A Minor Project Report

on

Deep Learning Techniques for Sentiment

Analysis on Social Media Text
submitted in partial fulfillment of the requirements for the award of degree

Bachelor of Technology
in
Computer Science and Engineering
by

21211A05Z0 Yanala Ashrita

21211A05V1 Sudhagoni Abhiram
21211A05W7 Vadranapu Hemanth
22215A0528 Adunuri Srinivas

Under the kind guidance of

Mr.V Punna Rao
Laboratory: Artificial Intelligence and Machine Learning

Department of Computer Science and Engineering

B V Raju Institute of Technology
(UGC Autonomous, Accrediated by NBA and NAAC)
Vishnupur,Narsapur, Medak(Dist), Telangana State, India-502313.
 CERTIFICATE OF APPROVAL

This project work (E2) entitled ” Deep Learning Techniques for Sentiment Analysis on
Social Media Text ” by Ms. Yanala Ashrita, Registration No. 21211A05Z0 under the su-
pervision of Mr. V Punna Rao in the Department of Computer Science and Engineering, B
V Raju Institute of Technology, Narsapur, is hereby submitted for the partial fulfillment of
completing Minor Project during II B.Tech II Semester (2022 - 2023 EVEN). This report
has been accepted by Research Domain Computational Intelligence and forwarded to the
Controller of Examination, B V Raju Institute of Technology, also submitted to Department
Special Lab ” Artificial Intelligence Machine Learning” for the further procedures.

V. Punna Rao Dr.CH.Madhu Babu

Assistant Professor and Supervisor Professor and Dept.Head
Department of CSE Department of CSE
B V Raju Institute of Technology B V Raju Institute of Technology
Narsapur. Narsapur

External Examiner Internal Examiner

Dr. Y C A Padmanabha Reddy

Lab Incharge - Artificial Intelligence Machine Learning Laboratory
Department of Computer Science and Engineering
B V Raju Institute of Technology
Narsapur

i
DECLARATION

We, the members of Research Group domain Computational Intelligence under Artifi-
cial Intelligence Machine Learning special lab , declare that this report titled: Deep Learn-
ing Techniques for Sentiment Analysis on Social Media Text is our original work and has
been submitted in whole or in parts for International conference or journal International
Conference for Contemporary Computing and Informatics (IC3I-2023). All sources of
information used in this report have been acknowledged and referenced respectively.

This project was undertaken as a requirement for the completion of our II B.Tech II
Sem Minor project in Department of Computer Science and Engineering at B V Raju
Institute of Technology, Narsapur. The project was carried out between 31-March-2023
and 11-August-2023. During this time, we as a team were responsible for the process model
selection, development of the micro document and designing of the project.

The Bidirectional LSTM is designed to perform text classification tasks, categorizing tex-
tual data into predefined categories using natural language processing and deep learning
techniques. It facilitates efficient data preprocessing, model training, and inference. The
code aims to enhance decision-making processes, automate content categorization, and ex-
tract valuable insights from textual information. It offers versatility, accuracy, and scalability,
making it a valuable tool for researchers, data scientists, and developers in various domains.

We would like to express our gratitude to our project supervisor Mr. V. Punna Rao for his
guidance and support throughout this project. We would also like to thank our Department
Head Dr.CH.Madhu Babu for his help and efforts.

We declare that this report represents Our own work, and any assistance received from
others has been acknowledged and appropriately referenced.

Yanala Ashrita (21211A05Z0)

Sudhagoni Abhiram (21211A05V1)

Vadranapu Hemanth (21211A05W7)

Adunuri Srinivas (22215A0528)

Supervisor Project Coordinator LAB Incharge HOD/CSE

ii
ACKNOWLEDGEMENT

This project is prepared in fulfilment of the requirement for the Artificial Intelligence
Machine Learning Lab under Research Domain Computational Intelligence. We owe our
deepest gratitude to the Department of Computer Science and Engineering
B V Raju Institute of Technology, Narsapur for providing us with an opportunity to work
on this project.

We also extend our gratitude towards our project supervisor Mr.V.Punna Rao and De-
partment Head Dr.CH.Madhu Babu , whose guidance and expertise have been invaluable
in steering us towards success. We also thank other faculty members of the Department for
their valuable feedback and suggestions.

Finally, we would like to thank our family and friends for their continuous support and
encouragement throughout the project. We acknowledge the contributions of everyone who
supported us in the creation of this project report.

Thank you all for your assistance and support.

The experience of working on this project will surely enrich our technical knowledge and
also give us hands on experience of working on a project and help develop our team’s skill
set to a great extent.

iii
 ABSTRACT

Social media enables individuals to express their opinions on a range of contacts with the
outside world and topics that interest them. Many people share their thoughts or opinions
through text, images, audio, and video. As a result, networking platforms cause a significant
volume of unstructured data to be produced on the Internet. Sentiment analysis is a tool
that can be used to quickly evaluate data in order to understand human psychology. When
compared to current feature-based techniques, deep learning models for sentiment analysis
offer detailed representation capabilities and improved performance. The automatic process
of sentiment analysis determines the author’s attitude and determines whether it is neutral,
positive, or negative. Analysis of these emotions without consideration for voice and facial
expressions is crucial and necessitates a supervisory technique for accurate emotion inter-
pretation. Despite these difficulties, it is important to recognise human emotions as they
increasingly communicate via social media platforms like Facebook, Twitter, etc. This arti-
cle examines the difficulties researchers have had studying sentiment analysis in social media
as well as other potential issues, and it introduces methods that have been tested through a
series of experiments using real data-sets and show significant advantages over other ap-
proaches currently in use for classifying text in social media.

Keywords: Sentimental Analysis, Natural Language Processing, Deep Learning, Supervisory Tech-
niques, Emotion Interpretation

iv
List of Figures
4.1 Use Case Diagram for Text Classification using BiLSTM Model. . . . . . . . . . . . . . 13
4.2 Data Flow Diagram Level 0 for Text Classification using BiLSTM Model. . . . . . . . . 14
4.3 Data Flow Diagram Level 1 for Text Classification using BiLSTM Model. . . . . . . . . 15
4.4 Data Flow Diagram Level 2 for Text Classification using BiLSTM Model. . . . . . . . . 16
4.5 Class Diagram for Text Classification using BiLSTM Model. . . . . . . . . . . . . . . . 17
4.6 Sequence Diagram for Text Classification using BiLSTM Model. . . . . . . . . . . . . 18
4.7 Activity Diagram for Text Classification using BiLSTM Model. . . . . . . . . . . . . . 19
4.8 State Chart Diagram for Text Classification using BiLSTM Model. . . . . . . . . . . . . 20
5.1 System Block Diagram for Text Classification using BiLSTM Model. . . . . . . . . . . 22
6.1 System Architecture Diagram for Text Classification using BiLSTM Model. . . . . . . . 24
7.1 Gann Chart of Project Timeline. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
7.2 Results showing accuracy and validation accuracy across epochs. . . . . . . . . . . . . 27

