POLAR: Multilingual Polarization Detection

Overview

three interconnected subtasks:

1. Subtask 1: Polarization Detection (Binary Classification)

   • Detect if text contains polarized content
   • Models: SVM, XLM-RoBERTa, RemBERT

2. Subtask 2: Polarization Type Classification (Multi-label)

   • Classify types: Political, Racial/Ethnic, Religious, Gender/Sexual, Other
   • Multi-label classification with 5 categories

3. Subtask 3: Manifestation Identification (Multi-label)

   • Identify how polarization is expressed
   • 6 manifestation types: Stereotype, Vilification, Dehumanization, Extreme Language, Lack of Empathy, Invalidation

Models

We use three models for all subtasks:

1. SVM (Statistical Model - Baseline)

   • TF-IDF feature extraction
   • Fast training and inference
   • Good baseline for comparison

2. XLM-RoBERTa (122M parameters)

   • Multilingual transformer (100+ languages)
   • Good balance of quality and speed
   • ~30-60 min training on GPU

3. RemBERT (568M parameters)

   • Google's Retrieval-based multilingual BERT
   • Highest quality predictions
   • ~1-2 hours training on GPU

Quick Start

Installation

# Create virtual environment
python -m venv venv
source venv/bin/activate  # Windows version: venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

Train First Model

# Run from project root (not from subtask1 directory)
python subtask1/src/train.py --config subtask1/config/config.yaml

View Results

# Check logs
tail logs/training_*.log

# View trained models
ls models/

Project Structure

NLP_Proj/
├── subtask1/               # Polarization Detection
│   ├── config/config.yaml  
│   ├── data/              
│   ├── src/train.py        # script
│   ├── models/             
│   └── logs/               
├── subtask2/               # Polarization Type (Multi-label)
├── subtask3/               # Manifestation Identification (Multi-label)
├── shared_utils/           # Shared utilities
│   ├── utils.py            # Configuration, logging
│   ├── data_loader.py      # Data loading & preprocessing
│   ├── metrics.py          # Evaluation metrics
│   └── inference.py        # Inference utilities
├── requirements.txt        # Python dependencies
└── README.md              

Configuration System

All parameters are configurable via YAML files :

SVM Configuration --- No need to change this I think (?)

• Feature extraction (TF-IDF settings)
• Kernel selection (linear, rbf, poly, sigmoid)
• Regularization parameters

Transformer Configuration (XLM-RoBERTa & RemBERT)

• Model selection
• Tokenizer settings (max_length, padding)
• Training parameters:
  • Batch size
  • Learning rate & scheduler
  • Number of epochs
  • Early stopping
  • Gradient accumulation
  • Optimizer selection
• Mixed precision training
• GPU/CPU device configuration

Training

Subtask 1 (Polarization Detection)

python subtask1/src/train.py \
    --config subtask1/config/config.yaml \
    --data-root subtask1/data \
    --seed 42

Output:

• Trained models in subtask1/models/
• Training logs in subtask1/logs/
• Per-language metrics

Subtask 2 (Polarization Type)

python subtask2/src/train.py --config subtask2/config/config.yaml

Multi-label classification for 5 polarization types

Subtask 3 (Manifestation Identification)

python subtask3/src/train.py --config subtask3/config/config.yaml

Multi-label classification for 6 manifestation types

Make Predictions

python shared_utils/inference.py \
    --config subtask1/config/config.yaml \
    --input-file new_data.csv \
    --model-type xlm_roberta \
    --output-dir predictions/

References

• Competition: https://www.codabench.org/competitions/10669/
• XLM-RoBERTa: https://huggingface.co/xlm-roberta-base
• RemBERT: https://huggingface.co/google/rembert