GitSum guidelines

This repository contains the source code implementation of GitSum and the datasets used to replicate the experimental results of our paper:

Too long; didn’t read: Automatic summarization of GitHub README.MD with Transformers

Introduction

GitSum is a novel approach to the summarization of README.MD, automatically filling the blank “About” field for repositories. It is built on top of BART and T5, two cutting-edge deep learning techniques, learning from existing data to perform recommendations for repositories with a missing description. We test its performance using two datasets collected from GitHub. The evaluation shows that GitSum can generate highly relevant predictions, outperforming a well-established baseline.

Pre-requisites:

Installing neccessary packages:

pip install datasets transformers rouge-score nltk

Fine-tuning pre-trained models

Run the script in src/fine-tune-model.py to fine-tune pre-trained BART model with the default configuration in src/args_bart.json

Change reference variable json_file in the script to src/args_t5.json in order to fine-tune pre-trained T5 model for this task

Running the script src/fine-tune-model.py also perform evaluation on the test dataset. All output is saved in models directory as default. Every settings can be changed by manipulating the json configuration file, regarded as src/args_t5.json or src/args_bart.json