1

Hybrid CNN and RNN for Twitter Sentiment Analysis

Adish Padalia1, Pratik Pujari 2, Tanish Shah3 and Dr. Kailas Devadkar4
1, 3, 4
Department of Information Technology, Sardar Patel Institute of Technology, University Of Mumbai, Mumbai, India
2
Department of Computer Engineering, Sardar Patel Institute of Technology, University Of Mumbai, Mumbai, India
1
[email protected], [email protected], [email protected], [email protected]

Abstract- Online social media networks have developed into public uses Twitter, it is a tremendously rich data source that
a widespread and significant platform for people to freely may be utilised to examine a variety of sentiment data. Twitter
express their thoughts and emotions in this era of digital is a reliable OSN (online social network) for user conversations
communication. This change in communication habits has and information sharing. Various elements of our lives are
important ramifications for brands looking to connect with significantly influenced by Twitter mood [11]. In order to
and understand their target market. Social media is one of extract textual information, SA and text classification further
the biggest sources of unstructured data, but it takes time to classify the polarity as positive (P) or negative (N). Information
analyse it and extract its meaning. This paper attempts to may frequently be retrieved from text or tweet content using
forecast sentimental analysis based on Twitter user's tweets. NLP algorithms. The process by which the machine (computer)
We used a Twitter dataset which contains around 520k determines the meaning of each statement produced by a human
Tweets, to predict whether the tweet has a positive or a is known as NLP-based sentiment classification. TSA (Twitter
negative connotation which can further help us to Sentiment Analysis) manual analysis takes longer and
understand the mental state of the user. Then, using this additional professionals are needed for tweet labelling. Due to
dataset, we develop different deep learning models the availability of big datasets, some of the existing research
(including RNN and CNN). We examine the outcomes of struggles to attain efficient processing time, complexity, and
applying recurrent neural network (RNN) and accuracy. Additionally, the efficiency of the classifier is
convolutional neural network (CNN) to these models. decreased by the extraction of irrelevant and low-level
Finally, we suggested an approach that combines RNN and characteristics. Additionally, the use of all extracted
CNN to fully exploit each technology's advantages: RNN characteristics takes up a lot of space. These flaws render the
can learn temporal and context features, notably long-term current algorithms unsuitable for processing data efficiently.
dependency between multiple entities, while CNN is capable These flaws present a research opportunity for an efficient
of catching numerous potential features. The result integrated algorithm for the processing of Twitter data.
demonstrates that our method is superior to the majority of Consequently, an automated model is created to address these
the existing methods. issues.
The analysis of online sentiments has made use of the
Index Terms- Sentiment Analysis, Deep Learning, advances of ML (Machine learning) algorithms [22][10],
Convolutional Neural Network (CNN), Recurrent Neural including SVM (Support Vector Machine), MNB (Multinomial
Network (RNN) Naive Bayes), LR (Logistic Regression), NB (Naive Bayes),
etc. Although these techniques showed good effectiveness, they
1. INTRODUCTION are relatively slow and require more time to complete the
To assess the user's sentiments, sentiment analysis training process. To effectively classify Twitter attitudes, the
(SA) employs text analysis, NLP (natural language processing), DL model is introduced. DL is a subclass of ML that employs
and statistics. Sentiments are emotions, ideas, or attitudes that numerous techniques to resolve challenging issues. DL allows
are conveyed in relation to a person, circumstance, or object. the machine to handle massive amounts of data with little
SA is used to determine whether the data or information human input by using a series of progressive events. The
obtained is positive or negative. Business experts commonly accurate outcomes of DL-based sentiment analysis may be used
utilise it to track attitudes, interpret social data, and assess brand for a number of tasks, such as emotion detection [25], product
reputation and client demands[14][7].The volume of prediction, movie recommendation, and other uses. Several
information created or uploaded online has quickly increased as academics have introduced DL in Twitter sentiment analysis as
a result of the significant growth in Internet usage in recent a result of these advancements.
years[23][7]. Since the advent of technology, people have There are numerous issues with employing DL
utilised social media platforms [18][17]like Twitter, Instagram, approaches for Twitter sentiment analysis. The author of [6]
Facebook, LinkedIn, YouTube, etc. to communicate their ideas performed sentiment classification from Twitter data using the
or opinions about various goods, occasions, or objectives. The DL model. This strategy examined each user's behavioural data
most popular global microblogging site nowadays for users to to categorise such data. However, this approach has had
express their ideas in the form of quick messages called tweets difficulty extracting precise tweet words from the enormous
is Twitter[12].On average, 6,000 tweets are sent on Twitter twitter corpus; as a result, the effectiveness of a classification
every second, which equates to more than 350,000 tweets sent algorithm has decreased. However, it is unable to lower the
every minute, 500 million tweets every day, and around 200 dimension that the extracted features take up. As a result, a
billion tweets every year[9]. Given how frequently the general number of useful traits are inside the local optimum. In[6], they
 2

