VISVESVARAYA TECHNOLOGICAL UNIVERSITY

Belagavi-590018

A
Progress Report Phase 1
on

“ Cyber Bullying and Toxic Comment Detection for Indian

Languages”

Submitted in partial fulfillment as per VTU curriculum for VI

semester for the award of degree of

Bachelor of Engineering
In
Artificial Intelligence & Data Science

Submitted By Guide
GOVARDHAN REDDY - 1EP23AD011 Prof. SUDHAKARA K
KRISHNA SAI M - 1EP23AD020 ASSISTANT PROFESSOR
SHASHANK - 1EP23AD046 Dept. Of AI&DS,EPCET
JEEVAN VP - 1EP23AD015

Department of Artificial Intelligence & Data Science

Jnana Prabha Campus, Virgo Nagar Post,
Bidarahalli.
Bengaluru – 560049

2025-26
 Department of Artificial Intelligence and Data Science

Certificate

This is to certify that GOVARDHAN REDDY(1EP23AD011), KRISHNA SAI M(1EP23AD020), SHASHANK

(1EP23AD046) & JEEVAN VP (1EP23AD015) students of East Point College of Engineering and
Technology have completed the Major Project Phase 1 in partial fulfilment for the award of Bachelor of
Engineering in Artificial Intelligence & Data Science of the Visvesvaraya Technological University, Belgaum
during the year 2025-2026. It is certified that all corrections/suggestions indicated for Internal Assessment have been
incorporated in the Report. This Major Project Phase 1 Report has been approved as it satisfies the academic requirements
in respect of Major Project Phase 1 work carried out for 6th semester of Bachelor of Engineering degree.

----------------------------- ------------------------
Sign. of Guide
Sign. of
[Link] K
HOD Dr.
Anand R

Assistant professor Professor & HOD

Department of AI & DS Department of AI & DS
EPCET, Bengaluru-49 EPCET, Bengaluru-49
 Cyber Bullying and Toxic Comment Detection for Indian languages

TABLE OF CONTENTS

[Link] Content Page no.

Ⅰ Abstract 1

1 Problem Statement 2

2 Literature Review 3

3 Gap Identification 5

4 Methodology 6

5 Data Acquisition 8

6 Purpose 10

6.1 Objective of the Study 10

6.2 Measurement Parameters 10

6.3 Measurement Method 11

6.4 Tools and Equipment 11

6.5 Error Sources and Mitigation 11

6.6 Data Analysis 12

7 Experimental Framework 13

8 Result, Analysis and Discussion 14

9 Significance and Implications 15

10 Potential Applications 16

Dept. of AI&DS, EPCET 2025-26

 Cyber Bullying and Toxic Comment Detection for Indian languages

ABSTRACT
The rapid growth of social media platforms has significantly increased online
communication, enabling users to share opinions and interact globally. However, this
expansion has also led to a rise in cyberbullying, hate speech, and toxic comments,
which negatively impact individuals’ mental well-being and create unsafe digital
environments. Manual detection and moderation of such harmful content is challenging
due to the massive volume and unstructured nature of user- generated data.
Most existing toxic comment detection systems are primarily designed for the English
language and fail to effectively handle Indian regional languages and code-mixed text
such as Hinglish. Additionally, current approaches often struggle with informal language,
slang, misspellings, and contextual understanding, reducing their effectiveness in real-
world scenarios.
This project proposes a machine learning–based system for cyberbullying and toxic
comment detection in Indian languages. The system utilizes Natural Language
Processing (NLP) techniques for preprocessing text, including tokenization,
normalization, and stopword removal. Feature extraction is performed using methods such
as TF-IDF vectorization, converting textual data into numerical representations suitable
for machine learning models.
Multiple classification algorithms, including Logistic Regression, Naïve Bayes, Support
Vector Machine (SVM), and Random Forest, are trained and evaluated to identify the
most effective model for toxicity detection. The system is designed to support multilingual
and code-mixed inputs, improving its applicability in Indian social media contexts.
An experimental framework is established to evaluate model performance using metrics
such as accuracy, precision, recall, and F1-score, along with confusion matrix analysis.
The proposed system aims to achieve high classification accuracy while maintaining low
prediction time, enabling real-time detection of toxic content.
Furthermore, a simple user interface is developed to allow users to input comments and
receive instant predictions. The system can be integrated into social media platforms,
online forums, and communication systems to assist moderators in filtering harmful
content.
Overall, this project contributes to the development of an efficient, scalable, and practical
solution for automated toxic comment detection, promoting safer and healthier online
communication environments.

