networks, such as Twitter, has become a powerful means of
media has become a big part of people’s daily life. It is a learning about the users’ opinions and has a wide range of
communication channel where attitudes and opinions are shared applications. However, the efficiency and accuracy of sentiment
and as such, it is a tool that shapes the future generation. Social analysis is being hindered by the challenges encountered in
networks open up the possibility for people everywhere to natural language processing (NLP). Deep learning models are
connect and make a difference and through that, it has the promising solution to the challenges of NLP. In [11] The
power to push instant change. Social media is an ideal platform authors showed reviews the latest studies that have employed
to study these aspects thanks to the large amount of data that is deep learning to solve sentiment analysis problems such as
available [6]. sentiment polarity. Models using term frequency-inverse
Therefore, in this paper, we proposed a deep learning model document frequency (TF-IDF) and word embedding have been
based on the use of Recurrent Neural Network (RNN) with a applied to a series of datasets. In [12], the study was proposed a
state memory and multilayer cell structure Long Short-Term sentiment analysis model-a SLCABG model for sentiment
Memory (LSTM) which can increase accurately and more analysis on product reviews is constructed used sentiment
efficient pre-processing techniques to remove ambiguity from dictionary, BERT model, CNN model, BiGRU model, and
the text. In addition to presentation of data that expresses text in attention mechanism. The sentiment lexicon is used to enhance
a more understandable way. the sentiment features in the reviews. Then the CNN and GRU
The rest of this paper is organized as follows: the next Section networks are used to extract the main sentimental and
2 presents related works. In Section 3, we provide the materials contextual features of the reviews, and attention mechanism is
and methodology. Section 4, presents the main results and used to weight them, the weighted sentiment features are
discussion. Finally, Section 5 concludes this paper. classified by analyzed the experimental results, it can be found
that the model has better classification performance than other
II. RELATED WORKS sentiment analysis models .They used model to analyze user
review to help merchants on e-commerce platforms to obtain
Given the aim of this work, we recall some representative
user feedback in time to improve their service quality and
and related existing works to examine the posts to see if there is
attract more customers to patronize. In[6] BERT-based models
any relation between the future stock price and users’
outperform the rule-based models VADER and TextBlob when
sentiment. In other words, we want to see if we can predict a
it comes to precision, recall, and F1-score. Rule-based models
future stock price based on the current sentiment of many users.
add a bias in the classifications when using data collected with
For example, In [7]Zucco et.al illustrated different techniques keywords due to the sentiment score of the keyword itself.
in SA approach for extracting and analyses sentiments More complex deep-learning models like BERT do not result in
associated with the polarity of positive, negative, or neutral on the same bias since it is based on context rather than rules.
the topic selected. Social networks SA can be a useful source of By adding a more comprehensive filter to Twitter API the data
information and data. SA acquires important in many areas of could be more suitable for the analysis. This could decrease the
business, politics, and thought. So, this study contains a amount of noise and outliers in the dataset. In all of the
comprehensive overview of the most important studies in this
field from the past to the recent studies till. In [8], They review
paper highlights latest studies regarding the implementation of

