Multilingual Cyberbullying Detection

A machine learning project to detect cyberbullying across multiple languages using mBERT (Multilingual BERT) and sentiment analysis. This project focuses on providing a context-aware and language-agnostic solution for detecting harmful online behavior in Hindi, Marathi, and Malayalam.

Table of Contents

• Introduction
• Screenshots
• Features
• Project Architecture
• Datasets
• Running
• Preprocessing
• Model
• Evaluation Metrics
• Installation
• Usage
• Future Work
• Authors

Introduction

With the rise of social media, cyberbullying has become a critical issue affecting individuals across diverse backgrounds and age groups. This project aims to create a robust solution capable of identifying cyberbullying in various languages, providing insights into harmful interactions and promoting safer online spaces.

Screenshots

Features

• Multilingual Support: Supports cyberbullying detection in Hindi, Marathi, and Malayalam.
• Sentiment Analysis Integration: Analyzes the emotional tone of input text to improve accuracy.
• Real-Time Classification: Classifies text as cyberbullying or non-cyberbullying instantly.
• Scalable Architecture: Ready for deployment on platforms like Heroku or AWS.

Project Architecture

The model architecture consists of several stages:

1. Input Layer: Receives text data in Hindi, Marathi, or Malayalam.
2. Preprocessing: Cleans and normalizes the text by removing special characters, stop words, and performing lemmatization.Cleans and normalizes the text by removing special characters, stop words, and performing lemmatization.
3. Feature Extraction:
   • TF-IDF: Extracts term frequency-inverse document frequency values.
   • mBERT Embeddings: Generates multilingual embeddings to capture language context.
4. Logistic Regression Classifier: Combines TF-IDF and mBERT features for binary classification.
5. Output Layer: Outputs binary labels for cyberbullying or non-cyberbullying.

Datasets

• Source: Data collected from social media platforms for sentiment analysis, with samples labeled as cyberbullying or non-cyberbullying.
• Languages: Hindi, Marathi, and Malayalam.
• Size: Over 30,000 labeled text samples to ensure diverse representation and robust model training.

Preprocessing

The preprocessing pipeline includes:

• Text Cleaning: Removing special characters, URLs, hashtags, and mentions.
• Lowercasing: Converting text to lowercase.
• Stop Words Removal: Eliminating common, irrelevant words.
• Lemmatization: Reducing words to their base forms.

Model

The project uses:

• mBERT (Multilingual BERT): A transformer model that provides robust, multilingual embeddings.
• Logistic Regression: A classifier trained on combined TF-IDF and mBERT embeddings to distinguish cyberbullying from non-cyberbullying effectively.

Evaluation Metrics

The model is evaluated on:

• Accuracy
• Precision
• Recall
• F1-Score

These metrics ensure balanced performance across classes and highlight the model’s ability to minimize false positives and negatives.

Installation

1. Clone the repository

   git clone https://github.com/yourusername/Multilingual-Cyberbullying-Detection.git
   cd Multilingual-Cyberbullying-Detection

2. Install Dependencies

   pip install -r requirements.txt

3. Download mBERT Model (if not included in the repository)

   from transformers import AutoTokenizer, AutoModel
   tokenizer = AutoTokenizer.from_pretrained("bert-base-multilingual-cased")
   model = AutoModel.from_pretrained("bert-base-multilingual-cased")

Usage

1. Run the Preprocessing Script

    python preprocess.py

2. Train the Model

   python train.py

3. Test the Model Input text samples to see classification results for cyberbullying detection.

Future Work

• Deep Learning Integration: Experiment with LSTM, CNN, or hybrid models for enhanced accuracy.
• Real-Time Monitoring: Develop capabilities for live monitoring of social media content.
• Expanded Language Support: Extend the model to detect cyberbullying in additional languages.

Authors

• Manish Kumar
• Sumriddhi Tonk
• Vrishin Gaikwad
• Sivaram Vinod