Toxic Comment Classification and Discord Bot

Table of Contents

• Overview
• Models
• Dataset
• Word Embedding
• Experimental Tracking
• Best Model
• Discord Bot Deployment
• Usage
• Results
• License
• Acknowledgements
• Contact

Overview

This repository presents a comprehensive toxic comment classification project that encompasses preprocessing, model training, experimental tracking, and real-time deployment as a Discord bot. The primary objectives are:

• Train and evaluate three distinct models (LSTM, BiLSTM, BiLSTM+CNN) for toxic comment classification.
• Employ the Kaggle dataset, comprising comments categorized into six levels of toxicity.
• Leverage Sentence Transformer for word embedding, enhancing model performance.
• Utilize MLflow for efficient experimental tracking and model comparison.
• Deploy the best-performing model (BiLSTM) as a Discord bot capable of identifying and deleting toxic messages in real-time.

Models

The project explores three advanced deep learning architectures:

1. LSTM Model: Implements a Long Short-Term Memory (LSTM) network for comment classification.
2. BiLSTM Model: Utilizes a Bidirectional LSTM (BiLSTM) network to capture bidirectional context.
3. BiLSTM + CNN Model: Combines a Bidirectional LSTM with a Convolutional Neural Network (CNN) for enhanced feature extraction.

Dataset

The dataset is sourced from Kaggle and consists of comments labeled across six toxicity levels. The dataset contains a total of 160,000 samples, each with associated toxicity labels.

Word Embedding

The project harnesses Sentence Transformer for word embedding. This technique enhances model performance by capturing intricate semantic relationships between words.

Experimental Tracking

Experimental tracking and comparison of different model iterations are facilitated through MLflow. This ensures comprehensive monitoring and evaluation of various metrics.

Best Model

After extensive evaluation, the BiLSTM model has emerged as the top-performing architecture, showcasing remarkable accuracy, F1-score, precision, and recall.

Discord Bot Deployment

The most successful model, BiLSTM, has been deployed as a Discord bot. This bot effectively identifies and promptly deletes toxic messages, contributing to a healthier online conversation environment.

Results

Performance evaluation of the models yielded the following key metrics:

Model	Accuracy	F1-Score	Precision	Recall
LSTM	0.97	0.97	0.971	0.97
BiLSTM	0.973	0.972	0.971	0.973
BiLSTM + CNN	0.969	0.97	0.971	0.969

Acknowledgements

We extend our gratitude to Kaggle for providing the dataset and the open-source community for the invaluable tools and libraries used throughout this project.

Contact

For inquiries or feedback, please contact Samman Shrestha(mailto:shresthasamman125@gmail.com).