Automated Content Moderation Using Machine Learning

Overview

In the digital age, managing online discourse is a growing challenge due to the prevalence of toxic and harmful content. This project focuses on developing a machine learning pipeline to detect and classify toxic comments with high accuracy. We implemented multiple models, from traditional classifiers to deep learning architectures, to identify toxicity across different categories.

---

Dataset

The dataset consists of Wikipedia comments labeled by human raters for toxicity types:

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

---

Data Preprocessing

• Stopword Removal: Eliminated unnecessary words using NLTK.
• Text Cleaning: Removed URLs, special characters, and extra whitespaces via regex.
• Feature Extraction: Utilized TF-IDF vectorization and GPT-based embedding techniques.
• Class Balancing: Addressed imbalance using LeastSampledClassSampler to ensure fair model training.

---

Models Implemented

1. Traditional Machine Learning Classifiers

Naïve Bayes Classifier

• Implemented Multinomial Naïve Bayes with OneVsRestClassifier.
• Performance: Struggled with minority classes, resulting in low F1 scores.

Logistic Regression

• Used TF-IDF features with OneVsRestClassifier.
• Performance: Improved over Naïve Bayes, yet failed to capture complexity in severe toxicity categories.

Linear Support Vector Classifier (SVC)

• Performance: Performed significantly better than previous models.
• Best F1 scores in detecting general toxicity.

2. Deep Learning Models

Custom Neural Network

• Architecture: Three linear layers with ReLU activations.
• Loss Function: BCEWithLogitsLoss for multi-label classification.
• Training: Used AdamW optimizer with a learning rate of 0.001.
• Performance: Much better than traditional models but struggled with minority classes.

Pre-trained Transformers (BERT-based Model)

• Architecture: Fine-tuned BERT model using BertForMultiLabelSequenceClassification.
• Tokenizer: BertTokenizer from Hugging Face.
• Training: 5 epochs, AdamW optimizer, learning rate 1e-5.
• Performance:
  • Best results overall, outperforming all previous models.
  • Struggled with identity hate and severe toxicity, similar to the neural network.

---

Performance Comparison

Macro-Averaged F1 Scores (Test Data)

Model	Toxic	Severe Toxic	Obscene	Threat	Insult	Identity Hate
Naïve Bayes	0.28	0.00	0.19	0.00	0.07	0.00
Logistic Regression	0.73	0.35	0.73	0.24	0.63	0.24
Linear SVC	0.77	0.37	0.78	0.36	0.68	0.37
Neural Network	0.85	0.66	0.81	0.55	0.80	0.73
BERT Transformer	0.96	0.41	0.83	0.58	0.78	0.73

---

Results and Insights

• Traditional classifiers were not effective for minority classes.
• Neural networks and BERT significantly improved performance, particularly for common toxic categories.
• BERT performed the best, but still struggles with highly imbalanced data.

---

Future Work

• Fine-tune hyperparameters for transformers.
• Explore ensemble models combining deep learning with traditional classifiers.
• Improve dataset balance using advanced oversampling techniques.
• Test alternative architectures like RoBERTa or DistilBERT for better efficiency.

---

Installation & Usage

Requirements

• Python 3.x
• Jupyter Notebook / Google Colab
• Libraries: pandas, numpy, nltk, scikit-learn, torch, transformers

Setup

pip install -r requirements.txt

Run the Project

jupyter notebook

Open Each Notebook and Run it. To obtain the original dataset for this project, you can email at [email protected]

---