have only used deep convolutional neural networks in order to The framework integrated an NLP and Bi-LSTM model to
classify tweet sentiments. The paper proposed a novel CNN optimally classify the sentiments of the tweets of users.
architecture which initialises the parameter weights. But it also The study conducted by Bello et al[5] presents a
proposes difficulties like having limited capabilities for word sentiment analysis method using Bidirectional Encoder
embeddings, difficulty in handling noises and loss in sentiment Representations from Transformers (BERT). The study
continuity. In another paper, published by Stanford [16], the proposes a text classification method that uses BERT in
neural net structures they experimented included one-hidden- combination with other variants like CNN, RNN, and BiLSTM.
layer Recursive Neural Net (RNN), two-hidden-layer RNN and The experimental findings show that these combinations
Recursive Neural Tensor Net (RNTN). RNNs are well suited perform well in terms of accuracy rate, precision rate, recall
for handling sequential information processing, handling rate, and F1-score compared to when BERT was used with
varying length and effective handling of noise but on the other Word2vec and when it was used with no variant.
hand the proposed methodology using RNNs contains a few The authors Ahmad et al [1] used SVM, a widely used
flaws that are limited context window, long training times, supervised machine learning algorithm for textual polarity
insufficiency in handling short texts, etc. All these flaws can be detection, to analyse the performance of sentiment analysis.
mitigated by integrating CNNs and RNNs in a hybrid model, The SVM was trained with two pre-classified datasets of
one can take advantage of the superior local feature extraction tweets. However, the authors also pointed out some problem
capabilities of CNNs and the sequential learning abilities of areas in the training data such as multiple occurrences of tweets,
RNNs. This hybrid approach can provide a more holistic opinion spamming, and dual opinion tweets which could affect
understanding of sentiment in tweets, effectively addressing the the accuracy. But on the contrary, it didn't consider the use of
limitations of each individual architecture. Additionally, it can additional features or different feature extraction techniques.
enhance the accuracy of sentiment analysis by fusing both local Mandloi and Patel [13] proposed various Machine
and contextual information, resulting in more robust and Learning methods like the Naïve Bayes, SVM, and Maximum
nuanced sentiment predictions in the context of tweet data. Entropy methods are compared. The paper discusses how
With these features, the suggested classifier performed better sentiment analysis is done by these classification algorithms
across the huge dataset and also obtained more accuracy while and what is the accuracy and precision in these cases. However,
making less classification errors. it is important to note that these ML classification models can
be effective for sentiment analysis but may not always perfectly
2. RELATED WORK capture the nuances and complexities of human language and
Sergiu Cosmin Nistor et al [16] developed a sentiment sentiment. Also, it is very sensitive to the quality and
analysis framework using Recurrent Neural Networks for tweet representativeness of the training data used.
classification, achieving an 80.74% accuracy rate. The method E et al[8], has taken a dataset including US airline
uses an attention mechanism to localise emotion features, online review and discusses the importance of the internet in
enhancing the network's performance. Tested on a large corpus decision-making, highlighting the need for sentiment analysis
of 1.5 million tweets, the methodology effectively analyses to help customers choose the best US airlines. It introduces a
emojis and offers improved results. new Adaboost approach for sentiment analysis, employing
In 2022, Rakin Mostafa [15] and his team developed a machine learning algorithms for performance analysis. The
sentiment analysis method for tweets using Natural Language research aims to bridge gaps between customer views and
Processing (NLP) and Bidirectional Long Short Term Memory airlines, potentially claiming it can perform well in other
(Bi-LSTM). The study aimed to classify the types of sentiment domains as well.
in tweets, which was based on how people feel about a
company’s products through microblogging.

