News Text Classification

A multi-class text classification project that categorizes news articles into World, Sports, Business, and Sci/Tech using the AG News dataset. Three models were built and compared: Logistic Regression, CNN, and DistilBERT.

Dataset

  • Source: AG News (fancyzhx/ag_news on Hugging Face)
  • Training set: 120,000 articles (30,000 per class)
  • Test set: 7,600 articles (1,900 per class)
  • Classes: 0 = World, 1 = Sports, 2 = Business, 3 = Sci/Tech

Models

Model Approach
Logistic Regression TF-IDF vectorization (10,000 features)
CNN Embedding + Conv1D + GlobalMaxPooling
DistilBERT Fine-tuned distilbert-base-uncased, trained on a balanced 30,000-row subset

Results

Model Accuracy
Logistic Regression 91.25%
CNN 91.14%
DistilBERT 93.30%

Model Comparison

DistilBERT performed best, benefiting from contextual language understanding that TF-IDF and basic embeddings don't capture.

Project Structure

  • lr_model.pkl # Trained Logistic Regression model
  • tfidf_vectorizer.pkl # TF-IDF vectorizer
  • cnn_model.keras # Trained CNN model
  • tokenizer.pkl # Keras tokenizer for CNN
  • distilbert_model/ # Trained DistilBERT model + tokenizer
  • plots/ # All generated plots (confusion matrices, ROC curves, training curves)
  • REPORT.md # Full project report

Requirements

  • datasets
  • scikit-learn
  • tensorflow
  • transformers
  • accelerate
  • matplotlib, seaborn

Usage

Load any saved model with its corresponding preprocessing tool (vectorizer/tokenizer) to run predictions on new text. See REPORT.md for full details, including confusion matrices and per-class performance.

Built With

Share this project:

Updates