ISSN: 2347-8578 www.ijcstjournal.org Page 27

International Journal of Computer Science Trends and Technology (IJCST) – Volume 11 Issue 1, Jan-Feb 2023
different SDGs, the majority of tweets were positive and all had
a similar pattern in the distribution. Most of the keywords had
between 60-70% positive tweets. In[13], authors focused on the
basics of sentiment /opinion mining and its levels. There are Data Collection
various approaches and methods to identify sentiment from
content. They examined represents machine learning
procedures. From various classification methods, Sentiment
Analysis indicates the results into positive, negative and neutral
scores. They showed that machine learning methods, such as
SVM, Naive Bayes, and neural networks had the highest Normalizing the tweets
accuracy and can be considered as the baseline learning
methods as well as in some cases lexicon based methods are
very effective. Neural network and deep learning shows more
accuracy than usual machine learning algorithms. They
recommended in future work and discovering the result of
Text Preparation
various other combinations of text data and other on prediction
accuracy can be done. In addition to more work in the future is
needed to advance performance measures.
From the presented related works, there are many studies that
have classified sentiment focused one aspect of the analysis
sentiment analysis paradigms , we can conclude that obtaining Feature Selection
high accuracy in the classification. Hence, the comparison of
our proposed model and the state-of-the-art proved the validity
and efficiency of our proposed model which exceeds all
proposed models' results and demonstrated its robustness in Sentiment Classification
coping with the limited availability of training data. In fact, in (LSTM Model)
the sentiment classification, the accuracy or speed are difficulty
to remove semantic ambiguity in sentiment analysis from the
text. but there are many machine learning that have a lot of
similarities in features. Therefore, in this paper, we have expand
the sentiment into sentence analysis. We used more efficient
pre-processing techniques and find a more accurate and fast Presentation of Results
way to analyse sentiment from text, we are using deep learning
models by a recurrent neural network (RNN) with a state
memory and multilayer cell structure Long Short-Term
Memory (LSTM). Fig .1 The five steps of sentiment analysis

III. MATERIALS AND METHODS 1) Data Collection

Though data sets for the experiments are widely available
In this paper, we used deep learning models in natural across various social networking platforms, the manual
language processing to analyse the emotion present in a given determination of sentiments of different tweets is a challenging
English language text. Using data on social media platforms task. It requires a high level of expertise in the area to perform a
such as Twitter. After working on processing the text and vigilant analysis of these tweets and generate accurate results.
putting it in an understandable way for the classification model. Existing public datasets having URL are rather limited due to
In this section, the sentiment analysis methodology differs these challenges.
according to the need for the results and their accuracy .Five This dataset consists of a nearly 3150 Amazon customer
main steps, ranging from the place of data collection as texts to reviews (input text), star ratings, date of review, variant and
the appearance of the results will be introduced (See Figure 1). feedback of various amazon Alexa products like Alexa Echo,
In the rest of this Section, we describe the main datasets, the Echo dots, Alexa Firesticks etc. for learning how to train
employed deep learning techniques, as well as performance Machine for sentiment analysis, our use of this data to analyse
evaluation of the proposed model. your Amazon Alexa product; Discover insights into consumer
reviews that are aided by deep learning models. In Features of
dataset descriptions of the features that were utilized to generate
the data sets are provided below, each row of the log contains
the information. In Table 1 we provide a description of the
newly created dataset, and in Figure 3 samples of the dataset are
depicted.