Author & year Methodology Pros Cons

Sergiu Cosmin Can classify tweets with an

Nistor et al. 2021 RNN (Recurrent Neural Network) 80.74% accuracy rate, Overfitting
[16] considering a binary task

Rakin Mostafa et NLP-BiLSTM (Natural Language Performance of word embedding Lower classification and
al. 2022 [15] Processing and Bidirectional Long Short techniques is good retrieval accuracy
Term Memory)
 3

Author & year Methodology Pros Cons

Abayomi Bello et BERT model (BERT-CNN, BERT-RNN, Comprehensive approach Lower classification
al. 2023 [5] BERT-BiLSTM) combining BERT with other accuracy
techniques

Munir Ahmad et al. SVM (Support Vector Machine) Able to build a baseline ML Not tested on large dataset
2017 [1] model

Lokesh Mandloi et Naive Bayes, SVM and Maximum Naive Bayes as baseline model Lower performance
al. 2020 [13] Entropy Method over SVM

E. Prabhakar et al. Improved Adaboost approach Construct a robust pipeline using Lower classification
2019[8] basic machine learning models accuracy

3. METHODOLOGY emoji/emoticons), pre-processing the data is a required step.

In order to identify the user’s sentiment and analyse it Below are the steps taken to ensure that noise and unwanted
effectively, our proposed work makes comparison of three deep data is removed from the raw tweets.
learning models. We pre-process the tweets from the dataset
efficiently, eventually converting into sequence of integers  Removal of links: Using Regular expression,
using tokenizer. The output from the tokenizer is split into train, hyperlinks, urls and user mentions can be removed
test datasets and is the input to the model. from the raw tweets
 Removal of emoticons: The tweets are processed for
3.1 Dataset removal of emojis, symbols, pictographs, transport
The dataset used for our proposed work was obtained from and map symbols using their respective unicodes.
Kaggle Repository [19] consisting of 1 million tweets collected  Removal of html related content: Tweets might
from Twitter randomly. The dataset titled “Sentiment Dataset sometimes contain html embeddings to render image
with 1 Million Tweets” is mainly recognized as a sentiment or render some iframe content. All of the html related
dataset and has been annotated mainly with labels (positive, content are filtered out in this step of process
negative, uncertainty, litigious). The proposed work mainly
 Removal of Stop Words: In order to remove stopwords
deals with positive and negative tweets from the dataset with a
(“the '', "a ","an ") from the tweets, python library
total of 5,26,765 tweets in English language. Statistical analysis
“nltk” is used. Its dictionary of English stop words are
of the dataset is shown in Table 1: Statistical Data from the
used to remove all the stop words which don’t
Dataset.
contribute to the meaning of the sentence
After all these processes are performed, the raw tweets are
Description Positive Negative converted into cleaned sentences. It is then split into words and
Tweets Tweets all of the individual letters are transformed into lowercase for
further processing. The most high frequency words represented
Number of Tweets 262220 264545 in the word cloud is shown in Fig 1.1 and 1.2