v
List of Tables
A.1 Deep Learning Techniques for Sentiment Analysis on social Media Text project timeline 34

vi
 LIST OF ACRONYMS AND ABBREVIATIONS

LSTM Long-Short Term Memory

BiLSTM Bidirectional Long-Short Term Memory

NLTK Natural Language Toolkit

NLP Natural Language Processing

TF-IDF Term Frequency - Inverse Document Frequency

vii
TABLE OF CONTENTS
CERTIFICATE OF APPROVAL i

DECLARATION ii

ACKNOWLEDGEMENT iii

ABSTRACT iv

LIST OF FIGURES v

LIST OF TABLES vi

LIST OF ACRONYMS AND ABBREVIATIONS vii

1 INTRODUCTION 1
1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.3 Problem statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.4 Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.5 Scope of Project . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

2 LITERATURE SURVEY 4

3 REQUIREMENT ANALYSIS & SPECIFICATION 1

3.1 Feasibility Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
3.1.1 Market Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
3.1.2 Technology Assessment . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
3.1.3 Operational Requirements . . . . . . . . . . . . . . . . . . . . . . . . . . 2
3.1.4 Financial Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
3.1.5 Risk Assessment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
3.2 Selection of Process Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
3.2.1 Process Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
3.2.2 Why Agile . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
3.2.3 Why Not . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
3.3 Software Requirements Specification . . . . . . . . . . . . . . . . . . . . . . . . . 5
3.4 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
3.4.1 Purpose . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
3.4.2 Scope . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
3.4.3 Definitions, Acronyms and Abbreviations . . . . . . . . . . . . . . . . . . 5
3.4.4 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
3.5 Overall Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
3.5.1 Product Perspective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

viii
 3.5.2 Product Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
3.5.3 User Characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
3.5.4 Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
3.5.5 Assumptions & Dependencies . . . . . . . . . . . . . . . . . . . . . . . . 8
3.6 Specific Requirements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
3.6.1 External Interfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
3.6.2 Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
3.6.3 Performance Requirements . . . . . . . . . . . . . . . . . . . . . . . . . . 9
3.6.4 Logical Database Requirements . . . . . . . . . . . . . . . . . . . . . . . 9
3.6.5 Design Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
3.6.6 Software System Quality Attributes . . . . . . . . . . . . . . . . . . . . . 10
3.6.7 Object-Oriented Models . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.6.8 Appendices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

4 DESIGN SPECIFICATION 12
4.1 Use Case Diagram . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
4.2 Data Flow Diagram . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
4.2.1 DFD LEVEL 0: . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
4.2.2 DFD LEVEL 1: . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
4.2.3 DFD LEVEL 2: . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
4.3 Class Diagram . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
4.4 Sequence Diagram . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
4.5 Activity Diagram . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
4.6 State Chart Diagram . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

5 METHODOLOGY 21
5.1 Modules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
5.2 System Block Diagram . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

6 IMPLEMENTATION DETAILS 23
6.1 Technology Stack . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
6.2 System Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
6.3 User Interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
6.4 Integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
6.5 Security . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
6.6 Testing and Deployment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

7 OBSERVATIONS 26
7.1 Time Domain - Gann Chart . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
7.2 Results and Comparitive Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

8 CONCLUSION 28

ix
9 LIMITATIONS AND FUTURE ENHANCEMENTS 29
9.1 Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
9.2 Future Enhancements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

A APPENDIX 32
A.1 Project Timeline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
A.2 Coding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

x
 CHAPTER - 1

1. INTRODUCTION

In this report, we present a comprehensive Python script for text classification using a combination of
natural language processing techniques and deep learning models.

The primary goal of this project is to classify text data into predefined categories or labels. To achieve
this, we have employed a multi-step approach that includes data preprocessing, feature extraction us-
ing TF-IDF (Term Frequency-Inverse Document Frequency), and the development of a deep learning
model for text classification. The project utilizes popular libraries and frameworks such as Tensor-
Flow, scikit-learn, and NLTK (Natural Language Toolkit).

1.1. Background

In today’s data-driven world, vast amounts of textual information are generated daily, creating a
pressing need for efficient text classification. Text classification involves categorizing text data into
predefined groups, enabling automated organization, analysis, and decision-making. Historically,
rule-based and statistical methods fell short in handling the complexity of language. However, the
emergence of deep learning and neural networks has revolutionized text classification. This project
is grounded in the convergence of escalating demand for accurate text classification and the capabil-
ities of deep learning. By offering a Python-based framework that combines preprocessing, feature
extraction, and deep learning, this code empowers users to effectively address a wide range of text
classification challenges [? ].

1.2. Motivation

This project is inspired by the urgent need to harness potential of the textual data that is constantly
produced nowadays. The motivation here is to automate the critical task of text classification, a funda-
mental cornerstone in data-driven decision-making. Whether it’s organizing information, improving
user experiences, empowering research, or enabling data-driven business strategies, the power of text
classification cannot be overstated. By providing a comprehensive codebase for this purpose, this
project aims to equip users across domains with the tools to navigate the information deluge effi-
ciently, unlocking valuable insights and facilitating innovation in natural language processing. This
code offers practical solutions for harnessing the power of text data across diverse domains.

1
1.3. Problem statement

• Networking technologies generate enormous amount of unstructured data, making it challeng-

ing to concentrate on text’s primary ideas.

• The ambiguity present in the data make it challenging for the algorithm to categorize the text
and make reliable predictions.

• Complex data and long-term dependencies are difficult to model with typical deep learning
models, so an efficient algorithm is needed to address these issues.

1.4. Objectives

1. To make the model more accurate and tackle massive amounts of data with very less human
interaction.

2. To examine the Long Short-Term Memory (LSTM) methodology and incorporate new features
to increase its effectiveness.

3. To analyse the results obtained and on comparing with the existing model and to state the
more accurate algorithm that can generalize to new, unseen text data and accurately classify
the sentiment expressed in the text.

4. To create a reliable and efficient system that can automatically classify large volumes of text
data for various applications.

1.5. Scope of Project

To gain an insight into the scope of our project, let us first understand what the term scope of a project
means in software development, the scope refers to the boundaries and limitations of a project. It also
defines the features, functions, and requirements of the software being developed. The scope of a
software project describes what the software will and will not do, and what is included and excluded
from the project.