ISSN: 2347-8578 www.ijcstjournal.org Page 28

 International Journal of Computer Science Trends and Technology (IJCST) – Volume 11 Issue 1, Jan-Feb 2023
3) Pre-processing
Table.1 Description of the Features of dataset
The messages from Twitter are too informal and has
Index Rat Date Variation Verified_reviews Feed different styles of using tweets based on the nationality, origin,
ing back age and the gender of the user. Therefore, tweets taken from
1 5 29-Jul-18 Charcoal Love my Echo! 1 twitter generally result in a noisy data set of unwanted
Fabric emoticons and symbols. Twitter users will use different kinds of
2 5 29-Jul-18 Charcoal Loved it! 1 special characters like sending the tweets again, which are
Fabric termed as re-tweets, emoticons, personalized wording etc. all of
3 4 29-Jul-18 Walnut "Sometimes while 1 which are to be suitably extracted. Hence for creating new
Finish playing a game, datasets, raw twitter data has to be normalized for easiness of
yo… the classifier algorithms to perform the emotion analysis. There
4 5 29-Jul-18 Charcoal "I have had a lot of 1 are different pre-processing methods available for tweets[2] .
Fabric fun with this The main reason for using the pre-processing task is to reduce
thing…
the noise and size of the messages. Initially we perform few
5 5 29-Jul-18 Charcoal Music 1 general and mandatory pre-processing task which include
Fabric
conversion of words to lower case. Also, there are tweets with
6 5 29-Jul-18 Heather "I received the 1
Gray... echo as a gift. I more dots as we use informal conversation. So, the additional
neede... dot is replaced with a space. At the end, we remove unwanted
7 3 29-Jul-18 Sandstone "Without having a 1 space and special characters. Multiple spaces can be avoided by
Fabric cellphone, I replacing them with a single space. The URL from the twitter
canno... messages will not be used in the analysis as it leads to a sparse
8 5 29-Jul-18 Charcoal I think this is the 1 result. Hence, we replaced all the URL in the tweets with the
Fabric 5th one I've expression ((www\. [\S] +))) . Handle is another special
purchas... character used in tweets to mention other user names in their
9 5 29-Jul-18 Heather looks 1 messages. The users commonly use the tagging of another user
Gray... great as @name_user.
10 5 29-Jul-18 Heather "Love it! I’ve 1
As we are eliminating all unwanted symbols from the data sets,
Gray... listened to songs I
have… we replace all user tagging like the @name_user with the word
... .. … … … … USER-MENTION. The equivalent expression used to match
user tagging is @[\s] +.2. It is very common to use different
3148 5 29-Jul-18 Black Dot "I do love these 1 types of images in the tweets. These include smiley symbols,
things, i have hand gestures, etc. Since the social media sites have a wide
them .. range of such images, it is very difficult to find equivalent
3149 5 29-Jul-18 White Dot “Only complaint I 1 expression for each among them.
have is that the… However, we cannot exclude the text as they play a major role
3150 4 29-Jul-18 Black Dot Good 1 in the emotion transfer and therefore, we use two types of
substitutions for positive and negative sentiment as emoj-
positive and emoj-negative respectively. Hashtags are used to
Normalizing tweets makes the data suitable for pre-processing. categorize tweets and messages on similar or common area so
Pre-processing steps are mandatory to avoid spam codes and
that people can easily find the area of interest. They are phrases
get tweets in a natural way, which in turn defines the data that without any space and preceded by a # symbol. When we click
needs to be handled and configured for the pre-processing on the hashtags it will move to other tweets that also have the
process.
same hashtag. So, these are from the same area of interest. For
replacing and giving a common format for these hashtags, we
2) Text preparation will remove the hash symbol and replace the hashtag with the
Text preparation examines the data before analysing it. word which comes after the hash symbol. The regular
Some reviews and conversations in the communication sites expression used to match the hashtag is # (+).
contain offensive and inappropriate words, so they are It is common to resend the tweets which we have already
examined and preparation to be the result more reliable analysis. received from other users for many reasons. This gives more
This process selects the contents that are not related to the flexibility to the use of tweets and make this platform more
analysis and then removes it. Objective of the process is the interesting. Normally these retweets begin with the letter RT
removal of spam and inappropriate reviews before sent to and during the pre-processing phase these RT will be removed
automated analysis. We are using NLTK techniques. as it is not relevant for the classification of tweet sentiment. The
retweets are represented as \brt\b. After the pre-processing of
the URL, Hashtags and retweets, the main part of tweet
processing is the clearing of punctuations in the tweets. That is,
in a tweet message like “Oh my God!! I cannot believe this.”