Mean Tweet length 23.52 26.06

(in words)
Table 1: Statistical Data from the Dataset

3.2 Data Preprocessing

Since most of the tweets contained in the dataset have unfiltered

information (urls, hashtags, user mentions and
Figure 1.1- High Frequency Positive Words
 4

c = σ ( W.X
i + b) ∈ R
(i:i+ j−1) (1)
In the above expression, σ represents the convolution operator,
X(i:i+ j−1) denotes the concatenation of word embeddings from
ith word to (i+j-1)th word. B will be the bias for W, W be the
weight vector. σ denotes the rectified Linear unit (ReLU)
activation function. Convolutional operation is performed on
the whole text X, for obtaining the feature map.

c = [ c , c ,..., c ,..., c ] ∈ R
1 2 i n− j+1
n-j+1
(2)
The max-pooling layer is then utilised for obtaining the
significant features from the feature map. The output produced
Figure 1.2 : High Frequency Negative Words from the convolution layer is then passed into the softmax-
activation function to predict the final output of the model.
Lemmatization: The main idea of this technique is to reduce the
word into its root form in order to simplify the meaning of the Using Convolutional Neural Networks (CNNs) for Twitter
actual sentence. The words from the previous process is sentiment analysis has its limitations and potential
lemmatized and is reduced to its root form and it then passed to shortcomings. CNN algorithms frequently miss the sequential
the tokenizer context of the data, but they are more effective at identifying
local patterns in text categorization. When it comes to sentiment
analysis, Transformer-based models like Recurrent Neural
Tokenization: In this pre-processing step, vectorizing into a
Networks (RNNs) do better at collecting contextual information
sequence of integers using the text corpus is executed.
and longer-range relationships that are necessary to grasp the
Lemmatized words are then treated with the tokenization
meaning of a tweet. CNNs often have trouble constructing good
process. The tokenizer is first fitted on the cleaned text and then
generalisations from Twitter data since it might be noisy and
padded up to a certain max length.
sparse. To function well, CNNs require a large volume of
training data. CNNs are typically seen to be less interpretable
than models such as RNNs. Interpretability is important in
sentiment analysis because you may wish to understand why a
certain sentiment forecast was made. RNNs provide a more
intuitive layer-by-layer interpretation since you can track how
information travels across the network over time.

Recurrent Neural Networks (RNN):

Figure 2: Flow of the Pre-processing
An RNN (Recurrent Neural Network) is a neural network
designed for processing sequential data or time series data.
Fig.2 shows the overall flow diagram of pre- Unlike feedforward networks, RNNs possess a limited memory
processing starting from the raw tweets until it undergoes the capacity, which means that the current output at a specific time
tokenization process. The processed data is split into ratios 80- step is dependent on the previous inputs as well. This makes
10-10 % for training, testing and validation datasets. RNNs an ideal tool for handling sequential data or data where
order is significant.

3.3 Architecture In the proposed model, LSTM and Bi-directional LSTM are
utilised to implement RNN in the architecture. The long-term
Convolution Neural Networks (CNN): dependency problem in RNNs is solved through the use of
Convolutional Neural Networks (CNNs) are a set of neural LSTM. LSTM utilises a memory cell to store information over
networks widely employed in image processing, image a long period of time, and three gates (input, output, and forget)
classification, and speech recognition. While traditionally used to regulate the flow of information into and out of the cell.
in computer vision, CNNs [22] are gaining momentum in the
field of Natural Language Processing (NLP), showing The equations for the input gate, forget gate, and output gate are
promising results. The convolution layer obtains an input as follows:
sequence of t-words [t ,t ,t ,… ,t ] and is denoted as X
1 2 3 n

concatenated word vector where d represents the (3)

dimensions of the individual word. A convolution filter (4)
is applied on a window of j words to generate new (5)
features . The new features obtained
after performing convolution operations are denoted by c. → represents the input gate.
→ represents the forget gate.
 5

→ represents output gate.

