Abstractive text summarisation of GitHub issues

Overview

A dataset of GitHub Issues' titles, bodies and URLs has been used to create a Sequence to Sequence model with GRUs to summarize the GitHub issue body. The machine generated title is a more compact yet accurate representation.

The project includes:

• RNNs to create a sequence to sequence model for abstractive text summarisation.
• Teacher forcing algorithm is used to train the decoder model.
• A recommender that suggests GitHub issues with similar titles. The Spotify ANNOY package is used for this purpose.
• The model's performance determined through it's BLEU score.

Architecture

Dataset

• The dataset used has over 8M entries and hence the model requires sufficient training time.
• You can find the dataset here.