ISSN: 2347-8578 www.ijcstjournal.org Page 29

International Journal of Computer Science Trends and Technology (IJCST) – Volume 11 Issue 1, Jan-Feb 2023
we will remove the punctuation and rephrase the tweets. For learning models we often used two different sets. A train set is
getting better features for the classification we need to remove used for training the model, A test set that is used to evaluate
all the punctuation, like’’‘! ()-[]{};:’"\, <>. ? @%ˆ* ”’ /. the final model accuracy before deployment.
Twitter is a platform where users can convey their feelings
through different expressions. Many people will post tweets 5) Sentiment Classification by LSTM model Constructor
like “I am sooooo surprised” to highlight their sentiment. But
during the word level processing of tweets we must ignore these LSTM helps us to classify our tweets by also using the
duplications. The other word level processing includes the long range of dependencies. The LSTM has flexibility to get rid
removal of “-” and “’” which is commonly used in typing t- of or append information to the cell which keeps track of the
shirts and can’t. These are represented as ”tshirt” and ”can”t in sate information, and these are controlled by gates. Gates allow
a more generalized manner [2] . the information to pass through the network. The segment layer
has the outputs range between 1 and 0 which denotes the
4) Features Selection information which are passing through. If the obtained value is
0, it tells us that nothing is passing through the gate and a value
In features selection we used Term Frequency Inverse of 1 describes that everything can be pass through. Initially, we
Document Frequency (TF-IDF) is used in natural language have to identify the features to be fed to the cell state. The forget
processing for identifying the important words or rare words in layer in the LSTM architecture is responsible for this decision
a text data. Term frequency converts the words that are in string whether to forget or accept the features. By considering the
format for numerical formatted data so that the machine value of ht-1 and xt it will give two values as output. That is, the
output values can either be a 0 or a 1 for each cell state. If the value is
learning models can understand the information. Term 1, it represents the acceptance of the feature and 0 represents the
Frequency TF is used to find out the frequency of occurrence of avoidance of the information. In this work we must accept or reject the
words that we have for the classes. In our data set the feature is feature words from the data sets to classify the emotion. In each cell
state, the new emotion word has to be accepted or rejected based on the
words. The frequency of each word in the dataset which is emotion of the tweets (See Figure 2).
being calculated using term frequency.

IDF helps us to take meaningful words from the document. TF

gives us the highest degree words while IDF helps us to get the
lowest occurred words by taking the logarithm of the values.

We find that the product of both the TF and IDF matrices is the
normalized weights which is the TF-IDF output. In this way we Fig .2 LSTM model constructor
get the numerical input for the machine learning model. TF-IDF
is used to represent text with a BoW (Bag of Words). Test cases for positive and negative emotions are demonstrated
Converting the Reviews into Numerical Vectors (doc2vec). here with an example.
The doc2vec vector algorithm performs considerably well for Case 1: Positive:
sentence similarity tasks. However, if the input corpus includes For example: “omg!!! It is surprising,” it is categorized as
a lot of words with misspellings like tweets, this algorithm may positive. Further, it is categorized based on the percentage in
not be an ideal choice. We used Doc2Vec method which is used correspondence to each emotion. Here we get 40.91% for
for the vectorization of documents. This is an improved version surprise, 28.27% of relief, 11.64% for fun, 7.22% of neutral and
of Word2Vec. This method was not desirable in the works less than 5 percentage for another three emotions each, which
where a lot of misspellings of words occur. It is better to shows that the emotion surprise is more dominant in this
convert words to vectors and then use these vectors to create the comment. Hence it is categorized as a surprise.
vector format of the whole document. Doc2Vec is used to Case 2: Negative:
represent text with word vectors. For example: “I am so panic these days”, it is categorized as
Train-Test Split Date ,One of the golden rules in deep learning negative emotion statement. Further, it is categorized based on
is to split your dataset into train, validation, and test set. The the percentage of correspondence to each emotion. Here we get
reason we do that is very simple. If we would not split the data 78.43 percentage to worry, only 14.82% is for sadness, 5.49%
into different sets the model would be evaluated on the same of empty and other three emotions together obtain only 1.26%
data it has seen during training, Back before using deep of the total. So, it is categorized as worry.