→ represents sigmoid function.
→ denotes the weight for the respective gate(x) neurons.
→ denotes the output of the previous LSTM block (at
timestamp t - 1).
→ denotes the input at the current timestamp.
→ denotes the biases for the respective gates(x).
Figure 3: CNN + Bidirectional LSTM Model Architecture
The equation for the control gate is as follows:
The proposed model is a hybrid of Bidirectional LSTMs and
(5)
CNNs, as shown in the Fig.3. This model aims to mitigate the
where is the value of the control gate. problems of both CNNs and LSTMs, resulting in a more robust
model that can accurately predict the sentiment of tweets.
The equation for the current memory cell state is as follows:
(6) The model is composed of six layers: an embedding layer, a
where is the value of the current memory cell state. CNN layer, a max pooling layer, a Dropout layer, a
Bidirectional layer, and a dense layer. These layers work
Finally, the equation for the output of the LSTM block is as cohesively to predict the final output of the deep learning
follows: architecture. The embedding layer takes an input sentence
(7) converts into its embeddings and then passes it to the CNN
where is the output of the LSTM block. layers. The CNN part of the architecture extracts the high-level
features from the sequence of words [w , w , w ,…, w ],
1 2 3 n

These equations are used to calculate the values of the input associated with embedding vectors of dimension d which were
gate, forget gate, output gate, control gate, memory cell state, received from the embedding layer. Then the output is sent to
and output of the LSTM block at each time step. The LSTM the max pooling layer.
architecture is trained using backpropagation through time to
minimise the loss function and improve the accuracy of the The Max Pooling layer in this hybrid architecture helps by
model. downsampling the feature maps generated by the CNN, which
reduces computational complexity and retains the most crucial
Bidirectional Long Short-Term Memory (BiLSTM) [26] local features. This processed data is then fed into the LSTM
effectively addresses the limitations of traditional part of the architecture. This LSTM part of the model consists
unidirectional LSTMs by processing sequential data in both of Bidirectional LSTM layers in which both past and future
forward and backward directions simultaneously. Bidirectional contexts have been modelled by considering temporal
LSTM design consists of two unidirectional LSTMs that information flow in both directions that is forward and
process the sequence in both forward and backward directions. backward. This helps the model to understand the entire context
This design can be viewed as having two independent LSTM of the tweet and make accurate predictions. Finally, the
networks, one receiving the token sequence as it is and the other architecture also includes some dropout layers [21]which helps
receiving it in reverse order. This enables a more to overcome the overfitting problem caused due to the
comprehensive understanding of context, the modelling of complexity of the model.
long-range dependencies, and improved feature extraction,
making BiLSTMs particularly suitable for tasks where word 4. EXPERIMENTAL RESULTS
order and context in both directions are critical, such as In this step, our proposed work evaluates results from three
sentiment analysis and language understanding. different deep learning models i.e CNN, LSTM and hybrid of
CNN and Bidirectional LSM. The results were compiled using
LSTM networks have a more complex architecture compared Google Collab’s GPU. For the training purpose, the training
to CNNs, with more gates and parameters. This complexity can data consist of 80%, testing data 10 % and validation data 10 %
make them computationally expensive and more challenging to of the original dataset. We have observed that the ensemble
train and tune. Due to their complex architecture, LSTMs are model (CNN+Bidirectional LSTM) model has achieved an
often prone to overfitting which causes the LSTM model to accuracy of 98.16 %. But other two deep learning models are
perform exceptionally well on the training data but significantly quite close to the accuracy model with CNN with an accuracy
worse on new data, where its performance is not as robust due of 92.42 % and Bidirectional LSTM model with an accuracy of
to its tendency to capture noise and peculiarities from the 91.26 %. The performance metrics used to evaluate the different
training dataset. Training LSTMs is inherently sequential and models are accuracy, precision, recall and f1-score as shown in
less suited for parallel processing. This can result in slower Table 2.
training and inference times compared to highly parallelized
CNNs. Fig.3 shows our proposed Model Architecture with
CNN and Bidirectional LSTM layers
 6

