News Dataset Text Preprocessing Experiment

This project investigates the impact of different text preprocessing techniques on the fine-tuning of a RoBERTa model for a news dataset. We conduct an experiment to compare two preprocessing approaches and analyze their effectiveness.

Experiment Setup

• Original Dataset: News Category Dataset
• Sub-dataset Creation: 18,981 items were randomly selected from the original dataset in a stratified manner to ensure proportional representation of different news categories.
• Preprocessing Techniques:
  • Dataset 1: Removal of numbers, stop words, and punctuation.
  • Dataset 2: Removal of numbers, stop words, and punctuation, plus stop words removing.
• Models:
  • Two Roberta models were fine-tuned, each on one of the preprocessed datasets.
  • A final model was trained based on results of the 2 model.

Project Structure

• experiments/
  • Contains 4 Jupyter notebooks detailing the experimental process:
    • data-preprocessing-nostopwords-vs-stopwords.ipynb: Creation of the two experimental datasets, one with and one without stop words.
    • roberta-fine-tuning-stopwords.ipynb: Fine-tuning of the Roberta model on the dataset with stop words.
    • roberta-fine-tuning-nostopwords.ipynb: Fine-tuning of the Roberta model on the dataset without stop words.
    • dataset-preprocessing.ipynb: Processes the entire dataset based on the insights gained from the fine-tuning experiments.
• news-classification-big-dataset.ipynb: Trains and evaluates the final model on the full dataset using the best preprocessing technique identified in the experiments.
• README.md

Experimental Results and Analysis

• Compare the performance of the two fine-tuned models.
• Discuss which preprocessing technique proved more effective.
• Explain the rationale behind the final model training.