Dept. of AI&DS, EPCET 2025-26

 Cyber Bullying and Toxic Comment Detection for Indian languages

II

Dept. of AI&DS, EPCET 2025-26

 Cyber Bullying and Toxic Comment Detection for Indian languages

Chapter:1
PROBLEM STATEMENT
The exponential growth of social media platforms and online communication channels has
led to an unprecedented increase in user-generated content. While these platforms
facilitate interaction and information exchange, they have also become a medium for
cyberbullying, hate speech, and toxic comments. Such harmful content can lead to
emotional distress, psychological trauma, and negative societal impacts, especially among
vulnerable users.

Detecting toxic comments manually is a complex and time-consuming task due to the
vast volume of data generated every second. Although automated systems have been
developed to address this issue, most existing solutions are limited to English language
processing and do not effectively support Indian regional languages or code-mixed text.
This limitation results in a significant portion of toxic content remaining undetected.

Furthermore, social media language is often informal, containing slang, abbreviations,

misspellings, and contextual nuances, making it difficult for traditional models to
accurately classify comments. Therefore, there is a need to develop an intelligent,
scalable, and efficient system capable of detecting toxic comments across multilingual
and code-mixed Indian language contexts.

The proposed system aims to automatically identify toxic and abusive comments, reduce
manual moderation efforts, and promote a safe and respectful online communication environment.

Dept. of AI&DS, EPCET 2025-26 1

 Cyber Bullying and Toxic Comment Detection for Indian languages

Chapter 2:
LITERATURE REVIEW

Cyberbullying detection and toxic comment classification have become important

research areas due to the rapid growth of social media platforms and online
communication. Researchers have explored various machine learning and deep learning
techniques to automatically identify harmful content in user-generated text. In recent
years, Natural Language Processing (NLP) has played a significant role in improving the
accuracy and efficiency of such systems.

Author Title Year Key Merit Limitation

Contribution
Proposes a
Cyberbullyin MURIL- Limited to
g Detection based model for Hinglish;
Devesh 2025 detecting toxic Good
in Hinglish accura requires
Kumar
Text Using comments in large
cy for
MURIL and Hinglish text with dataset;
code-
Explainable explainable AI high
mixed
AI support. computatio
text;
supports nal cost.
multiling
ual
input;
provides
interpret
ability.
Uses BERT and High
Securing Requires
BiLSTM accura
Social high
Rohan 2024 models to cy
Spaces: resources;
Biswas et classify toxic (~89%
al. Harnessing large training
comments );
Deep data needed;
using strong
Learning to not ideal for
contextual context
Eradicate real- time
understanding. understand
Cyberbullying use.
i
ng;
effective
for large
datasets.
Dept. of AI&DS, EPCET 2025-26 2
 Cyber Bullying and Toxic Comment Detection for Indian languages
Detecting Combines CNN Supports Depends on
and
Advaitha Gendered 2024 BiLSTM to detect multiple datase
Vetagiri et Abuse in abusive language in languages t
al.
Hindi, Tamil, multilingual ; good qualit
and Indian datasets. y;
English struggles
with
feature slang;
extraction computationa
l
; ly expensive.
improved
performa
nce.
Compares ML and Easy to
Detection of DL models for toxic implement; Limited
Cyberbullying comment detection good multilingu
Deepana 2023 baseline
Using Machine using NLP al support;
M.E. et al. Learning and techniques. performance;
lower
Deep Learning highlights
role of accuracy
than

