A PROJECT REPORT

ON

Cyberbullying Detection in Social

Media Using Machine Learning: A
Hinglish Language Approach

Submitted by

KARAN RANA [Reg No:RA2111003030446]

SUHAIL SAIFI [Reg No: RA2111003030439]
ZUFAR ALVI [Reg No:RA2111003030435]
SOHAIL [Reg No: RA2111003030436]

Under the guidance of

Mr. MAYANK GUPTA
(ASSISTANT PROFESSOR, DEPARTMENT OF COMPUTER

SCIENCE & ENGINEERING) in partial fulfillment for

the award of the degree of

BACHELOR OF TECHNOLOGY
in

COMPUTER SCIENCE & ENGINEERING

of

FACULTY OF ENGINEERING AND TECHNOLOGY

NOV 2024
SRM INSTITUTE OF SCIENCE & TECHNOLOGY
(Under Section 3 of UGC Act, 1956)

BONAFIDE CERTIFICATE

Certified that this project report titled “ Cyberbullying Detection

in Social Media Using Machine Learning: A Hinglish
Language Approach” is the bonafide work of “ KARAN RANA
[Reg No: RA2111003030446], SUHAIL SAIFI [Reg No:
RA2111003030439], ZUFAR ALVI [Reg No: RA2111003030435] ,
SOHAIL [Reg No: RA2111003030436], ”, who carried out the project
work under my supervision. Certified further, that to the best of my
knowledge the work reported herein does not form any other project re-
port or dissertation on the basis of which a degree or award was conferred
on an earlier occasion on this or any other candidate.

SIGNATURE
SIGNATURE

Mr. MAYANK GUPTA

GUIDE Dr. Avneesh Vashistha
Assistant Professor
HEAD OF THE DEPARTMENT
Dept. of Computer Science & Engi- Dept. of Computer Science & Engi-
neering neering

Signature of the Internal Examiner Signature of the External Examiner

 ABSTRACT

Cyberbullying has emerged as a significant social issue with the rapid growth of

social media platforms, impacting mental health and well-being. Despite various

efforts to counter cyberbullying, detecting such harmful content in diverse and

multilingual environments remains challenging. This project presents a

machine-learning-based approach to cyberbullying detection, particularly focused on

the Hinglish language, a popular blend of Hindi and English widely used in social

media conversations in India. Traditional cyberbullying detection systems have

primarily focused on single languages, often overlooking the unique characteristics of

Hinglish, including code-switching, informal grammar, and slang.

The proposed approach is built on a dataset of Hinglish tweets, which includes

samples labeled as cyberbullying or non-cyberbullying. We applied a combination of

Natural Language Processing (NLP) and Machine Learning (ML) techniques to

analyze and classify these messages effectively. Key NLP preprocessing steps,

including tokenization, stop-word removal, and lemmatization, were employed to

refine the dataset. To enhance feature extraction, we used the Term Frequency-Inverse

Document Frequency (TF-IDF) model, capturing relevant features that represent the

nature of Hinglish cyberbullying messages.

In the machine learning phase, multiple algorithms were implemented and evaluated,

including Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Logistic

Regression, and Stochastic Gradient Descent (SGD) Classifier. These classifiers were

chosen for their varying strengths in handling classification tasks with

high-dimensional data. Performance metrics such as accuracy, precision, recall, and

F1 score were computed to compare the algorithms' effectiveness in detecting

cyberbullying. Among the classifiers tested, Support Vector Machine demonstrated

high accuracy, reflecting the model’s capability in identifying nuanced cyberbullying

language in Hinglish.

This project also introduces an interactive web-based application developed with

Streamlit, allowing users to input text and receive real-time feedback on potential

cyberbullying content. The application provides a user-friendly interface, making it

accessible for general use and further research.

