Turkish News Classification Project

This project provides two different models for classifying Turkish news articles into various categories such as sports, politics, economy, and more. It includes a classic machine learning baseline and a modern deep learning approach using a Transformer-based model.

Models

This repository contains two distinct models, each in its own directory.

1. v1_baseline: TF-IDF + Logistic Regression

This model serves as a simple and fast baseline for the text classification task.

• Architecture: It uses a TF-IDF (Term Frequency-Inverse Document Frequency) vectorizer to convert text into numerical features and a Logistic Regression classifier to perform the classification.
• Technology: Built with scikit-learn, pandas, and nltk.
• Performance: This model provides a solid starting point. It performs very well for distinct categories like "spor" (sports) but struggles with more nuanced or overlapping categories.
• Details: You can find the training script, saved model, and dependencies in the v1_baseline/ directory.

2. v2_bert: Fine-Tuned BERT Model

This is a more advanced model that leverages a pre-trained BERT model for higher accuracy and better semantic understanding.

• Architecture: It uses the dbmdz/bert-base-turkish-cased model, a BERT model pre-trained on a large corpus of Turkish text. The model is then fine-tuned on the specific news classification dataset.
• Technology: Built with Hugging Face Transformers, PyTorch, and datasets.
• Performance: This model is expected to significantly outperform the baseline, as it can understand the context and semantics of the news text more effectively.
• Details: The v2_bert/ directory contains everything needed to train the model, run predictions, and manage dependencies.

Project Structure

.
├── v1_baseline/
│   ├── train_baseline.py   # Script to train the TF-IDF + Logistic Regression model
│   ├── model/              # Saved baseline model
│   └── requirements.txt    # Dependencies for the baseline model
│
└── v2_bert/
    ├── train_bert_finetune.py  # Script to fine-tune the BERT model
    ├── predict_news.py         # Script to classify news with the BERT model
    ├── model/                  # Saved fine-tuned BERT model
    └── requirements.txt        # Dependencies for the BERT model

How to Use

First, clone the repository to your local machine.

Running the Baseline Model (v1)

1. Navigate to the baseline directory:

   cd v1_baseline

2. Install dependencies:

   pip install -r requirements.txt

3. Run the training script:

   python train_baseline.py

   This will train the model and save the baseline_model.pkl and vectorizer.pkl inside the v1_baseline/model/ directory.

Running the BERT Model (v2)

1. Navigate to the BERT directory:

   cd v2_bert

2. Install dependencies:

   pip install -r requirements.txt

3. Run the training script:

   python train_bert_finetune.py

   This will fine-tune the BERT model and save the best version to the v2_bert/model/fine_tuned_bert/ directory.

4. Make predictions: To classify new text, run the interactive prediction script:

   python predict_news.py