Dept. of AI&DS, EPCET 2025-26 3

 Cyber Bullying and Toxic Comment Detection for Indian languages

Chapter 3:
GAP IDENTIFICATION

Based on the literature review, several limitations in existing cyberbullying detection

systems have been identified:

 Limited Multilingual Support:

Most existing models are primarily designed for English or a few selected
languages and do not effectively support the wide variety of Indian regional
languages.
 Poor Handling of Code-Mixed Text:

Social media users in India frequently use code-mixed languages such as

Hinglish. Existing systems struggle to accurately process and classify such
mixed-language inputs.
 High Computational Requirements:

Advanced deep learning models like BERT and transformer-based approaches

require significant computational resources and large datasets, making them less
suitable for real- time and low-resource environments.
 Inability to Handle Informal Language:

Social media content often includes slang, abbreviations, misspellings, and

emojis, which are difficult for many models to interpret correctly.
 Lack of Real-Time Scalable Systems:

Many existing approaches are not optimized for real-time deployment and cannot
efficiently handle large volumes of streaming data.

Conclusion of Gap

To address these challenges, the proposed system focuses on developing an efficient and
scalable toxic comment detection model that supports multilingual and code-mixed
Indian text, while maintaining good accuracy and performance for real-time applications.

Dept. of AI&DS, EPCET 2025-26 4

 Cyber Bullying and Toxic Comment Detection for Indian languages

Chapter 4:
METHODOLOGY
The proposed system for cyberbullying and toxic comment detection follows a
structured pipeline consisting of multiple stages, starting from data collection to final
prediction. The methodology is designed to efficiently process multilingual and code-
mixed text and classify comments as toxic or non-toxic.

Stage 1 — Data Collection and Preprocessing

The dataset is collected from publicly available sources such as Kaggle and research
repositories. It contains social media comments labeled as toxic and non-toxic. The
dataset includes English, Hindi, and code-mixed (Hinglish) text to ensure better real-
world applicability.

Stage 2 — Data Preprocessing

Raw text data is cleaned and normalized to improve model performance. This step includes:

 Removal of punctuation, special characters, and URLs

 Conversion of text to lowercase

 Removal of stopwords

 Tokenization of sentences into words

Preprocessing ensures that the input data is consistent and suitable for feature extraction.

Stage 3 — Feature Extraction

The cleaned text is converted into numerical form using TF-IDF (Term Frequency–Inverse
Document Frequency). This technique assigns importance to words based on their
frequency in the dataset, allowing the model to focus on meaningful features.

Stage 4 — Model Development

Multiple machine learning algorithms are implemented to classify comments:

Dept. of AI&DS, EPCET 2025-26 6

 Cyber Bullying and Toxic Comment Detection for Indian languages

 Logistic Regression

 Naïve Bayes

 Support Vector Machine (SVM)

 Random Forest

Each model is trained on the dataset to learn patterns associated with toxic and non-toxic
comments

Stage 5 — Model Evaluation

The performance of each model is evaluated using standard metrics:

 Accuracy
 Precision
 Recall
 F1 Score
A confusion matrix is also used to analyze classification results and identify errors.

Stage 6 — Model Selection

Based on evaluation results, the best-performing model is selected by comparing accuracy

and overall performance metrics. The selected model is used for final deployment.

Stage 7 — System Deployment

A simple user interface is developed using tools like Streamlit or Flask. Users can input
comments, and the system provides real-time predictions indicating whether the comment
is toxic or non- toxic.