The findings from this research underline the feasibility of using machine learning for

Hinglish cyberbullying detection and highlight the importance of language-specific

models in creating safe online environments. Future research can build upon this

study to improve detection accuracy by incorporating advanced NLP techniques and

expanding the application to detect other forms of abusive content across additional

multilingual settings.
 ACKNOWLEDGEMENTS

I would like to express my deepest gratitude to my guide, Mr. MAYANK

GUPTA(ASSISTANT PROFESSOR, DEPARTMENT OF COMPUTER SCIENCE &

ENGINEERING) for his valuable guidance, consistent encouragement,

personal caring, timely help and providing me with an excellent

atmosphere for doing research. All through the work, in spite of his busy

schedule, he has extended cheerful and cordial support to me for

completing this research work.

karan rana

Author
 TABLE OF CONTENTS

ABSTRACT iii

ACKNOWLEDGEMENTS v

LIST OF TABLES AND FIGURES vii

ABBREVIATIONS ix

LIST OF SYMBOLS x

1 INTRODUCTION 1
1.1 background and problem statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 objective of the study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 scope and significance of project........................................................................2

2 LITERATURE SURVEY 3
2.1 challenges and existing approaches. . . . . . . . . . . . . . . . . . . . . . . . . 3
2.2 machine learning and NLP in cyberbullying detection....................................3

3 System Analysis....................................................................................................... 4

4 System Design. . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . 5
4.1 Tables and Figures............................................................................................ 5

5 Code and outputs...................................................................................................... 8

6 Conclusion................................................................................................................ 10

7 Future Enhancements............................................................................................. 12

8 References................................................................................................................ 14
 LIST OF TABLES AND FIGURES

1. Comparison of algorithm with count vectorizer...........................................5

2. Comparison of algorithm with TF-IDF........................................................5

3. Table depicting accuracy of various algorithms............................................5

4. pie chart depicting the distribution of bullying and non bullying dataset......5

5.Research methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . 7

6. functional block diagram.......................................................................................... 7

7. server-side working.................................................................................................. 7

8. comparison between CV and TF-IDF......................................................................7

 ABBREVIATIONS

AI - Artificial Intelligence

API - Application Programming Interface

BI - Business Intelligence

CSV - Comma-Separated Values

FN - False Negative

FP - False Positive

KNN - K-Nearest Neighbors

LR - Logistic Regression

ML - Machine Learning

NLP - Natural Language Processing

NB - Naïve Bayes

SVM - Support Vector Machine

SGD - Stochastic Gradient Descent

TF-IDF - Term Frequency-Inverse Document Frequency

TN - True Negative

TP - True Positive

UI - User Interface

UX - User Experience

MLP - Multilayer Perceptron

ROC - Receiver Operating Characteristic

 LIST OF SYMBOLS
α - Learning rate (used in gradient descent optimization)

Σ - Summation symbol, often used to denote the sum of multiple terms

|V| - Vocabulary size (total number of unique words or tokens)

μ - Mean or average value of a dataset

σ - Standard deviation, representing data dispersion around the mean

ω - Weight vector in machine learning models (used in algorithms like SVM)

f(x) - Prediction function or model function output

→ - Vector notation, e.g., x⃗\vec{x}x for feature vectors

x
 CHAPTER 1

INTRODUCTION

With the explosive growth of social media, communication barriers are shrinking,

enabling people to connect across the world in real time. However, alongside the

positive effects, this digital revolution has fueled the rise of cyberbullying—a form of

online harassment that can severely impact mental well-being. In India and similar

multilingual regions, online communication often occurs in a mix of Hindi and

English, popularly known as Hinglish. This blend of languages, informal grammar,

and frequent code-switching poses unique challenges to traditional cyberbullying

detection systems that typically focus on single-language texts. This project tackles

these challenges by building a machine learning model to detect cyberbullying

specifically in Hinglish, utilizing Natural Language Processing (NLP) for effective

feature extraction and classification.

