🏆 Wise@LT-EDI 2025: Multi-scale Ensemble Learning for Tamil-English Code-Mixed Hate Speech Detection

📋 Table of Contents

• 🎯 Overview
• 🏆 Competition Results
• 👥 Team
• 📊 Dataset
• 🔬 Methodology
• 🏗️ Architecture
• 📈 Results
• 🚀 Quick Start
• 📁 Repository Structure
• 🔧 Dependencies
• 💡 Key Findings
• ⚠️ Limitations
• 🙏 Acknowledgments
• 📄 License

🎯 Overview

This repository contains our award-winning submission to the LT-EDI 2025 shared task on caste and migration-based hate speech detection in code-mixed Tamil social media content. Our innovative multi-scale hybrid architecture combines classical TF-IDF features with state-of-the-art transformer embeddings, achieving 3rd place with an impressive F1-score of 0.818.

🔥 Key Highlights

• 🥉 3rd Place in LT-EDI 2025 Competition
• 📊 F1-Score: 0.818 on Tamil-English code-mixed hate speech detection
• 🤖 5 Transformer Models: Google BERT, XLM-RoBERTa (Base & Large), IndicBERT, Seanbenhur BERT
• 📐 Multi-scale Features: 512, 1024, and 2048 dimensional representations
• 🎯 Hierarchical Ensemble: 3-level majority voting strategy
• 🔬 Hybrid Approach: Classical TF-IDF + Modern transformers

🏗️ Multi-scale Hybrid Architecture Overview

🏆 Competition Results

Our solution achieved outstanding performance in the LT-EDI 2025 shared task:

👥 Team

📊 Dataset

📈 Dataset Statistics

🏷️ Label Distribution

📝 Sample Data

🔸 Non-hate: "Good morning anna, how are you doing today?"
🔹 Hate: "[Contains offensive content targeting specific communities]"

Note: The dataset contains Tamil-English code-mixed social media comments with varying degrees of code-switching and transliteration.

🔬 Methodology

Our approach implements a sophisticated multi-stage pipeline that combines classical and modern NLP techniques.

🛠️ Stage 1: Data Preprocessing

Input Text → Replace newlines → Convert emojis → Remove URLs → 
Normalize spaces → Convert to lowercase → Output

Input Text → Basic cleaning → Tamil Unicode to English → 
Remove repeated chars → Remove non-ASCII → Output

🎯 Stage 2: Feature Extraction

📊 TF-IDF Features

graph LR
    A[Raw Text] --> B[TF-IDF Vectorizer]
    B --> C[~22,000 dimensions]
    C --> D[Truncated SVD]
    D --> E[512/1024/2048 dims]
    E --> F[MLP Compression]
    F --> G[32/64/128 dims]

🤖 Transformer Embeddings

🏗️ Stage 3: Multi-level Ensemble Architecture

graph TD
    A[TF-IDF Features] --> D[Feature Fusion]
    B[BERT Embeddings] --> D
    C[RoBERTa Embeddings] --> D
    D --> E[ML Classifiers]
    E --> F[Level 1: Top-3 Voting]
    F --> G[Level 2: Cross-Model Voting]
    G --> H[Level 3: Multi-Scale Voting]
    H --> I[Final Prediction]

🎲 Machine Learning Classifiers

🌳 Tree-Based Models:

• Random Forest, XGBoost, LightGBM, CatBoost, Gradient Boosting

📏 Linear Models:

• Logistic Regression, Ridge Classifier, SGD Classifier

🔮 Other Models:

• SVM (Linear, RBF, Nu-SVM), KNN, AdaBoost, Neural Networks

🗳️ Hierarchical Voting Strategy

1. Level 1: Top-3 classifiers per feature set
2. Level 2: Cross-model consensus per dimension
3. Level 3: Multi-scale final decision

🏗️ Architecture

🔧 Comprehensive System Architecture

📈 Results

🎯 Performance by Dimension and Model

📊 Dimension: 512

📊 Dimension: 1024

📊 Dimension: 2048 ⭐

🏆 Final Competition Standings

🎯 Achievement: Our team secured 3rd place out of numerous international participants, demonstrating the effectiveness of our multi-scale ensemble approach!