The BiLSTM Model for Text Sentiment Analysis will consist of two primary functionalities:

The project aims to deliver the following key features:

• Data Preprocessing: The project covers data preprocessing tasks such as HTML tag removal,
special character elimination, and text stemming. These preprocessing steps ensure that the
text data is clean and ready for feature extraction.

2
 • Feature Extraction: Text data is transformed into numerical features using TF-IDF vectoriza-
tion. This project demonstrates the creation of a TF-IDF matrix, which represents the impor-
tance of words in the corpus. It allows users to convert textual information into a machine-
readable format.

• Deep Learning Model: The core of the project revolves around the development of a deep
learning model for text classification. The model architecture includes dense layers, bidirec-
tional LSTM layers, and softmax activation for multi-class classification. TensorFlow’s Keras
API is employed for model building and training.

The boundaries and limitations of BiLSTM model Text Classification can also be defined by its scope,
which also outlines the features, functions, and requirements of the system. Some of these boundaries
and limitations are:

• Data Dependency: The effectiveness of the text classification model heavily relies on the
quality and quantity of the training data. Limited or biased training data may lead to suboptimal
results or bias in the model’s predictions

• Computational Resources:Training deep learning models can be resource-intensive, and this

project assumes access to a machine with sufficient computational power. For larger datasets
or more complex models, distributed computing or cloud resources may be necessary.

• Language Specificity: The code and model developed in this project are tailored to English
text data. Adapting it to other languages may require additional preprocessing and language-
specific considerations.

Addressing these limitations through ongoing research, collaboration with dental professionals, and
continuous model refinement will be essential for maximizing the model’s potential impact on Senti-
ment Analysis and Text Classification.

Overall, the project report will provide a comprehensive understanding of the Bidirectional LSTM.
The report will serve as a valuable resource for researchers interested in Sentiment Analysis of Texual
Data.

3
 CHAPTER - 2

2. LITERATURE SURVEY

Using several machine learning and deep learning techniques, the data will be categorised according
to the feelings in a Multi Class Sentiment Analysis [1]. Positive and Negative polarities or classes,
previously known as Binary Classification, were used to categorise the sentiment. It would offer
several sectors, such as universities or businesses, to assess the users ideas in accordance with their
network. The authors propose a hybrid approach that combines lexiconbased and machine learning
methods to improve sentiment classification accuracy.The authors in [2] claim that social networking
platforms, e-commerce websites, and movie review websites are good places to gather opinions. They
have proposed a novel method that makes use of lexicons that are based on the semantic similarity
between text terms and lexicon word stock, worked with real samples and experimented with various
feature set thresholds.Through the use of several statistical techniques, the feature extraction is eval-
uated. The major objective of this is to merge embedding based representation with semantic feature
extraction.

A study of how social media affects the user’s own well-being is of utmost relevance because the
influence of online communities today influences money, governments, relationships, and many other
facets of daily life. [3] highlights the necessity to examine and comprehend people’s well being
using information on their own online interactions. At the aspect level, sentiment analysis is used
to ascertain opinions about a certain aspect or feature. Based on the frequency of co-occurrence of
an aspect with a feature indicator and by taking advantage of the link between opinionated language
and explicit aspects, there is an algorithm to extract implicit aspects from documents. The explicit
positioning of context in an opinionated statement and the many polarity of the several aspects in a
single sentence present two issues for aspect-level analysis. To address these problems, in [4] the
authors developed a two-stage LSTM model with an attention mechanism.

To conduct the sentimental analysis, a few models were considered in [5]. The method that is being
suggested in this paper focuses on classifying emotions from tweets utilising a domain-dependent
approach based on machine learning that uses a variety of textual properties.N-grams that incorpo-
rate data from Twitter. Additionally, they used three distinct weighting techniques to investigate how
weighing affected the classifier’s precision. Additionally, to improve the effectiveness a tweet senti-
ment score vector is utilised with an SVM classifier to provide additional information. The authors
sentiment analysis of social media texts in [6] contributes to a few predictions from business intelli-
gence. They processed the film information set utilizing the Twitter page’s reviews, comments, and
feedback using the Hadoop Framework. Using Twitter data, [7] Sani Kams¸ and Dionysis Goularas
assessed deep learning methods for sentiment analysis. Their study examined how different deep

4
learning models performed in sentiment categorization tasks. The study demonstrated how well deep
learning computers can extract complex sentiment from brief social media texts.

A thorough evaluation and analysis of the research on sentiment analysis in social media is pro-
vided by [8]. The study discusses the goals of sentiment analysis tasks, the overall implementation
procedure, the algorithms used, and how they are employed in various domains using the PRISMA
framework.Deep learning methods for sentiment analysis were compared by S. P. I. Saripalli et al [9].
The performance of different deep learning models, such as CNN, LSTM, and Transformer-based
models, was assessed by the authors when it came to sentiment categorization tasks. An anonymous
realtime analytics monitoring solution supported by sentiment analysis was presented by [10] G.A.
de Oliveira Jnior et al. Their study attempted to use sentiment analysis on social media data to assist
decision-making processes. To enhance sentiment analysis, hybrid deep learning models were pro-
posed by P. Akula et al. in [11]. Their work combined various deep learning architectures, including
CNN and LSTM, to improve the accuracy of sentiment classification. In order to improve sentiment
analysis performance, the study highlighted the possibility of merging various deep learning models.

Using the TexBlob package, the author proposed a sentiment analysis method that draws on the
concepts of Bag of Words (BOW), Part of Speech (POS), and Natural Language Processing (NLP)
in order to analyse suicidal patterns in blogs and tweets. The opinions, expectations, and feelings
of the tweeters about the US reopening during the COVID19 epidemic were examined in [12]. To
determine people’s perspectives on the pandemic, sentiment analysis was used to tweets posted by the
World Health Organisation (WHO) and the people in general. In the midst of COVID-19 epidemic
[13], N.K. Rajput et al. performed a word frequency and mood analysis of tweets. The authors
analyzed public opinion using data from Twitter to find frequently used terms associated with the
pandemic. Using a deep learning approach [14], T. Kaur and P. Sandhu concentrated on sentiment
analysis of COVID19 tweets. In their research, deep learning techniques were employed to examine
the sentiments posted on Twitter during the pandemic.

The relevant state-of-the-artwork is reviewed and expanded upon in this publication [15]. A thorough
explanation of deep neural networks, including RNN, LSTM, BiLSTM, is also provided. The authors
intend to deploy transformerbased neural language models, such as BERT and RoBERTa, in the fu-
ture to enhance the performance of the deep neural networks they have already used. To categorise
tweets into positive, negative, or neutral sentiment categories, the suggested method employs a com-
bination of convolutional neural networks (CNNs) and long short-term memory (LSTM) networks.
The proposed method uses a multihead self-attention mechanism to capture the important features
of the text and CNN network to classify the sentiment of the text [16]. Deep learning models were
used by Arnab Roy and Muneendra Ojha [17] to analyze sentiment on Twitter. In order to classify
sentiment, their study used LSTM networks and CNN.