This project report provides a comprehensive overview of the methodologies used,

results obtained, and the implications of this approach for creating safer digital

spaces. It demonstrates how a custom-built application allows users to test

cyberbullying detection in Hinglish messages, highlighting the need for language-

specific solutions.
1. Background and Problem Statement

In recent years, cyberbullying has emerged as a serious social issue, with cases

rising due to increased digital connectivity. This form of bullying differs from

traditional bullying as it can happen anonymously and across distances,

amplifying its psychological impact on victims. Detecting cyberbullying in

Hinglish, however, presents new challenges. The language is unique,

containing a mix of English and Hindi vocabulary, informal grammar, and

colloquial expressions. Traditional detection models often fail to recognize

these subtleties, making it essential to design a solution tailored to Hinglish.

Objective of this Subtopic: Explain why cyberbullying is harmful, especially in

multilingual regions, and how Hinglish creates specific detection challenges.

2. Objectives of the Study

The primary objectives of this study are threefold:

○ Develop an Accurate Detection Model: Use machine learning

algorithms that can handle high-dimensional data to identify

cyberbullying in Hinglish.

○ Compare Algorithm Performance: Evaluate multiple machine learning

algorithms like Support Vector Machine (SVM), Logistic Regression,

and K-Nearest Neighbors (KNN) to find the most effective model for

Hinglish.

○ Implement a User-Friendly Application: Build a web application

where users can enter text and receive real-time feedback on potentially

harmful content, enhancing accessibility and usability.

3. Objective of this Subtopic: Clearly state what the study aims to achieve and the
 practical goals of the project in both technical and user-centered terms.

4. Scope and Significance of Cyberbullying Detection

Detecting cyberbullying effectively is crucial to protecting individual mental

health and promoting positive online communities. By focusing on Hinglish,

this project fills a critical gap in current research, which primarily focuses on

monolingual or English-only detection systems. Through this Hinglish-specific

model, the project provides valuable insights into handling other language

blends or multilingual environments for cyberbullying detection. The broader

significance of this research lies in its potential application to other social

media platforms, enabling safer online spaces in linguistically diverse regions

 CHAPTER 2

LITERATURE SURVEY

The detection of cyberbullying, particularly in mixed-language environments like Hinglish, requires

a thorough understanding of prior work in various domains such as natural language processing
(NLP), machine learning (ML), and multilingual cyberbullying detection. This literature survey
aims to review the existing body of work in these areas, focusing on how machine learning
techniques and language-specific models have been applied to cyberbullying detection, particularly
in mixed-language scenarios.

1. Cyberbullying Detection: Challenges and Existing Approaches

Cyberbullying detection has gained significant attention in recent years due to its detrimental impact
on mental health and online safety. Early work in this domain focused on rule-based approaches and
keyword matching to identify bullying content in social media. However, these methods proved
inadequate in dealing with complex linguistic structures, sarcasm, and indirect bullying language. In
recent years, machine learning techniques have emerged as a more effective solution for automatic
cyberbullying detection.

Studies such as those by Dadvar et al. (2013) and Kwok & Wang (2013) leveraged supervised
learning algorithms (e.g., SVM, Decision Trees) to classify texts as abusive or non-abusive. Their
findings demonstrated that while traditional models could capture simple bullying patterns, they
struggled with more nuanced forms of bullying, like covert harassment or bullying through images
or indirect communication.

Recent advancements, however, incorporate deep learning models such as Recurrent Neural
Networks (RNNs) and Convolutional Neural Networks (CNNs) for improved feature extraction
from textual data. Gambäck & Sikdar (2017) proposed using character-level embeddings to
address the challenges of informal text such as slang, emojis, and code-switching. These
advancements have significantly improved detection accuracy, yet challenges remain for languages
like Hinglish, where traditional models are less effective.
Objective of this Subtopic: To highlight the evolution of cyberbullying detection techniques,
focusing on the shift from rule-based to machine learning and deep learning models, and to discuss
the limitations of existing approaches in multilingual contexts.

2. Machine Learning and NLP in Cyberbullying Detection

Machine learning, particularly NLP techniques, has proven essential in enhancing the accuracy of
cyberbullying detection systems. NLP methods such as tokenization, sentiment analysis, and part-
of-speech tagging are commonly employed to extract meaningful features from social media content.
Machine learning classifiers like Naive Bayes (NB), Logistic Regression (LR), and Support
Vector Machines (SVM) are often used to categorize texts based on whether they contain bullying
or non-bullying content.