Dept. of AI&DS, EPCET 2025-26 6

 Cyber Bullying and Toxic Comment Detection for Indian languages

Source: Generated using ChatGPT (AI-based diagram creation tool)

Dept. of AI&DS, EPCET 2025-26 7

 Cyber Bullying and Toxic Comment Detection for Indian languages

Chapter 5:
DATA ACQUISITION

The datasets used in this study are collected from publicly available and widely used
text repositories on Kaggle. These datasets provide labeled social media comments for
detecting cyberbullying and toxic behavior. The data is primarily composed of user-
generated content from platforms such as Twitter, online forums, and comment
sections.

1.1 Toxic Comment Dataset

Source: Kaggle (Jigsaw dataset)
Link: Toxic Comment Classification Dataset
Classes:
 Toxic
 Non-Toxic
Total Samples: ~150,000+ comments
Samples per Class (approx.):
 Toxic: ~20,000
 Non-Toxic: ~130,000
Data Type: Text data (user comments from online platforms)

1.2 Alzheimer’s Disease Dataset

Source: Kaggle / Research datasets
Classes:
 Abusive
 Non-Abusive
Total Samples: ~20,000+ comments
Samples per Class (approx.):
 Abusive: ~8,000
 Non-Abusive: ~12,000
Data Type: Code-mixed text (Hindi-English / Hinglish)
Dept. of AI&DS, EPCET 2025-26 8
 Cyber Bullying and Toxic Comment Detection for Indian languages

Source: Kaggle / Research datasets

Classes:
 Abusive
 Non-Abusive
Total Samples: ~20,000+ comments
Samples per Class (approx.):
 Abusive: ~8,000
 Non-Abusive: ~12,000
Data Type: Code-mixed text (Hindi-English / Hinglish)

Dept. of AI&DS, EPCET 2025-26 9

 Cyber Bullying and Toxic Comment Detection for Indian languages

1.3 Social Media Comment

Source: Public datasets
(Twitter/Online forums)
Classes:

 Offensive

 Normal

Total Samples: ~10,000+ comments

Samples per Class (approx.):
 Offensive: ~4,000
 Normal: ~6,000
Data Type: Informal text with slang, abbreviations, and emojis

1.4 Data Splitting Strategy

The dataset is divided using a stratified approach to maintain class balance:
 Training Set: 80%
 Validation Set: 10%
 Test Set: 10%
This ensures that both toxic and non-toxic classes are equally represented across all
subsets, improving model reliability and performance.

1.5 Text Context Consideration

Since the dataset consists of individual comments, contextual understanding is
important for accurate classification. To improve performance:
 Sequential text patterns are considered during preprocessing
 Important keywords and phrases are captured using TF-IDF
 Contextual meaning is partially handled through feature extraction
This helps the model better understand user intent and detect subtle forms of cyberbullying.

Dept. of AI&DS, EPCET 2025-26 10

 Cyber Bullying and Toxic Comment Detection for Indian languages

Chapter 6:
PURPOSE
The primary purpose of this project is to develop an efficient and automated system for
detecting cyberbullying and toxic comments in online platforms. With the increasing use
of social media and digital communication, harmful and abusive content has become a
major concern. This project aims to use machine learning and natural language processing
techniques to identify such content and classify it accurately..

1.6 Detection of Toxic Content

The system is designed to analyze user-generated text and identify whether a comment is
toxic or non-toxic. It focuses on detecting abusive language, offensive words, and
harmful expressions present in social media comments.

1.7 Support for Multilingual and Code-Mixed Text

The project aims to handle multilingual data, including English, Hindi, and code-mixed
text such as Hinglish. This ensures that the system is capable of working in real-world
scenarios where users frequently mix languages.

1.8 Reduction of Manual Moderation Effort

The system helps reduce the need for manual content moderation by automatically filtering
harmful comments. This improves efficiency and allows faster decision-making in online
platforms.

1.9 Improvement of Online Safety

By detecting and controlling cyberbullying, the system contributes to creating a safer and
more respectful online environment. It helps protect users from harassment and abusive
interactions.