Model Accuracy Precision Recall F1-score 5. CONCLUSION

CNN 92.42 92.31 92.14 92.63 In today's day and age, understanding and analysing the user’s
Bidirectional
LSTM
91.26 91.45 91.38 91.63 sentiment is crucial for determining the trajectory of businesses,
CNN + organisation and policy making. Our proposed work highlights
Bidirectional 98.16 98.16 98.16 98.16 the significance of using a hybrid deep learning model to
LSTM evaluate user’s perception. The CNN and Bidirectional LSTM
model having an accuracy of 98% has proved to capture the
Table 2- Classification report for proposed models local and the global contextual information. With the model’s
high accuracy, it can be highly effective in monitoring content
The accuracy during training of the model over the number of on social media and determining user behaviour from his
epochs is shown in Fig 4.1 and Fig 4.2 shows the loss curve current activities
over epochs. We can observe that the difference between the Future works can include having a customised dataset for
training and validation accuracy curves is not too large, training which solely focuses on vocabulary relating to positive
indicating that there is some but not major over fitting. Fig 4.3 and negative sentiment words which results in better classifying
shows the confusion matrix which confirms the ensemble the data into categories. Multilingual tweets can be trained to
model’s performance metrics predict sentiments in different languages. As an extra step in
pre-processing, emoji’s can be expressed to correlate their
literal meaning. The same process can be applied to slang
language and sarcastic language.

Abbreviations Definition

RNN Recurrent Neural Network

CNN Convolutional Neural Network

LSTM Long Short-Term Memory

Fig 4.1 Accuracy of training and validation set over epoch NLP Natural Language Processing

DL Deep learning

SVM Support Vector Machine

BERT Bidirectional Encoder Representations

from Transformers

BiLSTM Bidirectional Long Short-Term Memory

REFERENCES

Fig 4.2 Loss on training and validation set over epoch

[1] Ahmad, M., Aftab, S., & Ali, I. (2017). Sentiment Analysis
of Tweets using SVM. International Journal of Computer
Applications, 177.
[2] Ahmad, M., Aftab, S., Muhammad, S. S., & Ahmad, S.
(2017). Machine Learning Techniques for Sentiment
Analysis: A Review. International Journal of
Multidisciplinary Sciences and Engineering, 8(3).
[3] Alamoudi, E. S., & Alghamdi, N. S. (2021). Sentiment
classification and aspect-based sentiment analysis on yelp
reviews using deep learning and word embeddings.
Journal of Decision Systems, 30, 259-281.
[4] Antonakaki, D., Fragopoulou, P., & Ioannidis, S. (2020). A
survey of Twitter research: Data model, graph structure,
Fig 4.3 Confusion matrix of the ensemble model sentiment analysis and attacks. Expert Systems with
Applications, 164.
[5] Bello, A., Ng, S.-C., & Leung, M.-F. (2022). A BERT
Framework to Sentiment Analysis of Tweets. Sensors.
 7