A study by Founta et al. (2018) explored the use of TF-IDF (Term Frequency-Inverse Document
Frequency) for feature extraction and applied classifiers like SVM and Random Forests to detect
cyberbullying in social media posts. The results showed that SVM performed particularly well in
binary classification tasks, demonstrating its ability to handle high-dimensional feature spaces
effectively. However, these models still face challenges when dealing with informal, non-standard
text forms like Hinglish, which requires specialized preprocessing techniques for better
classification.

The need for custom language models has been emphasized in research focusing on languages with
significant code-switching, like Hinglish. Vyas et al. (2019) proposed a hybrid model combining
rule-based methods with machine learning to improve detection accuracy for Hinglish. Their
approach involved a custom dictionary of common Hinglish terms and slang expressions, which
helped boost the model’s performance in detecting bullying content in mixed-language texts.

Objective of this Subtopic: To examine the role of machine learning and NLP in detecting
cyberbullying, with a focus on the algorithms and feature extraction techniques that have been
employed in recent studies, as well as the limitations faced in multilingual environments.
 CHAPTER 3

SYSTEM

ANALYSIS

1. Functional Requirements

The system must meet the following functional requirements:

● Text Input and Preprocessing: The system must accept social media posts,
comments, or tweets in Hinglish as input. The text should undergo
preprocessing, including tokenization, language detection, and conversion of
Hinglish text into a usable format for the model.
● Cyberbullying Classification: The core functionality of the system is to
classify input text into "bullying" or "non-bullying" categories. This
classification will be done using machine learning models, with accuracy
optimized for Hinglish.
● Model Training and Evaluation: The system must train a machine learning
model on a labeled dataset of Hinglish text to identify bullying content. It
should also evaluate model performance using metrics such as accuracy,
precision, recall, and F1-score.
● User Interface: A simple, user-friendly interface (such as a web application)
will allow users to input Hinglish text and receive feedback on whether the text
contains cyberbullying content.
● Real-time Feedback: The system must provide real-time feedback, displaying
results almost instantly after a user submits text. The output should inform the
user whether the message is classified as cyberbullying, along with a
confidence score indicating the certainty of the classification.
● Multilingual Handling: The system should be able to detect Hinglish, with the
capability to handle mixed languages, informal vocabulary, and slang terms
common in Indian social media.

2. Non-Functional Requirements

In addition to functional requirements, the system must also fulfill the following
non-functional requirements:

● Scalability: The system should be scalable to handle a large volume of

requests, as social media platforms generate vast amounts of content daily. This
requires the ability to process multiple users’ inputs simultaneously without
performance degradation.
● Performance: The system must be optimized for fast response times, ensuring
that text input results in feedback within a few seconds. Performance can be
 affected by the size of the dataset and the complexity of the machine learning
models.
● Security and Privacy: Given the sensitivity of the data involved in detecting
cyberbullying, the system must ensure data privacy. User data should be
anonymized to avoid any breaches of privacy.
● Accuracy: One of the key goals of the system is to achieve high accuracy in
identifying cyberbullying, even in informal and code-switched text. The system
must ensure minimal false positives and false negatives to avoid misclassifying
non-harmful content as bullying and vice versa.
● Usability: The system should be easy to use for non-technical users, with clear
instructions and feedback. The user interface should be simple and intuitive,
especially for those who may not be familiar with complex machine learning
models.