1.10 Academic and Practical Learning

Dept. of AI&DS, EPCET 2025-26 11

 Cyber Bullying and Toxic Comment Detection for Indian languages
The project demonstrates the application of machine learning and NLP techniques
such as preprocessing, feature extraction, and classification. It provides practical
knowledge and experience in solving real-world problems.

1.11 Data Analysis

Data analysis is performed to understand the distribution and characteristics of the dataset.
Exploratory analysis is used to examine the proportion of toxic and non-toxic comments
and identify any class imbalance. Basic visualizations such as charts are used to represent
the data distribution.
Frequent words in toxic comments are analyzed to understand common abusive patterns.
The dataset is also reviewed after preprocessing to ensure that noise and unnecessary
elements are removed.

Dept. of AI&DS, EPCET 2025-26 12

 Cyber Bullying and Toxic Comment Detection for Indian languages

Chapter 7:
EXPERIMENTAL FRAMEWORK

Experimental Procedure

Step 1: Download toxic comment datasets from Kaggle; verify total samples and class
distribution (toxic vs non-toxic)

Step 2: Perform preprocessing — remove punctuation, URLs, and special characters;

convert text to lowercase; apply tokenization and stopword removal

Step 3: Prepare balanced dataset by handling class imbalance (sampling techniques if

required) and create training, validation, and test splits

Step 4: Apply feature extraction using TF-IDF to convert text data into numerical vectors

Step 5: Train multiple machine learning models (Logistic Regression, Naïve

Bayes, SVM, Random Forest) on the training dataset

Step 6: Evaluate each model on validation data and compare performance using metrics
such as accuracy, precision, recall, and F1 score

Step 7: Select the best-performing model based on evaluation results and fine-tune
parameters if necessary

Step 8: Test the final model on unseen test data and analyze classification performance
using confusion matrix

Step 9: Integrate the trained model into a pipeline for real-time prediction of user input comments

Dept. of AI&DS, EPCET 2025-26 13

 Cyber Bullying and Toxic Comment Detection for Indian languages
Step 10: Deploy the complete system using Streamlit or a web interface and

validate end-to-end functionality with sample inputs

Source: Generated using ChatGPT (AI-based diagram creation tool)

Dept. of AI&DS, EPCET 2025-26 14

 Cyber Bullying and Toxic Comment Detection for Indian languages

Chapter 8:
RESULT, ANALYSIS, AND DISCUSSION

The following results are anticipated based on the proposed architecture and datasets.
Actual results will be updated upon completion of training and evaluation.

Expected Gate Model Performance

• Target accuracy: ≥ 85–90% on the test dataset

• Precision (Toxic Class): ≥ 85%

• Recall (Toxic Class): ≥ 83%

• F1 Score: ≥ 84%

Expected Model Comparison

Model Expected Expected F1 Benchmark

Accuracy Score

Logistic Regression ≥ 88% ≥ 0.86% Standard baseline

model
Support Vector ≥ 90% ≥ 0.88% Common NLP
Machine benchmark

Naïve Bayes ≥ 82% ≥ 0.80% Fast text

classifier

Random Forest ≥ 85% ≥ 0.83% Ensemble

method

Class Imbalance Impact

Since the dataset contains more non-toxic comments than toxic ones, class imbalance may
affect model performance. Techniques such as resampling or weighting are expected to
improve recall for toxic class detection.

Dept. of AI&DS, EPCET 2025-26 15

 Cyber Bullying and Toxic Comment Detection for Indian languages

Text Complexity Analysis

Handling informal language, slang, abbreviations, and code-mixed text (Hinglish) is

expected to slightly reduce accuracy. However, preprocessing and TF-IDF feature
extraction help in improving model performance.

Dept. of AI&DS, EPCET 2025-26 16

 Cyber Bullying and Toxic Comment Detection for Indian languages

Chapter 9:
SIGNIFICANCE AND IMPLICATIONS

The proposed cyberbullying detection system has significant importance in both

academic and real-world applications. It provides an effective solution for identifying
and filtering toxic content from user-generated text on online platforms.