[6] D, G., & S, K. (2019). Evaluation of Deep Learning prevent neural networks from overfitting. Journal of
Techniques in Sentiment Analysis from Twitter Data,. machine learning research, 1929-1958.
2019 International Conference on Deep Learning and [22] Yadav, N., Kudale, O., Rao, A., Gupta, S., & Shitole, A.
Machine Learning in Emerging Applications (Deep-ML), (2020). Twitter Sentiment Analysis using Supervised
12-17. Machine Learning. Intelligent Data Communication
[7] Drus, Z., & Haliyana, K. (2019). Sentiment Analysis in Technologies and Internet of Things, 57, 631–642.
Social Media and Its Application: Systematic Literature [23] Zeglen, E., & Rosendale, J. (2018). Increasing Online
Review. Procedia Computer Science, 161, 707-714. Information Retention: Analysing the Effects of Visual
[8] E, P., M, S., Hari Krishnan, A., Kumar, T., & Sudhakar, R. Hints and Feedback in Educational Games. Journal of
(2019). Sentiment Analysis of US Airline Twitter Data Open, Flexible and Distance Learning,.
using New Adaboost Approach. International Journal of [24] Zeng, D., Liu, K., Lai, S., Zhou, G., & Zhao, J. (2014).
Engineering Research & Technology. Relation Classification via Convolutional Deep Neural
[9] Gupta, B., Negi, M., Vishwakarma, K., Rawat, G., & Network. Proceedings of COLING 2014, the 25th
Badhani, P. (2017). Study of Twitter Sentiment Analysis International Conference on Computational Linguistics:
using Machine Learning Algorithms on Python. Technical Papers, 2335–2344.
International Journal of Computer Applications, 165. [25] Zhang, L., Wang, S., & Liu, B. (2018). Deep Learning for
[10] Jain, P., Pamula, R., & Srivastava, G. (2021). A systematic Sentiment Analysis : A Survey. arXiv:1801.07883.
literature review on machine learning applications for [26] Zhang, S., Zheng, D., Hu, X., & Yang, M. (2015).
consumer sentiment analysis using online reviews. Bidirectional Long Short-Term Memory Networks for
Computer Science Review, 41. Relation Classification. 29th Pacific Asia Conference on
[11] Jindal, S., Tyagi, A. K., & Sharma, K. (2020). Twitter Language, Information and Computation, 73 - 78.
user’s Behavior and Events Affecting Their Mood.
International Journal of Engineering and Advanced
Technology (IJEAT), 9(3). AUTHORS
[12] Karami, A., Lundy, M., Weeb, F., & Triwedi, Y. (2020).
Twitter and Research: A Systematic Literature Review First Author – Adish Padalia, currently pursing B.Tech
Through Text Mining. IEEE Access, 8, 67698-67717. Information Technology Engineering at Sardar Patel Institute
[13] Mandloi, L., & Patel, R. (2020). Twitter Sentiments of Technology. (Email - [email protected])
Analysis Using Machine Learning Methods. 2020 Second Author – Pratik Pujari, currently pursing B.Tech
International Conference for Emerging Technology Computer Engineering at Sardar Patel Institute of Technology
(INCET). (Email: [email protected]).
[14] Medhat, W., Hassan, A., & Korashy, H. (2014). Sentiment Third Author – Tanish Shah, currently pursing B.Tech
analysis algorithms and applications: A survey. Ain Shams Information Technology Engineering at Sardar Patel Institute
Engineering Journal. of Technology (Email - [email protected]).
[15] Mostafa, R., Mehedi, M., Md, A., & Rasel, A. (2022).
Bidirectional LSTM and NLP based Sentiment Analysis of Correspondence Author – Dr. Kailas K. Devadkar, Ph. D.
Tweets. (Technology) in Computer Engineering Associate Professor
[16] Nistor, S. C., Moca, M., Moldovan, D., Oprean, D. B., & Coordinator ERP S.P.I.T, Department of Information
Nistor, R. L. (2021). Building a Twitter Sentiment Analysis Technology Sardar Patel Institute of Technology
System with Recurrent Neural Networks. Sensors by (Email - [email protected])
MDPI, 21.
[17] Osatuyi, B. (2013). Information sharing on social media
sites. Computers in Human Behavior, 29(6), 2622-2631.
[18] Qian, Y., Deng, X., Ye, Q., Ma, B., & Yuan, H. (2019). On
detecting business event from the headlines and leads of
massive online news articles. Information Processing &
Management, 56(6).
[19] Sentiment Dataset with 1 Million Tweets. (n.d.). Retrieved
October 15, 2023, from Kaggle:
https://www.kaggle.com/datasets/tariqsays/sentiment-
dataset-with-1-million-tweets
[20] Severyn, A., & Moschitti, A. (2015). Twitter Sentiment
Analysis with Deep Convolutional Neural Networks.
SIGIR '15: Proceedings of the 38th International ACM
SIGIR Conference on Research and Development in
Information Retrieval.
[21] Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I.,
& Salakhutdinov, R. (2014). Dropout: a simple way to