Toxic Comment Classification

Overview

This project tackles the Jigsaw Toxic Comment Classification Challenge, where the goal is to classify each comment into various categories indicating the level of toxicity. These categories include:

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

The dataset contains both toxic and non-toxic comments, and the task is to classify them into one or more of the aforementioned labels.

Results

Model and Architecture

The model is built using TensorFlow and Keras. It uses a deep learning architecture to perform text classification, employing a TextVectorization layer that converts raw text into numerical vectors using the TF-IDF (Term Frequency-Inverse Document Frequency) method. The architecture consists of the following layers:

1. Text Vectorization Layer:
   • Converts input text into numerical vectors using TF-IDF with bigrams.
2. Dense Layers:
   • First Dense layer: 256 units, ReLU activation, and L2 regularization.
   • Second Dense layer: 32 units, ReLU activation, and L2 regularization.
3. Output Layer:
   • Sigmoid activation for multi-label classification (6 classes).

The model is optimized using the Adam optimizer, and the loss function is binary cross-entropy. It is evaluated using categorical accuracy and AUC metrics.

Classification Results:

Label	Precision	Recall	F1-score	Support
toxic	0.92	1.00	0.96	31915
severe_toxic	0.99	0.00	0.00	31915
obscene	0.96	1.00	0.98	31915
threat	1.00	0.00	0.00	31915
insult	0.96	1.00	0.98	31915
identity_hate	0.99	0.00	0.01	31915

Notes on Classification Metrics:

• Precision: Measures the proportion of positive predictions that are actually correct.
• Recall: Measures the proportion of actual positives that are correctly identified.
• F1-score: A balanced measure of precision and recall, especially useful for imbalanced datasets.

Features

• Text Vectorization: The TF-IDF method with bigrams allows the model to capture important word pairs and their significance in the comment, improving the feature representation.
• Regularization: L2 regularization helps prevent overfitting by penalizing large weights.
• Multi-label Classification: Using sigmoid activation allows for predicting multiple labels per comment.

Sprints

Sprint 1: Data Preprocessing

• Deliverable: Preprocess text by tokenizing, converting to lowercase, and removing stop words.

Sprint 2: Model Architecture Design

• Deliverable: Build and configure the model with the TextVectorization layer, dense layers, and output layer.

Sprint 3: Training and Evaluation

• Deliverable: Train the model with binary cross-entropy loss and Adam optimizer. Evaluate using precision, recall, and F1-score metrics.

Conclusion

This model successfully classifies comments into multiple categories of toxicity, achieving high precision for most categories. However, the low recall for severe_toxic, threat, and identity_hate suggests the need for further model improvements, such as rebalancing the dataset or using more advanced techniques like class-weighting or ensemble methods.