From a social perspective, the system helps in reducing cyberbullying and promoting a
safer digital environment. By automatically detecting harmful comments, it protects
users from online harassment and encourages respectful communication.

From a technical perspective, the project demonstrates the practical implementation of

machine learning and natural language processing techniques such as text preprocessing,
feature extraction, and classification. It highlights how these techniques can be applied to
solve real- world problems efficiently.

From a practical perspective, the system can be integrated into social media platforms,
chat applications, and online forums to assist moderators in filtering inappropriate
content. This reduces manual effort and improves the speed and accuracy of moderation.
Overall, the system contributes to improving online safety, enhancing user
experience, and supporting automated content moderation systems.

Dept. of AI&DS, EPCET 2025-26 17

 Cyber Bullying and Toxic Comment Detection for Indian languages

Chapter 10:
POTENTIAL APPLICATIONS

The proposed cyberbullying detection system can be applied across various domains where
user- generated content is widely used. It is particularly useful in social media platforms
such as Facebook, Instagram, and Twitter, where large volumes of comments are
generated ежедневно.

The system can automatically detect and filter toxic content before it reaches other users.

In chat applications and messaging platforms, the system can monitor conversations in real
time and help prevent abusive or harmful interactions. This improves user safety and
promotes respectful communication.

The system can also be used in online gaming platforms, where toxic behavior is common,
to maintain a healthy and positive environment among players. Additionally, it can be
implemented in educational platforms to ensure safe interaction between students and
instructors.

Furthermore, the system can be applied in customer feedback and review analysis to
identify inappropriate or offensive comments. This helps organizations maintain a
professional environment and improve content quality.

Overall, the system has wide applicability in any domain that involves online
communication and user interaction.

Dept. of AI&DS, EPCET 2025-26 18

 Cyber Bullying and Toxic Comment Detection for Indian languages

REFERENCES

[1] J. H. Friedman, “Greedy function approximation: A gradient boosting machine,”

Annals of Statistics, vol. 29, no. 5, pp. 1189–1232, Oct. 2001.

[2] J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “BERT: Pre-training of deep
bidirectional transformers for language understanding,” in Proc. NAACL-HLT,
Minneapolis, MN, USA, 2019,
pp. 4171–4186.

[3] T. Davidson, D. Warmsley, M. Macy, and I. Weber, “Automated hate speech detection
and the problem of offensive language,” in Proc. ICWSM, 2017, pp. 512–515.

[4] Z. Zhang, D. Robinson, and J. Tepper, “Detecting hate speech on Twitter using a convolution-
GRU based deep neural network,” in Proc. ESWC, 2018, pp. 745–760.

[5] P. Badjatiya, S. Gupta, M. Gupta, and V. Varma, “Deep learning for hate speech detection in
tweets,” in Proc. WWW Companion, 2017, pp. 759–760.

[6] A. Joulin, E. Grave, P. Bojanowski, and T. Mikolov, “Bag of tricks for efficient text
classification,” in Proc. EACL, 2017, pp. 427–431.

[7] S. Bird, E. Klein, and E. Loper, Natural Language Processing with Python, O’Reilly Media,
2009.

[8] Jigsaw, “Toxic Comment Classification Challenge Dataset,” Kaggle, 2018. [Online]. Available:
[Link]/c/jigsaw-toxic-comment-classification-challenge

[9] Kaggle, “Multilingual Toxic Comment Classification Dataset,” 2020. [Online].

Available: [Link]
Dept. of AI&DS, EPCET 2025-26 19
 Cyber Bullying and Toxic Comment Detection for Indian languages

[10] Scikit-learn Developers, “Scikit-learn: Machine Learning in Python,” Journal of

Machine Learning Research, vol. 12, pp. 2825–2830, 2011.

Dept. of AI&DS, EPCET 2025-26 20