Abstract Summarizer

This repository contains various components for abstractive text summarization using the T5 (Text-to-Text Transfer Transformer) model. The project explores both using a pre-trained base T5 model and fine-tuning T5 for improved summarization, along with a demonstration application.

Repository Structure

The repository is organized into two main directories: no-tuning and tuning, each serving a distinct purpose in the summarization workflow, plus a main application in the root.

app.py (Root Directory)

This app.py in the root directory serves as the primary demonstration of the project's capabilities. It is a Streamlit application designed to showcase summarization using a pre-trained, fine-tuned T5 model hosted on Hugging Face (admin-sauce/t5-summarizer). This is intended to be the user-facing application for quick and efficient summarization.

no-tuning/

This directory focuses on demonstrating text summarization using a generic, pre-trained t5-base model without any custom fine-tuning. It provides a foundational understanding of how T5 can be used for summarization out-of-the-box.

• no-tuning/app.py: A Flask web application that provides a simple interface for abstractive summarization. Users can input raw text, and the application will generate a summary using the t5-base model. It also includes logic to handle and chunk larger input texts for processing.
• no-tuning/main.ipynb: A Jupyter Notebook illustrating the process of extracting text from various document formats (PDF, DOCX) and then summarizing the extracted content using the t5-base model. This notebook provides the experimental code for document processing, which can be integrated into summarization workflows.
• no-tuning/static/ & no-tuning/templates/: These subdirectories contain the static assets (CSS) and HTML templates for the Flask web application, defining its user interface and styling.

tuning/

This directory is dedicated to the aspects of fine-tuning and evaluating T5 models for specific summarization tasks. It contains scripts and resources related to training and benchmarking custom summarization models.

• tuning/evaluate.py: A Python script designed for evaluating fine-tuned summarization models. It typically uses benchmark datasets (e.g., big_patent) and metrics like ROUGE to assess the performance of a model, such as KipperDev/t5_summarizer_model. This script helps in understanding the effectiveness of fine-tuning efforts.
• tuning/t5/: (Hypothesized) This subdirectory is expected to contain the core scripts or Jupyter Notebooks used for the actual fine-tuning process of the T5 model. This is where the base T5 model would be adapted using specific datasets to create specialized summarization models.

Models Used

Throughout this project, the following T5 models are utilized:

• t5-base: The generic, pre-trained T5 model used for baseline summarization in the no-tuning components.
• admin-sauce/t5-summarizer: A fine-tuned T5 model, likely developed within this project or an external resource, used in the main Streamlit app.py for high-performance summarization.
• KipperDev/t5_summarizer_model: Another fine-tuned T5 model, evaluated by the tuning/evaluate.py script, showcasing different fine-tuning outcomes or external model comparisons.

Getting Started

To run the different applications:

• Streamlit Demo:

  streamlit run app.py

• Flask Web App (no-tuning):

  python no-tuning/app.py

  (Then navigate to http://127.0.0.1:5000 in your browser)

Further instructions for setting up environments and dependencies would be provided in individual component readmes or a requirements.txt file.