📊 Training Performance Graphs

📊 Dimension: 512

🔍 Click to view 512-dimensional training graphs

🤖 Model	📈 Training Graph
Google BERT
IndicBERT
Seanbenhur BERT
XLM-RoBERTa Base
XLM-RoBERTa Large

📊 Dimension: 1024

🔍 Click to view 1024-dimensional training graphs

🤖 Model	📈 Training Graph
Google BERT
IndicBERT
Seanbenhur BERT
XLM-RoBERTa Base
XLM-RoBERTa Large

📊 Dimension: 2048 ⭐

🔍 Click to view 2048-dimensional training graphs

🤖 Model	📈 Training Graph
Google BERT
IndicBERT
Seanbenhur BERT
XLM-RoBERTa Base
XLM-RoBERTa Large

🚀 Quick Start

📋 Prerequisites

# Python 3.8+ required
python --version  # Should be >= 3.8

⚡ Installation

# Clone the repository
git clone https://github.com/your-username/CasteMigrationHateSpeech.git
cd CasteMigrationHateSpeech

# Install dependencies
pip install -r requirements.txt

# Install additional dependencies
pip install torch torchvision transformers
pip install xgboost lightgbm catboost

🔧 Running the Pipeline

1️⃣ Data Preprocessing

# Navigate to preprocessors
cd "DL Feature set 1024/Processors"

# Run preprocessing notebooks
jupyter notebook with_transliteration_processor.ipynb
jupyter notebook no_transliteration_processor.ipynb

2️⃣ Feature Generation

# Generate TF-IDF features
jupyter notebook tf_idf_vectorizer.ipynb

# Apply vectorizer
jupyter notebook apply_vectorizer.ipynb

3️⃣ Model Training

# Navigate to trainers
cd "../Trainers"

# Train transformer models
python training_bert.py  # Configure model in script

4️⃣ Ensemble Classification

# Navigate to ML heads
cd "../../ML Classification Heads"

# Run ensemble classification
jupyter notebook code.ipynb

📁 Repository Structure

🏆 CasteMigrationHateSpeech-main/
│
├── 📊 DL Feature set 512/          # 512-dimensional experiments
│   ├── 🤖 Models/                  # Transformer model definitions
│   │   ├── google_bert.py
│   │   ├── indic_model.py
│   │   ├── tanglish_model.py
│   │   ├── xlm_roberta_base.py
│   │   └── xlm_roberta_large.py
│   │
│   ├── 📄 No trans data/           # Non-transliterated features
│   │   ├── test_features_no_trans.csv
│   │   ├── test_processed_no_trans.csv
│   │   └── tfidf_vectorizer_2.pkl
│   │
│   ├── 📊 Original Data/           # Raw dataset
│   │   └── test.csv
│   │
│   ├── 🔧 Processors/              # Data preprocessing scripts
│   │   ├── apply_vectorizer.ipynb
│   │   ├── no_transliteration_processor.ipynb
│   │   ├── tf_idf_vectorizer.ipynb
│   │   └── with_transliteration_processor.ipynb
│   │
│   ├── 📈 Train Data/              # Training results & visualizations
│   │   ├── Graphs/                 # Training performance plots
│   │   ├── Logs/                   # Training logs
│   │   └── Results/                # Model outputs
│   │
│   └── 🏋️ Trainers/                # Training infrastructure
│       ├── classification_features_generator.ipynb
│       ├── dataset.py
│       ├── logger.py
│       ├── trainer.py
│       └── training_bert.py
│
├── 📊 DL Feature set 1024/         # 1024-dimensional experiments
│   └── [Same structure as 512]
│
├── 📊 DL Feature set 2048/         # 2048-dimensional experiments  
│   └── [Same structure as 512]
│
├── 🔄 Trans data/                  # Transliterated features
│   ├── test_features.csv
│   ├── test_processed.csv
│   └── tfidf_vectorizer_1.pkl
│
├── 🖼️ Images/                       # Documentation assets
│   └── overview.png
│
├── 🎯 ML Classification Heads/     # Ensemble classification
│   ├── size 512/                   # 512-dim ensemble results
│   ├── size 1024/                  # 1024-dim ensemble results
│   ├── size 2048/                  # 2048-dim ensemble results
│   ├── code.ipynb                  # Main ensemble script
│   ├── final_1024.csv             # Final predictions
│   ├── final_2048.csv
│   ├── final_512.csv
│   └── best_submission.csv         # Competition submission
│
└── 📖 README.md                    # This file

🔧 Dependencies

📦 Core Dependencies

🏷️ Package	🔢 Version	📝 Purpose
	3.8+	Programming language
	2.0+	Deep learning framework
	4.0+	Pre-trained models
	1.0+	Machine learning

📊 ML Libraries

📥 Installation Command

pip install torch>=2.0 transformers>=4.0 scikit-learn>=1.0 \
    xgboost lightgbm catboost pandas numpy nltk emoji \
    matplotlib seaborn tqdm jupyter

💡 Key Findings

🎯 Technical Insights

📈 Performance Patterns

🔥 Best Performing: 2048-dimensional features
🎯 Most Consistent: Random Forest classifiers  
🚀 Biggest Surprise: Gradient Boosting on Google BERT
📊 Sweet Spot: 1024-dimensional IndicBERT