CoolFace
Apppublic

khaoula-ghz/Cyberbullying-Detection

sourceHugging Faceapache-2.0updated 6mo agoView on Hugging Face
0likes
App README

Multilingual Cyberbullying Detection System

Detecting Abusive Content in English, Arabic & Algerian Darija

An AI-powered system for detecting cyberbullying and abusive content across three languages. Built with transformer models, CNN feature extraction, contrastive learning, and ensemble learning on three datasets.

Performance: English 95.55% F1 • Arabic 68.31% F1 • Algerian Darija 85.84% F1

Try Live Demo • How to Use • Quick Start • Architecture • Datasets ---

Why This Project?

Most cyberbullying detection systems only work in English. This project provides detection for:

  • —English: Well-resourced language with abundant training data (95.55% F1)
  • —Arabic: Complex morphology and multiple dialects (68.31% F1)
  • —Algerian Darija: Extremely low-resource, no standardized written form, first dataset ever created (85.84% F1)

How to Use the App

No coding knowledge required. Just visit the live demo and start analyzing text.

Getting Started (3 Steps)

  1. 1.Open the demo: https://huggingface.co/spaces/khaoula-ghz/Cyberbullying-Detection
  1. 1.Enter your text in English, Arabic, or Algerian Darija
  1. 1.View results including confidence score and individual model votes

What You'll See

The app provides:

  • —Clear Verdict: Cyberbullying or Safe
  • —Confidence Score: 0-100% certainty
  • —Model Agreement: How many models agree
  • —Detailed Breakdown: Charts and visualizations

Quick Start

Installation & Running

bash
# Clone and install
git clone https://github.com/khaoula-ghz/Cyberbullying-Detection.git
cd Cyberbullying-Detection
pip install -r requirements.txt

# Run the app
python app.py

Open http://localhost:7860 in your browser.

Python API

python
from app import ensemble_predict

# Single prediction
result = ensemble_predict(
    text="You're awesome!",
    user_selected_lang="auto"  # auto-detect or specify 'en', 'ar', 'dz'
)

# Check result
if result['is_cyberbullying']:
    print(f"Detected: {result['label']}")
    print(f"Confidence: {result['confidence']:.0%}")
else:
    print("Safe text")

Architecture Overview

Key Components

Transformer Encoders

  • —Generate rich contextual embeddings
  • —Pre-trained on massive multilingual corpora
  • —Fine-tuned on cyberbullying-labeled data

CNN Feature Layer

  • —Sits on top of transformer embeddings
  • —Extracts local n-gram patterns
  • —Specializes in abuse-specific terminology
  • —Improves detection of obfuscated abuse

Contrastive Learning

  • —Pulls similar cyberbullying examples together
  • —Pushes normal text away in embedding space
  • —Improves generalization to unseen patterns

Ensemble Methods

For each language, we use a different ensemble approach:

  • —English: Majority voting from 5 models (BERT, BERTweet, RoBERTa, DistilBERT, HateBERT)
  • —Arabic: Stacking with meta-model from 3 models (AraBERT, MARBERT, QARiB)
  • —Algerian: Soft voting from 3 models (DziriBERT, DziriBERT-Sentiment, mBERT)

Datasets

All datasets are publicly available for research purposes.

English Dataset

  • —Source: Twitter/X cyberbullying detection corpus
  • —Size: ~47,000 labeled tweets
  • —Labels: Age, Ethnicity, Gender, Religion, Other Cyberbullying, Not Cyberbullying
  • —Link: Kaggle - Cyberbullying Classification

Arabic Dataset

  • —Source: Instagram comments and social media posts
  • —Size: ~47,000 labeled comments
  • —Labels: Neutral, Positive, Bullying, Toxic
  • —Dialects: Modern Standard Arabic + regional variations
  • —Link: Dataset Repository (Accessed: February 3, 2025)

Algerian Darija Dataset (Original Contribution)

  • —Source: Authentic Algerian social media comments
  • —Size: ~4,000 manually annotated comments
  • —Labels: Cyberbullying, Not Cyberbullying
  • —Language: Algerian Darija (low-resource dialect)
  • —Significance: First cyberbullying detection dataset for Algerian Darija
  • —Link: Kaggle - DzBullying Dataset

About the Algerian Dataset: The Algerian Darija dataset was custom-created for this thesis. Since Algerian Darija has no standardized written form, building this dataset required manual annotation of authentic social media comments.

Project Structure

Cyberbullying-Detection/
├── app.py                      # Gradio web interface
├── models.py                   # CNN-Transformer architecture
├── language_detector.py        # Language identification
├── text_preprocessing.py       # Text cleaning pipeline
├── final_meta_model.pkl        # Arabic stacking meta-model
├── requirements.txt            # Dependencies
└── README.md                   # This file

Deploy It

Live Demo: https://huggingface.co/spaces/khaoula-ghz/Cyberbullying-Detection

Deploy your own: Create Space on Hugging Face → Connect GitHub repo → Auto-deploys on push ✨

Contributing

Have ideas to improve? We welcome contributions!

Citation

If you use this work in research, please cite:

bibtex
@mastersthesis{ghimouze2025cyberbullying,
  author = {Ghimouze, Khaoula and Kemcha, Rania Rym},
  title = {Cyberbullying Detection on Social Media: A Transformer-Based Hybrid Approach},
  school = {Constantine 2 University},
  year = {2025},
  type = {Master's Thesis}
}

Datasets:

bibtex
@dataset{ghimouze2025dzbullying,
  author = {Ghimouze, Khaoula and Kemcha, Rania Rym},
  title = {DzBullying: Algerian Cyberbullying Detection Dataset},
  year = {2025},
  howpublished = {Kaggle},
  url = {https://www.kaggle.com/datasets/khaoulaghz23/dzbullying-algerian-cyberbullying-dataset}
}

License

MIT License – Free to use, modify, and distribute with attribution.

Authors

Khaoula Ghimouze Data Science & NLP Specialist, Constantine 2 University GitHub • Email • Hugging Face

Rania Rym Kemcha Research Collaborator & Co-author

<div align="center">

Built to make online spaces safer

Star us on GitHub • Try the Live Demo

</div>