Using deep learning methods, A. K. Yadav et al in [18] concentrated on sentiment analysis of Hindi
tweets. The authors made a contribution to the study of sentiment analysis in multilingual contexts
by using deep learning algorithms to examine the sentiments conveyed in Hindi tweets. In [19], deep

5
learning models such as LSTM and RNN are employed to classify human emotion in Twitter postings.
Each word from the input tweets was segregated into its own semantic word vector from the lexical
words using the word2vec technique. To improve sentiment categorization performance, the study
employs an improved methodology for word embedding and the establishment of a long short-term
memory (LSTM) network. In [20], using word embedding and an LSTM network that distinguishes
suffixes and prefixes from training data, they created a strong framework. A number of experiments
were performend utilising the dataset of Twitter coronavirus hashtags as well as datasets of open
reviews from Amazon and Yelp. An analysis of statistical data shows that these proposed models
perform better than other models in terms of categorization accuracy. Complex data and long-term
dependencies are difficult to model with typical deep learning models, so to solve these problems, an
effective algorithm is required.

6
 CHAPTER - 3

3. REQUIREMENT ANALYSIS & SPECIFICATION

This section of the project report is the most critical element as, it provides a foundation for the entire
project. It ensures that all stakeholders are aligned on the project’s objectives, scope, and deliverables,
and it provides a clear road map for the project team to follow.
This section is further divided into 3 sister sections:

• Feasibility Study

• Model Selection

• SRS

3.1. Feasibility Study

A feasibility study is the first stepping stone into the development of any project, including our BiL-
STM Model to perform Sentiment Analysis on Social Media Text. It involves assessing the potential
for the project to be successful, which in turn includes evaluating the market, technology, financial
aspects, and operational requirements.

3.1.1. Market Analysis

The market for text classification solutions is substantial and continues to grow rapidly. Businesses
across industries seek automated text analysis to enhance customer experiences, streamline opera-
tions, and extract actionable insights from textual data. This project, offering an open-source and
adaptable text classification framework, caters to this demand. It targets a wide range of potential
users, including data scientists, developers, and researchers. The flexibility and scalability of the
code make it suitable for both small enterprises and large corporations, ensuring its relevance in a
competitive market where efficient text analysis is paramount for informed decision-making.

3.1.2. Technology Assessment

This project leverages cutting-edge technologies and libraries in the field of natural language pro-
cessing (NLP) and machine learning. Key components include TensorFlow and scikit-learn for deep

1
learning and feature extraction, NLTK for text preprocessing, and Python for flexibility and ease of
use. The use of TF-IDF for feature extraction and LSTM-based deep learning models reflects current
best practices in text classification. The project’s compatibility with standard computing resources
and its documentation makes it accessible to a wide range of users, ensuring that it aligns well with
current technology trends and user needs in the NLP and machine learning domain.

3.1.3. Operational Requirements

The operational requirements encompass a compatible Python environment with essential libraries
like TensorFlow, scikit-learn. It should operate on standard hardware with reasonable computa-
tional resources. Additionally, a stable internet connection may be needed for accessing external
data sources or models. The code must handle various text data formats (e.g., CSV, JSON) and pro-
vide clear user instructions for data input and model usage. Error handling and logging mechanisms
are essential for robust operation, ensuring that issues are captured, reported, and addressed promptly.

3.1.4. Financial Analysis

Financial analysis will provide crucial insights into the financial viability and potential profitability of
the BiLSTM Model in Text Classification.This project offers substantial cost advantages, primarily
owing to its open-source nature. Users can leverage the project without incurring licensing fees,
making it a cost-effective solution. However, users should consider potential expenses related to
computing resources for model training and storage for large datasets. Additionally, there may be
personnel costs associated with data collection, preprocessing, and customization. Overall, the project
provides a cost-efficient means to develop and deploy text classification solutions, particularly when
compared to proprietary alternatives in the market.

3.1.5. Risk Assessment

A comprehensive risk assessment for the BiLSTM model addresses the data quality that stands as a
primary concern, with the potential for inaccurate classifications if the training data is insufficient or
of poor quality. Resource constraints are another significant risk, particularly regarding the compu-
tational resources required for training deep learning models, which may pose financial challenges.
Complex classification tasks could demand more advanced model architectures, extending develop-
ment timelines. Maintenance risks involve the need for code updates due to evolving libraries. Over-
fitting remains a concern, especially with smaller datasets, and ensuring robust privacy and security
measures is essential to safeguard sensitive data and prevent potential breaches or misuse.

2
3.2. Selection of Process Model

The software life cycle process model is a framework that outlines the various stages involved in the
development of a software application. So, choosing a life cycle process model is the stepping stone
into the development of a software product.

3.2.1. Process Models

The Agile software development process model is highly suitable for this Deep Learning Project. Ag-
ile’s iterative and flexible approach aligns well with the dynamic nature of text classification tasks. It
allows for continuous development, adaptation, and collaboration among team members, data scien-
tists, and stakeholders. In text classification, where data quality, model complexity, and requirements
may change over time, Agile empowers the project to efficiently respond to these shifts, address data
issues, and deliver incremental value to users.

3.2.2. Why Agile

Here are some reasons why the Agile model is the best choice for developing teeth segmentation
model:

• Adaptability to Changing Requirements: The Agile model accommodates changing data

sources, evolving linguistic patterns, and the need for ongoing model updates. It enables teams
to respond effectively to data shifts and adapt the classification approach accordingly.

• Iterative Development: Agile encourages iterative development cycles, which align perfectly
with the gradual refinement and optimization required for text classification models. Teams can
continuously improve model performance

• Continuous Stakeholder Engagement: Agile emphasizes delivering small, working incre-

ments of the project. This approach allows stakeholders to provide feedback early and often,
helping to identify and address data quality issues and model performance concerns.

• Flexibility: Text classification tasks often involve dynamic and evolving requirements. Ag-
ile’s flexibility allows teams to adapt to changing needs, making it well-suited for the iterative
nature of machine learning model development.

3.2.3. Why Not

Every coin has two sides thus, we can’t forget to consider that the waterfall model has some limitations
too such as:

3
 • Scope Management

• Resource Intensive

While Agile offers many benefits, it may not be suitable for projects with fixed, rigid timelines or
highly regulated environments, where extensive documentation and upfront planning are mandatory.
In such cases, adopting Agile might require careful planning and adaptation to address the specific
needs and constraints of the project.