3. System Architecture

The system architecture is designed to ensure seamless processing from text input to
classification output. The architecture includes several layers that handle different
tasks, as outlined below:

● Data Collection Layer: This layer is responsible for collecting Hinglish data
from social media platforms (e.g., Twitter, Facebook, Instagram). Data can be
scraped using APIs or collected from publicly available datasets of social
media posts. The dataset will contain both bullying and non-bullying messages
for training purposes.
● Data Preprocessing Layer: This layer is essential for cleaning and preparing
the Hinglish text data for machine learning. The preprocessing steps include:
○ Tokenization: Splitting the text into individual words or tokens.
○ Language Detection: Identifying the Hindi-English language mix and
normalizing the text.
○ Noise Removal: Removing unnecessary symbols, links, or stop words.
○ Slang Detection and Translation: Identifying Hinglish-specific slang
and regional variations and converting them into a more structured form.
● Machine Learning Layer: After preprocessing, the data is passed to the
machine learning models. Several algorithms such as Support Vector
Machine (SVM), Naive Bayes (NB), and Logistic Regression (LR) are
trained using the labeled data. These models are evaluated based on
performance metrics to determine the most effective one for cyberbullying
detection.
● Prediction and Feedback Layer: This layer handles user inputs and runs
predictions using the trained machine learning models. Once the model
classifies the text, feedback is generated, and the user is notified in real-time
whether the text contains cyberbullying content or not.
● User Interface Layer: The final layer is the user interface, which allows users
to interact with the system through a simple web application. The interface will
accept text inputs, process them, and display results to the user.
4. Tools and Technologies

The system utilizes several key technologies to achieve its objectives:

● Programming Languages: Python will be used for implementing the machine

learning algorithms and NLP tasks, given its rich ecosystem of libraries for data
analysis and natural language processing (e.g., scikit-learn, NLTK, Pandas).
● Web Framework: Tkinter will be used for building the user interface, making
it easy for users to submit text and receive feedback.
● Machine Learning Libraries: Libraries such as scikit-learn for model
training, TensorFlow or Keras for deep learning (if required), and XGBoost
for boosting algorithms will be used for building and evaluating the models.
● Data Sources: The system will make use of publicly available datasets for
training the models, including Twitter data and datasets from previous
cyberbullying detection research.
 CHAPTER 4

SYSTEM DESIGN

1. System Architecture
The system follows a layered architecture to separate different concerns and ensure
modularity. It consists of the following components:

● Data Collection Layer: This module is responsible for gathering social media
posts or comments that are written in Hinglish. The data is scraped from
publicly available datasets or fetched from social media APIs (such as Twitter
or Facebook) using keywords related to cyberbullying. The data consists of a
mix of text types, including informal language, slang, and code-switched
language. The collected data is stored in a structured format, such as a CSV file
or a database, for later preprocessing and model training.
● Preprocessing Layer: The raw text data collected from social media is
preprocessed to make it suitable for machine learning. This module performs
several steps:
○ Tokenization: Breaking down the text into individual words or phrases.
○ Noise Removal: Removing unnecessary characters, links, stop words, or
irrelevant data.
○ Text Normalization: Converting informal text or Hinglish into a
standard format. This includes converting slang terms into their
corresponding English meanings or standardized forms.
○ Feature Extraction: Using techniques like TF-IDF (Term Frequency-
Inverse Document Frequency) or Word2Vec for converting text into
numerical features that can be used by machine learning algorithms.
● Machine Learning Layer: This core component contains the machine learning
algorithms that are trained on the labeled data to classify text as
"cyberbullying" or "non-cyberbullying." The training process involves:
○ Data Splitting: Dividing the dataset into training, validation, and testing
sets.
○ Model Training: Using various classification algorithms like Support
Vector Machines (SVM), Naive Bayes (NB), and Logistic Regression
(LR). The models are trained using the processed and feature-extracted
data.
○ Model Evaluation: Assessing the model’s performance using accuracy,
precision, recall, and F1-score. The best-performing model is selected
 based on these metrics.
● Prediction and Feedback Layer: After training the models, this module takes
input text from the user and classifies it using the selected model. Once a user
submits a text (e.g., a comment or tweet), the system processes the text through
the preprocessing module and then feeds it into the trained model for
prediction. The model outputs a label indicating whether the text is classified as
"cyberbullying" or "non-cyberbullying." Additionally, a confidence score is
generated to indicate the certainty of the prediction.
● User Interface Layer: The system provides an interface where users can
interact with the cyberbullying detection system. This user-friendly interface is
built using web technologies such as HTML5, CSS3, and JavaScript, and is
implemented through a framework like Flask or Django. The user enters the
text into an input field, and the result is displayed within seconds. The interface
is designed to be simple and intuitive, with options to enter different forms of
social media text and get real-time results.