ISSN: 2347-8578 www.ijcstjournal.org Page 30

International Journal of Computer Science Trends and Technology (IJCST) – Volume 11 Issue 1, Jan-Feb 2023
Twitter represents a data source and contains a specific text
Model Accuracy format that enables easy text classification.
In the measurement of accuracy, there have chances of miss To analyze sentiment from sentence, we need many tools.
interpretation of the prediction. These tools are divided between tools for collecting data from
the data source, tools for processing sentence before the
classification process, and classification tools that represent
functions from a programming language that enables analysing
and extracting sentiment from the sentence. We present the
That is, the actual defected results are being identified as true experimental results and used the LSTM (Long-Short Term
cases and is termed as True Positive, represented as TN. Also, Memory) algorithm to classify the emotions from the text
some cases which are correct are being recognized as negative, (positive or negative) after training the model on the training
called as False positive and is represented as FP. data and testing the model with the test data. The experimental
Presentation of Output steps were a dataset loading the dataset and creating a new
The main objective of analysing a huge amount of data is to column ‘sentiment’ based on ‘rating’. This dataset encoded the
convert unstructured text into useful information and then to target variable with a 2-point ordinal scale: 0 = negative, 1 =
display it through charts such as a graph, line graph, and bar positive.
graph [5]. We are building the LSTM model are depicted in Figure.4
Performance Evaluation of the Model using the ‘Keras’ library. This step involves model initialization,
In this phase, performance evaluation will be builder the adding required LSTM layers, and model compilation.
LSTM model used the ‘Keras’ library. This step involves model
initialization, adding required LSTM layers, and model
compilation. Hence, in the measurement of accuracy, there have
chances of miss interpretation of the prediction. That is, the
actual defected results are being identified as true cases and is
termed as True Positive, represented as TN. Also, some cases
which are correct are being recognized as negative, called as
False positive and is represented as FP, we have to calculate the
following metrics for the test rating.
• True Positive (TP): Refers to the number of
predictions where the classifier correctly predicts the
positive category as positive.
• True Negative (TN): Refers to the number of
predictions where the classifier correctly predicts the
negative category as negative.
• False Positive (FP): Refers to the number of
predictions where the classifier incorrectly predicts the
negative class as positive. Fig .3 Model Design
• False Negative (FN): Refers to the number of
predictions where the classifier incorrectly predicts the We can see that the training the model using train data are
positive class as negative. depicted in Figure5 and Figure6 .The train-test split procedure
Before using deep learning models, we often used two different is used to estimate the performance of deep learning LSTM
sets. A train set is used for training the model and a test set that Model when they are used to make predictions on data not used
is used to evaluate the final model accuracy before deployment. to train the model.

6) RESULTS AND DISCUSSION

Python programming language and Keras packages were
used to train the proposed deep transfer learning models. It is
contains libraries of artificial intelligence models and ready-
made classification methods. And Jupyter notebook is the latest
web-based interactive development environment for notebooks,
code, and data. Its flexible interface allows users to configure
and arrange workflows in data science. API were used to
collected information from the source and sends it to the
classification model. All experiments were performed on

ISSN: 2347-8578 www.ijcstjournal.org Page 31

International Journal of Computer Science Trends and Technology (IJCST) – Volume 11 Issue 1, Jan-Feb 2023

Table .2 Performance measurements for all models

Number of Data Accuracy Loss Train time
layers

Data Set 0.9210 0.079 25m

5 API Data 0.9318 0.0682 10s

In[9] the authors present modeled a system for sentiment

analysis of twitter messages. The tweets we consider in the
analysis are, a mixture of different words and emoticons. We
Fig .4 Training the model modeled the classifier With deep learning techniques such as
RNN and LSTM. In order to gain better accuracy, we
incorporated different feature selection methods like TF-IDF
and Doc2Vect. The feature extraction generates a vector that is
been given as input to the classification model. Our model
achieved better results in the task of classifying the twitter
emotional messages. As a future work, the analysis of the
personality of the users from their tweets are need to be
investigated, so that the system can be more personalized.

IV. CONCLUSION & FUTURE SCOPE

In this paper, a deep learning model based on transfer
learning is proposed used the LSTM (Long- Short Term
Memory) algorithm to efficiently classify the emotions from the
text (positive or negative) after training the model on the
Fig .5 Training-Testing Accuracy training data and testing the model with the test data. We found
that the LSTM model gives more accurate results in
classification, the greater the number of data used in training,
Save the model allows to save your model to file and load it later in and the greater the number of layers of the model structure, the
order to make predictions. We use predictions model by new more accurate the model in classification, and that is inversely
data to obtain actual data to test the model's work on any other proportional to the training time and loss data. Evaluation of the
data, we obtain that through the API, data marked with model shows that the system gains emotion prediction on
(hashtags) from Twitter. LSTM model with 92.10 % accuracy for positive/negative
classification with train time 25m for dataset and 93.18 %
accuracy for positive and negative with train time 10s for API
data. In future work, we recommend to combine LSTM to get
less execution time with the layers for large twitter datasets.

