🎯 Masha'erohom (Arabic Sentiment Analysis)

An end-to-end Arabic Emotion Detection Web Application that leverages ensemble deep learning to accurately classify emotions expressed in Arabic text. The system is designed to handle Modern Standard Arabic (MSA) and multiple Arabic dialects, achieving 90% accuracy through a stacking ensemble of state-of-the-art models.

📋 Table of Contents

• Project Overview
• Key Features
• Architecture
• Models & Performance
• Dataset
• Results & Comparisons
• Future Work
• Team
• Citation
• Publication

🚀 Project Overview

With the massive growth of Arabic content on social media platforms, understanding public emotions has become both critical and challenging. Arabic sentiment and emotion analysis is particularly difficult due to:

• Complex morphology
• Dialectal diversity
• Limited high-quality annotated datasets

Mashaerohom addresses this gap by providing a production-ready web application that automatically detects emotions in Arabic text using a powerful ensemble of deep learning models.

The system supports batch analysis (CSV upload) and real-time analysis (Twitter keyword search), accompanied by interactive visual dashboards.

Key Achievement: 90% Accuracy - State-of-the-art performance on Arabic emotion detection

✨ Key Features

🔍 Dual Functionality

1. CSV File Processing

   • Upload Excel/CSV files containing Arabic text
   • Automatic emotion classification
   • Download enriched dataset with emotion labels
   • Interactive dashboard with visualizations

2. Real-time Twitter Analysis

   • Search for keywords on Twitter
   • Retrieve 50 most recent tweets
   • Real-time sentiment analysis
   • Visual sentiment distribution

📊 Visualization Dashboard

• Interactive pie charts and bar graphs
• Percentage distribution tables
• Real-time sentiment tracking
• Historical data analysis

🏗️ Architecture

Three-Tier System Design

Architecture.mp4

Model Pipeline

Input → Preprocessing → [Bi-LSTM, Bi-GRU, MARBERTv2] → Stacking → Random Forest → Output

🤖 Models & Performance

Ensemble Architecture

We combine three powerful models using Random Forest stacking:

Model	Description	Key Features
Bi-LSTM	Bidirectional Long Short-Term Memory	Captures long-term dependencies, bidirectional context
Bi-GRU	Bidirectional Gated Recurrent Unit	Efficient, computationally lighter than LSTM
MARBERTv2	Arabic-specific BERT model	Pre-trained on 1B Arabic tweets, 128GB text data
Random Forest	Ensemble Meta-learner	Combines predictions from all base models

Performance Metrics

Model	Accuracy	F1-Score	Recall	Precision
Bi-GRU	72%	71%	70%	72%
Bi-LSTM	72%	71%	71%	72%
MARBERTv2	81%	80%	79%	81%
Ensemble (RF)	90%	90%	90%	90%

Note: Ensemble model outperforms all individual models significantly!

📊 Dataset

Emotone_ar Dataset

• Size: 10,065 Arabic tweets
• Emotions: 8 categories (Sadness, Anger, Joy, Surprise, Love, Sympathy, Fear, None)
• Dialects: Multiple Arabic dialects
• Annotation: Manually annotated by 3 native Arabic speakers
• Balance: Approximately equal distribution across emotions

Sample Distribution:

📈 Results & Comparisons

State-of-the-Art Comparison

Name	Date	Dataset	Model	Accuracy	F1-score
Text Based Emotion Recognition in Arabic text	2019	Emotone-AR[4]	CNN	0.70	0.70
Textual Emotions	2023	Emotone-AR[4]	BI-GRU	0.73	0.74
Improved Emotion Detection Framework for Arabic Text using Transformer Models	2023	Emotone-AR[4]	arabic-bert-base model	0.74	0.74
Masha'erohom	2024	Emotone-AR[4]	BI-LSTM	0.72	0.71
Masha'erohom	2024	Emotone-AR[4]	BI-GRU	0.72	0.71
Masha'erohom	2024	Emotone-AR[4]	MARBERT	0.81	0.80
Masha'erohom	2024	Emotone-AR[4]	Ensemble RF	0.90	0.90

Confusion Matrix (Ensemble Model)

Actual \ Predicted	Ang	Fea	Joy	Lov	Sad	Sym	Sur	Non
Ang	1296	24	18	12	45	21	15	9
Fea	31	1082	35	18	28	9	2	0
Joy	22	19	1150	32	42	8	7	0
Lov	15	8	29	1120	21	15	5	0
Sad	38	21	45	25	1050	32	33	10
Sym	28	12	18	21	35	920	12	0
Sur	19	5	12	8	28	14	945	13
Non	21	8	15	12	35	18	25	1405

Emotion Abbreviations:

• Ang: Anger
• Fea: Fear
• Joy: Joy
• Lov: Love
• Sad: Sadness
• Sym: Sympathy
• Sur: Surprise
• Non: None

🔮 Future Work

Planned Enhancements

Model Improvements

• Integrate AraT5 for better text understanding
• Add dialect-specific models
• Implement sarcasm and irony detection

Feature Expansion

• Facebook keyword search integration
• Multi-platform social media analysis
• Real-time streaming analysis

Technical Upgrades

• Docker containerization
• Cloud deployment (AWS/Azure)
• Mobile application
• API rate limiting and scaling

Dataset Expansion

• Include more Arabic dialects
• Add news articles and blogs
• Cross-domain sentiment analysis

👥 Team

Supervisors

• Dr. Shaimaa Haridy - Lecturer, Information Systems Department, Ain Shams University

Development Team

• Ali Abdallah
• Mohamed Ali
• Karima Sobhi
• Ali Maher
• Abdulthman Abdelhalim
• Hany Mohamed

Institution

This project was developed as a Bachelor’s Graduation Project at:

Faculty of Computer and Information Sciences, Information Systems Department
Ain Shams University
Cairo, Egypt

📚 Citation

If you use this project in your research, please cite:

@article{arabicsentiment2024,
  title={Arabic Sentiment Analysis using Ensemble Deep Learning Model},
  author={Ali Abdallah, Mohamed Ali, Karima Sobhi, Ali Maher, Abdelrahman Abdelhalim, Hany Mohamed},
  year={2024},
  publisher={Ain Shams University},
  note={Bachelor's Graduation Project}
}

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

🙏 Acknowledgements

We acknowledge with gratitude:

🎓 Supervisory Excellence: Dr. Shaimaa Haridy for her exceptional mentorship, research guidance, and unwavering support that transformed this project into published research.

🏛️ Institutional Support: Ain Shams University for academic resources and Nile University for the Emotone_ar dataset.

⚡ Technical Enablement: Hugging Face, Google Colab, and Kaggle for providing the tools and computational power essential for this deep learning research.

🌍 Open Source Community: Countless contributors whose work forms the foundation of modern NLP research.

Gratitude turns what we have into enough, and research into impact.

---

⭐ If you find this project useful, please give it a star on GitHub!

📧 Contact: For questions or collaborations, please email: [email protected]

🔗 Live Demo:

📖 Documentation: Full Documentation

📚 Publication

Title: Arabic Sentiment Analysis using Ensemble Deep Learning Model
Journal: International Journal of Intelligent Computing and Information Sciences
DOI/Link: https://ijicis.journals.ekb.eg/article_406786.html
Status: ✅ Published