2. Data Flow and Interactions

The flow of data in the system follows these steps:

1. User Input: The user enters a social media post or comment written in
Hinglish into the system via the web interface.
2. Preprocessing: The text is passed through the preprocessing module, where it
is tokenized, normalized, and transformed into features that the machine
learning model can understand.
3. Prediction: The preprocessed text is sent to the trained machine learning
model, which classifies the text as "cyberbullying" or "non-cyberbullying."
4. Output: The prediction result is displayed to the user, along with a confidence
score.
5. Feedback: Users are given feedback to help understand the results. If the text
is classified as cyberbullying, a message can be provided with suggestions or
warnings.

4.1 Tables and Figures

 Comparison of Algorithms with Count Vectorizer

Comparison of Algorithms with Term Frequency-Inverse Document

Frequency
 accuracies of various machine learning algorithms

pie chart depicting the distribution of bullying and non bullying dataset

FIGURES
 RESEARCH METHODOLOGY

FUNCTIONAL BLOCK DIAGRAM

representation of how various algorithms can be worked over the dataseT.

 CHAPTER 5

CODE AND OUTPUT

5.1 CODE SNIPPET OF THE PROJECT

client GUI python code

Server side program

5.2 OUTPUT OF THE PROJECT

server side

working UI interface for two different users

 CHAPTER 6

CONCLUSION

In conclusion, this project aimed to develop a robust system for detecting

cyberbullying in Hinglish text on social media platforms using machine learning
techniques. The system successfully integrates multiple components, including data
collection, preprocessing, feature extraction, machine learning model training, and
real-time predictions, to classify text as either cyberbullying or non-cyberbullying.
Through the application of various classification algorithms such as Support Vector
Machine (SVM), Naive Bayes, and Logistic Regression, we achieved a functional
model capable of handling the complexities of informal language, slang, and code-
switching in Hinglish.
The project demonstrates the potential of machine learning in addressing the growing
concern of cyberbullying on social media, specifically targeting Hinglish, a blend of
Hindi and English, which poses unique challenges in natural language processing. By
utilizing techniques such as TF-IDF for feature extraction, the system can effectively
process and classify social media posts with a reasonable degree of accuracy.
The system's user-friendly interface allows users to easily input text and receive
instant feedback on whether the text qualifies as cyberbullying or not, providing
valuable insights into the nature of online interactions. This can serve as a preventive
tool for both individuals and organizations looking to identify harmful content on
social media platforms.
Overall, the project successfully meets its objectives and demonstrates the application
of machine learning in a socially relevant domain. While the system performs well for
the given dataset, future improvements could include enhancing the model's accuracy
by incorporating larger, more diverse datasets and exploring advanced deep learning
techniques. Additionally, real-time deployment on a larger scale could help in
preventing cyberbullying incidents by raising awareness and offering early detection
mechanisms.
This project has the potential to contribute to ongoing efforts to create safer online
spaces by providing a scalable solution to detect and mitigate cyberbullying,
especially in multi-lingual, culturally diverse environments like India, where Hinglish
is widely used.
 CHAPTER 7

FUTURE ENHANCEMENT

1. Model Improvement and Accuracy Enhancement

One of the key areas of improvement for the system is the accuracy of the machine learning
model. The current implementation uses classical machine learning algorithms like Support
Vector Machines (SVM), Naive Bayes (NB), and Logistic Regression (LR). These
models, while effective for small datasets, may not perform as well on larger and more
diverse social media text datasets. To enhance accuracy, the following steps can be taken:

