Toxic Comment Classification & Image Generation with GANs

A comprehensive NLP and Computer Vision project portfolio showcasing multiple machine learning techniques for toxic comment detection and image generation.

📋 Table of Contents

• Overview
• Projects
  • Toxic Comment Classification
  • Image Generation with GANs
• Technologies Used
• Project Structure
• Results
• Setup
• Usage

🎯 Overview

This repository contains two main projects:

1. Multi-label Toxic Comment Classification: A comprehensive NLP project comparing multiple techniques (TF-IDF+SVC, BiLSTM, DistilBERT) for detecting toxic content in online comments across 6 categories.
2. Image Generation with GANs: A generative adversarial network implementation for creating realistic handwritten digits using the MNIST dataset.

📊 Projects

Toxic Comment Classification

A multi-label classification system that identifies toxic comments across six categories:

• toxic
• severe_toxic
• obscene
• threat
• insult
• identity_hate

Methodology

1. EDA & Preprocessing (EDA.ipynb)

• Exploratory data analysis of the Kaggle Jigsaw Toxic Comment Classification dataset
• Text cleaning and normalization
• Label distribution analysis
• Stratified train/validation/test split (80/10/10)
• Feature engineering and statistics

2. Model Implementations

• BiLSTM Baseline (BiLSTM Model.ipynb)

  • Bidirectional LSTM with word-level embeddings
  • Mean + max pooling for sentence representation
  • Multi-label classification with BCE loss and class weighting
  • Per-label threshold tuning for optimal F1 scores

• DistilBERT (Final Model) (DistillBERT_ChosenTechnique.ipynb)

  • Fine-tuned DistilBERT transformer model
  • Subword tokenization for robust handling of slang/typos
  • Contextual understanding through self-attention mechanisms
  • Interactive Gradio interface for end-user demonstration
  • Comprehensive evaluation metrics

Key Features

• Class Imbalance Handling: Uses pos_weight in BCE loss to handle rare labels
• Threshold Optimization: Per-label threshold tuning on validation set
• Comprehensive Metrics: Micro-F1, Macro-F1, PR-AUC, Hamming Accuracy
• Reproducibility: Fixed seeds and consistent train/val/test splits
• Production-Ready Interface: Gradio demo for real-time inference

Performance

Model	Micro-F1	Macro-F1	Mean PR-AUC	Hamming Accuracy
DistilBERT	0.789	0.694	0.741	0.984
BiLSTM	0.674	0.567	0.555	0.976

DistilBERT was selected as the final model for its superior performance on rare labels and better contextual understanding.

Image Generation with GANs

A Generative Adversarial Network implementation for generating realistic handwritten digits from the MNIST dataset.

Architecture

• Generator: Maps random noise (100-dim) → 28×28 grayscale images

  • Fully connected layers with ReLU activations
  • Tanh output for normalized pixel values [-1, 1]

• Discriminator: Binary classifier for real vs. generated images

  • Deep feedforward network with dropout regularization
  • Sigmoid output for probability scores

Training Process

• Adversarial training with alternating updates
• Binary cross-entropy loss for both networks
• Adam optimizer (lr=0.0001)
• 30 epochs on MNIST training set

🛠 Technologies Used

NLP Project

• PyTorch: Deep learning framework
• Transformers (Hugging Face): Pre-trained DistilBERT model
• scikit-learn: Data splitting, metrics, and baseline models
• pandas & numpy: Data manipulation and analysis
• matplotlib & seaborn: Visualization
• Gradio: Interactive web interface

GAN Project

• PyTorch: Neural network implementation
• torchvision: MNIST dataset loading and transforms
• matplotlib: Visualization of generated images

📁 Project Structure

.
├── EDA.ipynb                          # Exploratory data analysis and preprocessing
├── BiLSTM Model.ipynb                 # BiLSTM baseline implementation
├── DistillBERT_ChosenTechnique.ipynb  # DistilBERT final model with Gradio interface
├── ImageGeneration.ipynb              # GAN-based image generation
└── README.md                          # This file

📈 Results

Toxic Comment Classification

• Successfully handles multi-label classification with 6 toxicity categories
• Achieves 98.4% Hamming accuracy with DistilBERT
• Strong performance on rare labels (threat, identity_hate) through class weighting
• Interactive demo interface for real-time classification

Image Generation

• Generator learns to produce realistic MNIST-like digits
• Adversarial training achieves equilibrium between generator and discriminator
• Generated images demonstrate learned distribution of handwritten digits

🚀 Setup

Prerequisites

pip install torch torchvision transformers
pip install pandas numpy scikit-learn matplotlib seaborn
pip install gradio

Quality checks

The lightweight CI suite exercises the tokenisation, vocabulary, encoding, text-cleaning, sigmoid, and threshold-tuning functions directly from the committed notebooks. It does not download the training dataset or model weights.

python -m pip install -r requirements-ci.txt
ruff check tests
mypy tests/notebook_loader.py
PYTHONPATH=tests python -m unittest discover -s tests
jupyter nbconvert --to html --ClearOutputPreprocessor.enabled=True \
  --output-dir /tmp/notebook-build *.ipynb

Data

For the toxic comment classification project:

• Download the Kaggle Jigsaw Toxic Comment Classification Challenge dataset
• Place train.csv in the working directory

For the GAN project:

• MNIST dataset is automatically downloaded by PyTorch

💻 Usage

Toxic Comment Classification

1. Run EDA.ipynb to preprocess the data and create train/val/test splits
2. Run BiLSTM Model.ipynb to train the baseline model
3. Run DistillBERT_ChosenTechnique.ipynb to train the final model and launch the Gradio interface

Image Generation

1. Run ImageGeneration.ipynb to train the GAN and generate images
2. Training takes approximately 30 epochs
3. Generated images are displayed in a 4×4 grid

📝 Notes

• All notebooks are designed to run on Google Colab with GPU support
• Model artifacts and metrics are saved for reproducibility
• Content warning: The toxic comment dataset contains offensive language
• Seeds are fixed for reproducible results across all experiments

🔗 Resources

• Kaggle Jigsaw Competition
• DistilBERT Paper
• GAN Paper

📄 License

This project is for educational and portfolio purposes.

---

Author: Luke Nguyen Repository: https://github.com/tuilachit/ToxicComment_NLP_Project