Toxic Comment Classification Challenge

Overview

This project is a solution to the Jigsaw Toxic Comment Classification Challenge, hosted on Kaggle. The goal is to develop a model that can classify online comments into different categories of toxicity.

Dataset

The dataset consists of comments labeled for different types of toxicity, including:

• Toxic
• Severe Toxic
• Obscene
• Threat
• Insult
• Identity Hate

The dataset is available on Kaggle and includes both training and test sets.

Methodology

The project follows these steps:

1. Data Exploration:

   • Checking dataset structure
   • Visualizing word frequency and toxicity distribution

2. Preprocessing:

   • Text cleaning (removing punctuation, stopwords, and special characters)
   • Tokenization and stemming/lemmatization
   • Converting text to numerical representations using TF-IDF or embeddings

3. Modeling:

   • Training machine learning models such as Logistic Regression, Naive Bayes, or deep learning models like LSTMs and Transformers (e.g., BERT)
   • Hyperparameter tuning and cross-validation

4. Evaluation:

   • Performance assessment using metrics like AUC-ROC, accuracy, precision, recall, and F1-score
   • Confusion matrices and classification reports

5. Prediction & Submission:

   • Making predictions on the test dataset
   • Formatting and submitting results to Kaggle

Requirements

To run this project, install the following dependencies:

pip install numpy pandas matplotlib seaborn scikit-learn tensorflow torch transformers

Usage

To execute the notebook:

1. Download the dataset from Kaggle.
2. Run the notebook step by step.
3. Modify hyperparameters and try different models for better performance.

Results & Findings

• Toxicity classes are highly imbalanced, most comments are not toxic, and some categories like threat and identity hate have very few positive examples.
• There are comments with misspellings, non-standard grammar, and slang that make preprocessing more crucial
• Single comment can belong to several categories.

Future Improvements

• Experiment with different architectures such as Transformer-based models.
• Improve text preprocessing techniques.
• Perform data augmentation to balance the dataset.

Acknowledgments

Thanks to Kaggle and the Jigsaw team for providing this challenge and dataset. This project was developed using open-source libraries like Scikit-learn, TensorFlow, and PyTorch.

Author

[Adilet Akimshe]