● Deep Learning Models: Transitioning from traditional machine learning models to

deep learning approaches like Recurrent Neural Networks (RNNs) or Long Short-
Term Memory (LSTM) networks can improve the system’s understanding of the
complex structure of natural language, especially for informal and code-switched text.
These models are particularly useful in text classification tasks as they are capable of
learning dependencies across sequences of words, which is critical for detecting
nuances in Hinglish.
● BERT and Transformer Models: Bidirectional Encoder Representations from
Transformers (BERT) has shown state-of-the-art performance in many NLP tasks.
Fine-tuning a pre-trained BERT model specifically for cyberbullying detection would
likely result in a significant improvement in classification performance. BERT is
effective in capturing the context of words in a sentence and can better understand the
nuances of Hinglish text, which often mixes multiple languages and slang.
● Ensemble Models: Combining the outputs of multiple models through ensemble
methods like Random Forests, AdaBoost, or XGBoost could improve the model’s
robustness. By aggregating the predictions of various models, the final prediction
tends to be more reliable, particularly in cases of ambiguity in the input data.
● Hyperparameter Tuning: Optimizing the hyperparameters of the models used can
lead to better performance. Techniques like Grid Search or Random Search can be
applied to find the best combination of parameters, such as kernel type in SVM or the
number of estimators in ensemble methods, to maximize accuracy.

2. Multi-Lingual and Code-Switching Support

Since the project focuses on Hinglish, a language mix of Hindi and English, one of the
challenges is handling code-switching—the practice of alternating between languages within
a sentence or phrase. The current system relies on standard preprocessing techniques that
may not fully capture the context or sentiment of Hinglish text. To address this, the
following enhancements could be made:

● Preprocessing with Hinglish-Specific Tokenization: A language-specific tokenizer

trained to identify Hinglish-specific words and patterns could improve preprocessing.
 This tokenizer would handle slang and transliterations, recognizing the semantic
meaning behind informal words used in Hinglish conversations.
● Custom Dictionaries and Language Models: Building a custom dictionary or using
pre-existing Hinglish datasets can help the system recognize commonly used words
and phrases in Hinglish, which are not typically found in standard language models.
This dictionary could be continuously updated as new slang or expressions emerge.
● Cross-lingual Embeddings: Using multilingual word embeddings, such as
mBERT or XLM-R, which are trained to handle multiple languages simultaneously,
can help in understanding Hinglish text better. These embeddings capture the
relationships between words in multiple languages, improving the model's ability to
understand the intent and meaning behind code-switched sentences.

3. Real-Time Prediction and Scalability

Enhancing the ability of the system to provide real-time predictions for large-scale data is
crucial for widespread adoption. A robust real-time detection mechanism would help in
actively monitoring social media platforms for harmful content. Some strategies to enhance
this functionality include:

● Web Scraping and API Integrations: Integrating the system with social media
platforms like Twitter, Facebook, or Instagram using their APIs would allow the
system to scan posts and comments in real-time. This can be achieved by setting up
scheduled scrapers or streaming APIs to collect social media data as it is posted.
● Batch Processing for Large Datasets: To ensure scalability, implementing batch
processing for large datasets can improve performance. In situations where real-time
analysis is not necessary, processing large amounts of historical data can be done in
batches, using distributed computing frameworks like Apache Spark or Hadoop.
● Cloud Integration: Deploying the system on a cloud platform like AWS, Google
Cloud, or Microsoft Azure would allow the system to scale as needed. The cloud
infrastructure can accommodate spikes in traffic and large data volumes, ensuring
seamless real-time performance. Cloud services can also facilitate the storage and
access of huge datasets, which is essential for handling social media content.

4. User Interface and Experience

The current user interface is basic and provides text input and result output. However, to
enhance the user experience, several features can be added:

● Sentiment Analysis: In addition to classifying text as cyberbullying or non-

cyberbullying, adding a sentiment analysis feature would provide users with insights
into the emotional tone of the post. For instance, users would be able to see if a post is
angry, sad, or neutral, which could further help in understanding the context of
cyberbullying.
● Visualization of Results: Adding visual elements like word clouds, graphs, or heat
maps to represent the severity or frequency of cyberbullying instances could make
the feedback more engaging and informative. This would also help in identifying
patterns and trends in cyberbullying across different social media platforms.
● Multi-Platform Support: Expanding the system to work not just on a web interface
but also as a mobile app or as a browser extension would make it more accessible to
users. By allowing users to analyze posts directly from social media platforms or
 blogs, the system would become more practical and widely used.