4
3.3. Software Requirements Specification

3.4. Introduction

3.4.1. Purpose

The purpose of this Document is to develop a comprehensive text classification solution using deep
learning techniques. It aims to enable efficient categorization of textual data into predefined classes
or categories. This project seeks to address the increasing demand for automated text analysis across
various domains, including sentiment analysis, content recommendation, and content moderation. By
providing a flexible and adaptable codebase, the project empowers users to leverage the power of
natural language processing and deep learning for accurate and scalable text classification, ultimately
enhancing decision-making processes and user experiences.

3.4.2. Scope

The Sentiment Analysis of Texual Data and BiLSTM model will include the following functionalities:

• Text Classification for Diverse Applications: The primary scope of the model is to provide a
versatile text classification framework. It enables users to apply text classification to a wide
range of applications, such as sentiment analysis, content categorization, spam detection, and
more. This adaptability allows the code to address the needs of various domains and industries.

• Emotion Recognition and Understanding: The model can focus on developing algorithms and
models to recognize and understand human emotions expressed in textual data. This includes
detecting sentiments, emotions, and mood from written text, such as social media posts, re-
views, or personal messages.

3.4.3. Definitions, Acronyms and Abbreviations

• Definitions:
LSTM: Long-Short Term Memory
BiLSTM: Bidirectional Long-Short Term Memory
SRS: Software Requirements Specification
NLTK: Natural Language Toolkit
TF-IDF Term Frequency - Inverse Document Frequency

5
 • Acronyms:
DL: Deep Learning
SA: Sentiment Analysis
API: Application Programming Interface
NLP: Natural Language Processing

• Abbreviations:
LSTM: Long-Short Term Memory
BiLSTM: Bidirectional Long-Short Term Memory
SRS: Software Requirements Specification
NLTK: Natural Language Toolkit
TF-IDF Term Frequency - Inverse Document Frequency

3.4.4. Overview

The document will mostly consist of two parts:

• Overall Description

• Specific Requirements

Overall description describes the major components of the system, assumptions and dependencies of
the system, while specific requirements describes the functions of the system and their roles in the
system and the constraints faced by the system.

3.5. Overall Description

3.5.1. Product Perspective

The product perspective of this deep learning model emphasizes its role as a versatile and user-friendly
tool for text classification. It’s designed to be modular and compatible with common libraries, ensur-
ing easy integration into existing workflows. Scalability and adaptability enable it to handle diverse
text classification tasks. Prioritizing performance, reliability, and security, the code offers a seamless
user experience with clear documentation and version control. This perspective positions your project
code as a valuable asset in the realm of text analysis and classification, catering to a wide range of
user needs and promoting long-term sustainability.

6
3.5.2. Product Functions

The main functions of the Sentiment Analysis of Texual Data and BiLSTM Model are as follows:

• Text Classification: The primary scope of the model is to provide a versatile text classification
framework. It enables users to apply text classification to a wide range of applications, such as
sentiment analysis, content categorization, spam detection, and more. This adaptability allows
the code to address the needs of various domains and industries.

• Emotion Recognition and Understanding: The model can focus on developing algorithms
and models to recognize and understand human emotions expressed in textual data. This in-
cludes detecting sentiments, emotions, and mood from written text, such as social media posts,
reviews, or personal messages.

3.5.3. User Characteristics

The user characteristics of the BiLSTM Model are diverse, spanning data scientists, machine learning
practitioners, developers, and researchers in various domains. Novice users may benefit from its
user-friendly documentation and ease of customization, while experts can leverage its flexibility and
adaptability to address complex text classification challenges. Users seeking efficient text analysis
solutions for tasks such as sentiment analysis, content categorization, and spam detection will find
value in this code, regardless of their level of expertise in natural language processing and deep
learning.

3.5.4. Constraints

The following constraints are considered for the development of the model:

• Data Quality: The model assumes the availability of high-quality, well-structured labeled data
for training and testing the text classification model. Poor data quality, including inaccuracies
in labeling or noisy data, can significantly impact the model’s performance and effectiveness.

• Computational Resources: While the model is designed to run on standard computing re-
sources, very large datasets or highly complex models may require more computational power
than what’s readily available on typical machines. This constraint needs to be considered when
working with extensive data or advanced model architectures.

7
3.5.5. Assumptions & Dependencies

The Text Classification and Deep Learning Model assumes the following:

• Quality Data Availability: The project assumes the availability of a sufficiently large and high-
quality labeled dataset for training and testing the text classification model. This dataset should
accurately represent the intended text classification task.

• Python Environment:It assumes users have access to a Python environment, with necessary
libraries such as TensorFlow, scikit-learn, NLTK, and matplotlib installed.

• Computational Resources: The project assumes access to standard computing resources for
model training and execution. However, it may require additional computational resources for
very large datasets or complex models.

3.6. Specific Requirements

3.6.1. External Interfaces

The Deep Learning Model for Text classification may include the following external interfaces:

• User Interface (if applicable): The model may offer user interfaces or command-line interfaces
(CLIs) for users to input text data or customize parameters for the classification task. User
interaction interfaces should be user-friendly and intuitive.

• Data Input Interface: The model shall acceot data from external data sources or APIs to fetch
text data for analysis. These sources could be social media platforms, websites, or databases
containing textual information relevant to the classification task.

3.6.2. Functions

The primary functions of the Sentiment Analysis of Texual Data and BiLSTM Model are:

• Text Classification: The primary scope of the model is to provide a versatile text classification
framework. It enables users to apply text classification to a wide range of applications, such as
sentiment analysis, content categorization, spam detection, and more. This adaptability allows
the code to address the needs of various domains and industries.

• Emotion Recognition and Understanding: The model can focus on developing algorithms
and models to recognize and understand human emotions expressed in textual data. This in-

8
 cludes detecting sentiments, emotions, and mood from written text, such as social media posts,
reviews, or personal messages.

3.6.3. Performance Requirements

The Text Classification and Deep Learing Model shall meet the following performance requirements:

• Model Accuracy: The primary performance metric is the accuracy of the text classification
model. It measures how well the model correctly predicts the class or category of text data.

• Loss Function: The loss function, such as categorical cross-entropy, is used during model
training to quantify how well the model is learning from the data. Lower loss values indicate
better learning.

3.6.4. Logical Database Requirements

The logical database requirements for this project involve designing a schema capable of storing text
data, labels, and metadata relevant to text classification. Ensuring data integrity through normaliza-
tion, implementing efficient indexing, and developing SQL queries for data retrieval and manipulation
are key considerations. Additionally, support for data partitioning, versioning, security, and backup
mechanisms is essential. These requirements aim to create a robust and scalable database system
capable of managing diverse text datasets while maintaining data quality and security.