ACKNOWLEDGMENT

We would like to thank the reviewers, whose valuable

feedback, suggestions and comments increased significantly the
overall quality of this survey. And we very grateful respected
professor Ashraf Gasem Elseed Abdalla for his guideline and
insightful feedbacks.

REFERENCES
Fig .6 Results Of Predictions [1] D. N. Mati, M. Hamiti, A. Susuri, B. Selimi, and J.
Ajdari, "Building Dictionaries for Low Resource
Results of model is content of accuracy and loss of dataset for
Languages: Challenges of Unsupervised Learning,"
new API data from twitter and execute time. The final result of
the model prediction for API data are depicted in Table 2.

ISSN: 2347-8578 www.ijcstjournal.org Page 32

International Journal of Computer Science Trends and Technology (IJCST) – Volume 11 Issue 1, Jan-Feb 2023
Annals of Emerging Technologies in Computing Interdisciplinary Reviews: Data Mining and
(AETiC), vol. 5, pp. 52-58, 2021. Knowledge Discovery, vol. 10, p. e1333, 2020.
[2] A. A. Q. Aqlan, B. Manjula, and R. Lakshman Naik, [8] Q. T. Ain, M. Ali, A. Riaz, A. Noureen, M. Kamran,
"A study of sentiment analysis: concepts, techniques, B. Hayat, et al., "Sentiment analysis using deep
and challenges," in Proceedings of International learning techniques: a review," International Journal
Conference on Computational Intelligence and Data of Advanced Computer Science and Applications, vol.
Engineering, 2019, pp. 147-162. 8, 2017.
[3] N. Yadav, O. Kudale, S. Gupta, A. Rao, and A. [9] P. Shilpa, R. Shereen, S. Jacob, and P. Vinod,
Shitole, "Twitter sentiment analysis using machine "Sentiment Analysis Using Deep Learning," in 2021
learning for product evaluation," in 2020 International Third International Conference on Intelligent
Conference on Inventive Computation Technologies Communication Technologies and Virtual Mobile
(ICICT), 2020, pp. 181-185. Networks (ICICV), 2021, pp. 930-937.
[4] D. Ramachandran and R. Parvathi, "Analysis of twitter [10] N. C. Dang, M. N. Moreno-García, and F. De la Prieta,
specific preprocessing technique for tweets," Procedia "Sentiment analysis based on deep learning: A
Computer Science, vol. 165, pp. 245-251, 2019. comparative study," Electronics, vol. 9, p. 483, 2020.
[5] T. Sajana, S. Jacob, P. Vinod, V. G. Menon, and P. [11] M. Kamyab, G. Liu, and M. Adjeisah, "Attention-
Shilpa, "Context-aware gender and age recognition based CNN and Bi-LSTM model based on TF-IDF and
from smartphone sensors," in 2022 International glove word embedding for sentiment analysis,"
Conference on Computing, Communication, Security Applied Sciences, vol. 11, p. 11255, 2021.
and Intelligent Systems (IC3SIS), 2022, pp. 1-6. [12] L. Yang, Y. Li, J. Wang, and R. S. Sherratt,
[6] E. Rosenberg, C. Tarazona, F. Mallor, H. Eivazi, D. "Sentiment analysis for E-commerce product reviews
Pastor-Escuredo, F. Fuso-Nerini, et al., "Sentiment in Chinese based on sentiment lexicon and deep
analysis on Twitter data towards climate action," 2023. learning," IEEE access, vol. 8, pp. 23522-23530, 2020.
[7] C. Zucco, B. Calabrese, G. Agapito, P. H. Guzzi, and [13] M. Rahul, R. Rajeev, and S. Shine, "Social media
M. Cannataro, "Sentiment analysis for mining texts sentiment analysis for Malayalam," Int. J. Comput.
and social networks data: Methods and tools," Wiley Sci. Eng, vol. 6, pp. 48-53, 2018.

ISSN: 2347-8578 www.ijcstjournal.org Page 33