5. Continuous Learning and Feedback Loop

To ensure that the system remains effective over time, it is essential to implement a feedback
loop and enable continuous learning. New data and user feedback can be incorporated into
the model, helping it stay up to date with evolving language patterns and new forms of
cyberbullying:

● Active Learning: Implementing an active learning framework where the model is

periodically retrained on newly labeled data can improve its performance. User
feedback on predictions can help the system learn from its mistakes and improve over
time.
● Crowdsourced Data Labeling: Involving users in labeling new data or correcting
misclassified data can accelerate model improvement. Crowdsourced labeling can be
done through a simple interface where users flag cyberbullying content, which can
then be reviewed and added to the training data.
 CHAPTER 8

REFERENCES

[1] B. Dean, “How many people use social media in 2021? (65+ statistics),”
Sep. 2021. [Online]. Available: https://backlinko.com/social-media-users

[2] J. W. Patchin, “Summary of our cyberbullying research (2004-

2016),” Jul. 2019.
[Online].
Available:https://cyberbullying.org/summary-of-our-cyberbullying-research

[3] S. M. Novianto, I. Isa, and L. Ashianti, “Cyberbullying

classification using text mining,” in Proc. 1st Int. Conf. on Informatics
and Computational Sciences (ICICoS), 2017, pp. 241–246.

[4] C. Van Hee et al., “Automatic detection of cyberbullying in social

media text,” PLoS One, vol. 13, no. 10, p. e0203794, 2018.

[5] M. A. Al-Garadi et al., “Predicting cyberbullying on social media

in the big data era using machine learning algorithms: Review of
literature and open challenges,” IEEE Access, vol. 7, pp. 70 701–70
718, 2019.

[6] K. Sahay, H. S. Khaira, P. Kukreja, and N. Shukla, “Detecting

cyberbullying and aggression in social commentary using NLP and
machine learning,” Int. J. Engineering Technology Science and
Research, vol. 5, no. 1, 2018.

[7] M. Di Capua, E. Di Nardo, and A. Petrosino, “Unsupervised

cyberbullying detection in social networks,” in Proc. 23rd Int. Conf. on
Pattern Recognition (ICPR), 2016, pp. 432–437.

[8] H. Hosseinmardi et al., “Detection of cyberbullying incidents on

the Instagram social network,” arXiv preprint arXiv:1503.03909, 2015.

[9] V. Banerjee, J. Telavane, P. Gaikwad, and P. Vartak, “Detection of

cyberbullying using deep neural network,” in Proc. 5th Int. Conf. on
Advanced Computing & Communication Systems (ICACCS), 2019, pp.
604–607.

[10]H. Watanabe, M. Bouazizi, and T. Ohtsuki, “Hate speech on

Twitter: A pragmatic approach to collect hateful and offensive
expressions and perform hate speech detection,” IEEE Access, vol. 6,
pp. 13 825–13 835, 2018.
 [11]J. Yadav, D. Kumar, and D. Chauhan, “Cyberbullying detection using
pre-trained BERT model,” in Proc. Int. Conf. on Electronics and Sustainable
Communication Systems (ICESC), 2020, pp. 1096–1100.

[12] A. Gaydhani, V. Doma, S. Kendre, and L. Bhagwat, “Detecting hate

speech and offensive language on Twitter using machine learning: An n-
gram and TF-IDF based approach,” arXiv preprint arXiv:1809.08651, 2018.

[13] L. Ketsbaia, B. Issac, and X. Chen, “Detection of hate tweets using

machine learning and deep learning,” in Proc. 19th Int. Conf. on Trust,
Security and Privacy in Computing and Communications (TrustCom), 2020,
pp. 751–758.