3.6.5. Design Constraints

The design constraints of the Deep Learning Model for Text classification include:

• Model Size: The model should be designed with consideration for deployment on various
platforms, including devices with limited computational resources.

• Resource COnstraints : The code should be optimized to run efficiently on standard comput-
ing resources. Very resource-intensive models or tasks may require additional computational
power.

• Compatibility: The code must be compatible with specific versions of libraries, frameworks,
or operating systems. Deviations from these compatibility requirements may result in issues.

9
3.6.6. Software System Quality Attributes

The quality attributes of the Deep Learning Model for Text classification include:

• Accuracy: The model’s ability to accurately classify text data into predefined categories is
paramount for its effectiveness.

• Performance: This includes considerations of speed, scalability, and resource efficiency, en-
suring that the code performs well with both small and large datasets.

• Robustness: The model should be robust enough to handle variations in data quality, input
formats, and potential outliers without crashing or producing erroneous results.

• Usability: The model should be user-friendly, with clear documentation, easy-to-use inter-
faces, and customization options.

• Security: Ensuring that sensitive data is handled securely, including access control and data
encryption, is crucial, particularly in applications involving confidential information.

3.6.7. Object-Oriented Models

The Deep Learning Model for Text classification can be represented using object-oriented models
with the following key components:

• Text Classifier: Encapsulates the classification model and provides a method to classify text.

• Data Preprocessor: Handles text preprocessing tasks like cleaning, tokenization, and stem-
ming, with options for customizing the preprocessing steps.

• Model Evaluator: Responsible for evaluating the classification model’s performance using
specified metrics.

3.6.8. Appendices

1. Glossary

• Text Classification: The process of categorizing text documents into predefined categories or
classes based on their content.

• Tokenization: The process of splitting a text document into individual words or tokens for
further analysis.

10
• Stemming: The technique of reducing words to their base or root form to simplify text pro-
cessing and improve feature extraction.

• Stopwords: Common words (e.g., ”the,” ”and,” ”is”) that are often removed from text data
because they don’t carry significant meaning for classification.

• TF-IDF (Term Frequency-Inverse Document Frequency): A numerical statistic that reflects

the importance of a word in a document relative to a collection of documents, often used for
feature extraction.

• Embedding: A vector representation of words or phrases in a continuous space, typically

learned from a large corpus of text, used to capture semantic meaning.

11
 CHAPTER - 4

4. DESIGN SPECIFICATION

The design specification for this BiLSTM Model is centered around text classification using a deep
learning approach. It begins with meticulous data preprocessing, including the removal of HTML
tags, special character cleansing, and stemming, ensuring high-quality input data. Tokenization and
padding are performed for consistent input dimensions. Feature extraction relies on TF-IDF vector-
ization, capturing word importance.

The neural network model architecture is thoughtfully crafted, consisting of Dense layers with ReLU
activation functions, a Reshape layer, and a Bidirectional LSTM layer for sequence modeling. The
final Dense layer employs softmax activation for multi-class classification. Model training, validation,
and evaluation follow best practices, with early stopping to prevent overfitting and Matplotlib for
visualization, ensuring robust performance.

This code’s design emphasizes modularity and maintainability, leveraging a comprehensive technol-
ogy stack that includes Python, Pandas, NumPy, TensorFlow/Keras, NLTK, Scikit-learn, and Mat-
plotlib. The modular approach allows adaptability to various text classification tasks, making it a
valuable foundation for natural language processing projects.

We understand all these requirements better by developing the following diagrams of our system:

• Use Case Diagram

• Data Flow Diagram

• Class Diagram

• Sequence Diagram

• Activity Diagram

• State Chart Diagram

12
4.1. Use Case Diagram

Figure 4.1: Use Case Diagram for Text Classification using BiLSTM Model.

13
4.2. Data Flow Diagram

4.2.1. DFD LEVEL 0:

Figure 4.2: Data Flow Diagram Level 0 for Text Classification using BiLSTM Model.

14
4.2.2. DFD LEVEL 1:

Figure 4.3: Data Flow Diagram Level 1 for Text Classification using BiLSTM Model.

15
4.2.3. DFD LEVEL 2:

Figure 4.4: Data Flow Diagram Level 2 for Text Classification using BiLSTM Model.

16
4.3. Class Diagram

Figure 4.5: Class Diagram for Text Classification using BiLSTM Model.

17
4.4. Sequence Diagram

Figure 4.6: Sequence Diagram for Text Classification using BiLSTM Model.

18
4.5. Activity Diagram

Figure 4.7: Activity Diagram for Text Classification using BiLSTM Model.

19
4.6. State Chart Diagram

Figure 4.8: State Chart Diagram for Text Classification using BiLSTM Model.

20
 CHAPTER - 5

5. METHODOLOGY

5.1. Modules

Data Collection: Data collection involves gathering relevant information from various sources, such
as databases, sensors, surveys, or web scraping, to create a dataset that can be used for analysis or
machine learning. It is a crucial step in any data-driven project, as the quality and quantity of data
collected directly impact the effectiveness of subsequent tasks.

Data Pre-processing: Data pre-processing is the process of cleaning, transforming, and preparing
raw data for analysis or model training. This step includes tasks such as handling missing values,
removing outliers, standardizing or normalizing data, and encoding categorical variables to ensure
the data is in a suitable format for further analysis.

Feature Extraction: Feature extraction involves selecting or creating relevant features (variables)
from the dataset that will be used as input for machine learning models. This step aims to reduce
dimensionality, improve model performance, and capture the most important information in the data.

Data Splitting: Data splitting is the practice of dividing the dataset into two or more subsets: typically
a training set and a testing (or validation) set. This division allows for model training on one portion
of the data and evaluation on another, helping to assess the model’s performance and generalization.

Model Training: Model training involves using a machine learning algorithm or deep learning archi-
tecture to learn patterns, relationships, or rules from the training data. During this process, the model
adjusts its parameters to minimize a chosen error metric, making it capable of making predictions or
classifications on new, unseen data.

Model Evaluation: Model evaluation is the assessment of a trained machine learning model’s perfor-
mance on a separate testing dataset. Common evaluation metrics include accuracy, precision, recall,
F1-score, and mean squared error, among others. The goal is to measure how well the model gener-
alizes to new data and to identify any potential issues, such as overfitting or underfitting.

21
5.2. System Block Diagram

Figure 5.1: System Block Diagram for Text Classification using BiLSTM Model.

