Google Extension for Effortless Text Summarization

Description

The Text Summarization Google extension simplifies online reading by condensing articles into concise summaries. With an intuitive interface featuring a text area and 'Summary' button, it efficiently processes news articles, providing quick access to key information. The extension works by reading the current website's URL, extracting content via a server, and summarizing it using the PEGASUS_X model. In my work, I implemented abstractive text summarization techniques based on two papers: PEGASUS: Pre-training with Extracted Gap-sentences for Abstractive Summarization and Investigating Efficiently Extending Transformers for Long Input Summarization. Specifically, I will utilize the pre-training objectives (GSG) proposed by PEGASUS for the pre-training phase, and I will incorporate the improvements in network architectures and training strategies for long text summarization.

Prerequisite

• CUDA/CUDNN
• Python3
• Packages found in requirements.txt

Getting started

• git clone https://gitlab.com/DuongKien2001/summary_page
• cd summary_page
• pip install -r requirements.txt

Run training and testing

Prepare dataset for training model with task GSG

python model/prepare_data.py

Example of pre-training a model with C4 dataset on a single gpu

python model/train.py --src_len 512 --tgt_len 256 --epochs 6

Example of fine-tuning a model with PubMed dataset

python model/train.py --src_len 6400 --tgt_len 256 --epochs 3 --no-pretrain -r checkpoint.pth

Evaluate the model with the PubMed test set

python model/evaluatePegasusX.py --start_idx 0 --end_idx 7000

Start backend server

python backend/main.py

Pretrained models

Pretrained model for PubMed can be downloaded at: (Link), and should be unzipped in the 'model/checkpoint/finetune' folder.

Usage

1. Go to a webpage that has an article you want to shorten.
2. Click on the extension icon or use the keyboard shortcut to turn on the extension.
3. On the extension, click the 'Summary' button.
4. Read the brief summary.