Skip to content

Latest commit

 

History

History
160 lines (120 loc) · 6.21 KB

README.md

File metadata and controls

160 lines (120 loc) · 6.21 KB

Research Paper Classification System

A comprehensive machine learning pipeline for classifying research papers into scientific disciplines using a fine-tuned BERT model. This system includes web scraping capabilities, data preprocessing, and a high-accuracy classification model.

Target Scientific Disciplines

The system classifies research papers into 9 major scientific categories with detailed subfields:

Category Subfields
Computer Science Artificial Intelligence, Cloud Computing, Cyber Security, Quantum Computing, Software Engineering
Medicine Cardiology, Surgery, Neurology, Oncology, Pediatrics, Psychiatry
Business Marketing, Finance, Management, Entrepreneurship, E-commerce
Biology Genetics, Ecology, Zoology, Biochemistry, Physiology
Physics Quantum Mechanics, Astrophysics, Particle Physics, Cosmology
Chemistry Organic Chemistry, Analytical Chemistry, Biochemistry, Medicinal Chemistry
Mathematics Algebra, Calculus, Statistics, Probability, Optimization
Psychology Clinical Psychology, Cognitive Psychology, Social Psychology, Neuropsychology
Environmental Science Climate Change, Conservation, Sustainability, Renewable Energy

Dataset Statistics

Data Collection & Cleaning

  • Original Records: 155,882 research papers
  • Clean Non-duplicate Records: 140,004 papers
  • Cleaning Ratio: 89.81%

Final Dataset Distribution

  • Psychology: 16,821 records (12.0%)
  • Chemistry: 16,675 records (11.9%)
  • Physics: 15,941 records (11.4%)
  • Business: 15,929 records (11.4%)
  • Mathematics: 15,464 records (11.0%)
  • Medicine: 15,361 records (11.0%)
  • Computer Science: 14,776 records (10.6%)
  • Biology: 14,729 records (10.5%)
  • Environmental Science: 14,308 records (10.2%)

Quick Start

Prerequisites

  • Google Colab (recommended) or Jupyter Notebook
  • Google Drive account for data storage

Installation & Setup

  1. Clone this repository to your Colab environment:
!git clone https://github.com/Emran025/Research_Paper_Classification_model.git
%cd Research_Paper_Classification_model
  1. Download all project assets using our downloader notebook:

Open and run gdrive_folder_downloader.ipynb to download:

  • Pre-trained BERT model (bert_text_classifier)
  • Training checkpoints and logs
  • Cleaned dataset (cleaned_dataset.csv)
  • Raw scraped research papers
  1. You're ready to go! All project components are now available locally.

Project Structure

Research_Paper_Classification_model/
├── gdrive_folder_downloader.ipynb      # Download project assets from Google Drive
├── mendeley_scraper.ipynb              # Scrape research papers from Mendeley
├── research_papers_bert_finetuning.ipynb # BERT model training pipeline
├── bert_text_classifier/               # Pre-trained model (downloaded)
├── bert_classification_output/         # Training artifacts (downloaded)
├── cleaned_dataset.csv                 # Processed dataset (downloaded)
└── Mendeley_Research/                  # Raw scraped data (downloaded)

Usage Options

Option 1: Use Pre-trained Model (Recommended)

After running the downloader notebook, you'll have immediate access to:

  • Fine-tuned BERT model (95.39% accuracy)
  • Pre-processed dataset (140,004 research papers)
  • Training logs and checkpoints

Option 2: Full Pipeline from Scratch

For complete reproducibility:

  1. Data Collection: Run mendeley_scraper.ipynb to scrape fresh data from Mendeley
  2. Model Training: Execute research_papers_bert_finetuning.ipynb to fine-tune BERT
  3. Evaluation: Use the trained model for inference and analysis

Model Performance

The fine-tuned BERT model achieves state-of-the-art performance:

  • Evaluation Loss: 0.184
  • Overall Accuracy: 95.39%
  • Evaluation Runtime: 428.03 seconds
  • Samples Processed: 65.306 samples/second

Detailed Classification Report

Category Precision Recall F1-Score Support
Biology 0.94 0.93 0.94 3,177
Business 0.96 0.97 0.97 3,179
Chemistry 0.94 0.96 0.95 3,073
Computer Science 0.96 0.93 0.95 2,987
Environmental Science 0.95 0.94 0.95 2,850
Mathematics 0.93 0.96 0.95 3,091
Medicine 0.97 0.96 0.96 3,067
Physics 0.97 0.95 0.96 3,181
Psychology 0.97 0.97 0.97 3,348

Model Integration

Hugging Face Hub

The model is available on Hugging Face Hub for easy integration:

from transformers import AutoTokenizer, AutoModelForSequenceClassification

tokenizer = AutoTokenizer.from_pretrained("Emran025/bert_text_classifier")
model = AutoModelForSequenceClassification.from_pretrained("Emran025/bert_text_classifier")

Model Card: https://huggingface.co/Emran025/bert_text_classifier

Google Drive Resources

All datasets and model artifacts are available via Google Drive:

  • Main Drive Folder: Google Drive Link
  • Raw datasets organized by scientific discipline
  • Pre-trained model and training checkpoints
  • Cleaned and processed datasets

Data Sources

Data was collected from Mendeley research catalog across 9 major disciplines:

  • Computer Science, Medicine, Business, Chemistry, Mathematics
  • Psychology, Environmental Science, Biology, Physics

Each category contains multiple specialized subfields for comprehensive coverage.

Technical Specifications

  • Framework: PyTorch with Transformers
  • Base Model: BERT (bert-base-uncased)
  • Training Environment: Google Colab
  • Dataset Size: 140,004 research papers (after cleaning)
  • Classes: 9 scientific disciplines with 5-12 subfields each
  • Training Samples: 27,953 papers (evaluation set)

Note

This project is optimized for Google Colab. All file paths and dependencies are configured for Colab's environment. When running locally, adjust the file paths accordingly.

Contributors: Emran Nasser, Mohammed Alyafrosy, Ryadh Alizi
License: MIT
Last Updated: October 2024