22
 CHAPTER - 6

6. IMPLEMENTATION DETAILS

In this Deep Learning model for Text Classification, we first load and preprocess text data from a CSV
file, cleaning it by removing HTML tags and special characters, and applying stemming. We then
tokenize the text, pad sequences to a consistent length, and extract TF-IDF features for classification.
The data is split into training and testing sets. A deep learning model is defined, featuring Dense
and Bidirectional LSTM layers for sequence modeling. The model is compiled with appropriate
loss and metrics for multi-class classification and trained on the training data with early stopping to
prevent overfitting. Model accuracy is evaluated, and training metrics are visualized using Matplotlib.
This code demonstrates a complete pipeline for text classification, from data preprocessing to model
training and evaluation, using Python, Pandas, NumPy, TensorFlow/Keras, NLTK, Scikit-learn, and
Matplotlib libraries.

6.1. Technology Stack

The technology stack employed in this BiLSTM Model encompasses various Python libraries and
frameworks for natural language processing (NLP) and deep learning. It includes Pandas for data
manipulation, NumPy for numerical operations, TensorFlow/Keras for deep learning model construc-
tion, NLTK for text preprocessing, Scikit-learn for data splitting and one-hot encoding, and Matplotlib
for visualization. Additionally, the code leverages the Pickle library for serialization and the Tf-
idfVectorizer from Scikit-learn for feature extraction. This comprehensive stack enables data loading,
cleaning, feature engineering, model development, training, evaluation, and visualization, making it
a robust solution for text classification tasks.

23
6.2. System Architecture

Figure 6.1: System Architecture Diagram for Text Classification using BiLSTM Model.

The system architecture of the model is designed for text classification using deep learning tech-
niques. It commences with data preprocessing, which includes cleaning the text data by removing
HTML tags, special characters, and stemming words. The tokenized and padded sequences are then
transformed into TF-IDF features. The neural network model, comprising Dense layers, Reshape
layer, and a Bidirectional LSTM layer, is constructed for sequence modeling and multi-class clas-
sification. The model is trained with early stopping and evaluated for accuracy. A comprehensive
technology stack, including Pandas, NumPy, TensorFlow, NLTK, Scikit-learn, and Matplotlib, is uti-
lized to streamline the entire process, ensuring efficiency and effectiveness in text classification tasks.

6.3. User Interface

The model primarily focuses on text classification and does not include a graphical user interface
(GUI) component. It is designed to operate as a backend script for automated text classification tasks.
Users interact with the code by providing input data in the form of a CSV file containing text and
categories for classification. The code processes the data, trains a deep learning model, and provides
results and visualizations, but this interaction occurs via command-line or script execution rather than
through a traditional graphical user interface.

6.4. Integration

The integration aspect of this model involves the seamless combination of various libraries, frame-
works, and technologies to create a cohesive and functional text classification system. It integrates
data preprocessing techniques from Pandas, NLTK, and regular expressions to clean and prepare the

24
text data. Feature extraction relies on Scikit-learn’s TF-IDF vectorization, while deep learning capa-
bilities are harnessed through TensorFlow/Keras for building and training the neural network model.
The code effectively integrates these components to enable end-to-end text classification, from data
loading to model evaluation, providing a robust and comprehensive solution for NLP tasks.

6.5. Security

Security considerations in this model primarily revolve around data privacy and model deployment.
To enhance security, sensitive data, if present, should be handled with care, and access control mech-
anisms must be implemented. Additionally, when deploying the model, securing APIs and endpoints
against potential attacks and ensuring data encryption during transmission are crucial. Regular model
updates and monitoring for vulnerabilities can help maintain a secure and robust system.

6.6. Testing and Deployment

Testing involves evaluating the trained model’s performance on the testing dataset to assess its gener-
alization. Once satisfied, the model can be deployed for real-world use cases. Deployment methods
vary; it can be deployed as a web service, integrated into an application, or hosted on cloud servers.
Regular monitoring and maintenance are essential to ensure the model’s continued accuracy and reli-
ability in a production environment.

Taking all the subsections into consideration, we were successfully able to develop a robust and
efficient Deep Learning model BiLSTM for Sentiment Analysis that can streamline segmentation
operations, reduce the manual work and enhance applications in various dental fields.

25
 CHAPTER - 7

7. OBSERVATIONS

7.1. Time Domain - Gann Chart

Figure 7.1: Gann Chart of Project Timeline.

7.2. Results and Comparitive Study

In our research the BiLSTM (Bidirectional Long Short-Term Memory) model effectively leveraged
its bidirectional mechanism to capture the sequential and contextual nature of input data. By con-
sidering both past and future dependencies, the model was able to focus on crucial segments of the
text, enabling more accurate sentiment categorization. The trend of increasing accuracy and valida-
tion accuracy across epochs reflects the model’s improving ability to correctly classify sentiments.
Simultaneously, the decreasing loss and validation loss throughout epochs indicate a reduction in
prediction inaccuracies, highlighting the model’s continuous refinement and enhanced performance.
These observations collectively underscore the model’s proficiency in text classification tasks.

26
Figure 7.2: Results showing accuracy and validation accuracy across epochs.

27
 CHAPTER - 8

8. CONCLUSION

The BiLSTM Model for Sentiment Analysis provides numerous benefits in Text Classification and
Emotion detection.

In conclusion, this project has successfully implemented a robust text classification pipeline using
deep learning techniques. We began by meticulously preprocessing the text data, including HTML
tag removal, special character cleaning, and stemming. The data was tokenized, and sequences were
padded to ensure consistent input dimensions. TF-IDF feature extraction was employed to transform
the text into numerical vectors suitable for machine learning.

Our deep learning model, composed of Dense and Bidirectional LSTM layers, demonstrated strong
performance in classifying text into multiple categories. The model was trained with early stopping
to prevent overfitting, and its accuracy on the training data was impressive.

Serialization of the tokenizer ensures consistent text preprocessing for future inference. Visualizations
of training and validation metrics aided in understanding model performance.

This project highlights the power of natural language processing and deep learning in automating text
classification tasks, making it valuable for a wide range of applications, including sentiment analysis,
content categorization, and more. Future work may involve fine-tuning hyperparameters, exploring
different neural architectures, and scaling up for larger datasets to further improve model accuracy
and robustness.

28
 CHAPTER - 9

9. LIMITATIONS AND FUTURE ENHANCEMENTS

9.1. Limitations

This project is highly planned and acted upon from the beginning. Nevertheless, the project had to
face some of the limitations due to various factors. Different aspects of the projects such as nature of
data, visualisation methods, data storage method and so on have their own limitations. Some of the
limitations faced by the project are:-

1. The code lacks extensive data exploration and analysis. Understanding the data distribution,
class imbalances, or outliers could lead to better preprocessing and model selection.

2. The model does not include advanced techniques like dropout or regularization to combat
overfitting. It might not generalize well to unseen data, especially if your dataset is small.

3. The hyperparameters of the model, such as the number of LSTM units or the learning rate, are
not optimized. Fine-tuning these parameters could significantly improve model performance.

4. If the dataset has imbalanced classes, the model may not perform well on minority classes.
Techniques like oversampling, undersampling, or using different loss functions need to be
considered.

5. The code evaluates the model on the training data, which might not reflect its real-world per-
formance. It’s essential to evaluate it on a separate test set to assess generalization.

6. While NLTK and Scikit-learn are useful, they may not cover all aspects of NLP. Using more
advanced NLP libraries like spaCy or transformers (e.g., Hugging Face’s Transformers) could
enhance text processing.

7. The code lacks techniques for model interpretability. Understanding why the model makes
specific predictions is essential, especially for applications in sensitive domains or when regu-
latory compliance is required.

29
 8. The tokenizer’s fixed vocabulary size and padding length might not be optimal for all datasets.
Dynamic tokenization approaches, like subword tokenization, could be more flexible.

9. There is no mechanism to track changes to the dataset. In real-world scenarios, datasets can
change over time, which can affect model performance and reproducibility.

10. The code lacks robust error handling and exception catching. Real-world applications should
include mechanisms to handle data anomalies or unexpected issues gracefully.

30
9.2. Future Enhancements

Future enhancements for the BiLSTM Model for Text Classification can be planned to further improve
its capabilities and address emerging needs. Some potential areas for enhancement include:

1. Optimize model hyperparameters such as learning rate, batch size, LSTM units, and the num-
ber of layers to improve model performance and convergence.

2. Implement data augmentation techniques like synonym replacement, word embeddings, or

back-translation to increase the diversity of the training dataset, which can help improve model
generalization.

3. Combine multiple models, such as LSTM, CNN, and Transformer-based models, into an en-
semble for improved classification accuracy and robustness.

4. Explore the use of pre-trained transformer models like BERT or GPT for text classification
tasks, which have shown state-of-the-art performance in NLP.

5. Experiment with pre-trained word embeddings (e.g., Word2Vec, FastText, GloVe) to capture
more nuanced semantic information in the text data.

6. Develop a user-friendly interface for users to interact with the model, input text data, and
receive predictions.

7. Extend the model to handle multiple languages or implement language detection to adapt to
different languages automatically.

8. Perform in-depth error analysis to understand the model’s weaknesses and iteratively improve
its performance.

31
A. APPENDIX

References

[1] Kanika Jindal, Rajni Aron,” A systematic study of sentiment analysis for social media data”,2214-
7853,January 2021.

[2] Shilpa P C , Rissa Shereen, Susmi Jacob, Vinod P,” SENTIMENT ANALYSIS USING DEEP
LEARNING”, . IEEE Xplore Part Number: CFP21ONG-ART; 978-0- 7381-1183-4, (ICICV May
2021).

[3] Suvarna G Kanakaraddi, Ashok K Chikaraddi, Karuna C. Gull, P S Hiremath,” Comparison Study
of Sentiment Analysis of Tweets using Various Machine Learning Algorithms”, IEEE Xplore Part
Number:CFP20F70-ART; ISBN:978-1-7281- 4685-0 ,(ICICT- June 2020)

[4] Nirmal Varghese Babu, E. Grace Mary Kanaga,” Sentiment Analysis in Social Media Data for
Depression Detection Using Artifcial Intelligence: A Review”, SN Computer Science (2022) 3:74,
October 2021.

[5] A.S. Alamri, S. A. Mahmoud, and M. A. AlZain, ”A Deep Learning Approach for Sentiment
Analysis on Twitter Da-ta,” IEEE Access, vol. 8, pp. 128,960-128975, 2020.

[6] Pansy Nandwani, Rupali Verma,” A review on sentiment analysis and emotion detection from
text”, Social Network Analysis and Mining (2021) 11:81,July 2021.

[7] Wang Dawei, Rayner Alfred, Joe Henry Obit, Chin Kim ,” A literature review on text classification
and sentiment analysis approaches”, 305-323, 2021August.

[8] Loukas Ilias, Ioanna Roussaki,” Detecting malicious activity in Twitter using deep learning tec-
niques”, 2021.107360 1568-4946/© , 2021

[9] Gild´asio Antonio de Oliveira J´unior , Rafael Tim´oteo de Sousa Jr. , Robson de Oliveira
Albuquerque , Luis Javier Garc´ıa Villalba ,” Adversarial attacks on a lexical sentiment analysis
classifier”,Computer Communications 174 (2021)154- 171, April 2021

[10] Lang,X.liu,Z.Chen,Y,Zhang,”A Sentiment Analysis Approach Based on Domain Knowledge and

Hireearchical Attention Networks “,IEEE Access,vol.8,pp.69,054-69,062, 2020

[11] Qianwen Ariel Xu , Victor Chang b , Chrisina Jayne,” A systematic review of social media-based
sentiment analysis: Emerging trends and challenges”, Decision Analytics Journal 3 (2022),2772-

32
6622,June 2022.

[12] Ahmed Alsayat,”Improving Sentiment Analysis for Social Media Applications Using an Ensem-
ble Deep Learning Language Model”,Arabian Journal for Science and Engineering (2022) 47:2499–2511,Oc-
tober 2021.

A.1. Project Timeline

33
Table A.1: Deep Learning Techniques for Sentiment Analysis on social Media Text project timeline

Date Topic Discussion Action Taken

03-02-2023 Discussion regarding the problem Clarified the problem statement
statement with the team
09-02-2023 Discussion about the title and re- Found an appropriate title for our
search journals project
12-02-2023 Explanation of our understanding Shared various journals and their
about the journals research findings
21-02-2023 Discussion about how to propose The methodology was chosen
new methodology
12-04-2023 Discussion of microdocument Drafted the micodocument
15-04-2023 Presentation of micro document Reviewed and presented microdoc-
ument with team
17-04-2023 Verification of micro document and Reworked the contents of mi-
ppt crodocument
17-05-2023 Discussion regarding the dataset Dataset is chosen and finalized
26-05-2023 Discussion about the architecture Determined the design and model
design and model methodology
05-06-2023 Discussion about the model flow di- Drawn the data flow diagrams
agrams
06-06-2023 Discussion on the overview of the
project
23-06-2023 Discussion regarding the code im- Viewed and various suitable codes
plementation for implementation
03-07-2023 Discussion about the evaluation
metrics and expected results
04-07-2023 Verification of microdocument and Addressed any potential plagiarism
plagiarism check concerns in the projects

A.2. Coding

GitHub Repository : https://github.com/ashrita17/DSentimentAnalysisOnSocialMediaText

34