# PEGASUS: Pre-training with Extracted Gap-sentences for Abstractive Summarization

**Authors:** Jingqing Zhang, Yao Zhao, Mohammad Saleh, Peter J. Liu  
**arXiv:** [1912.08777](https://arxiv.org/abs/1912.08777)  
**Date:** 2019-12-18  
**Summary page:** https://www.summarizepaper.com/en/paper/1912.08777v3/

## Key points

- The paper proposes a new self-supervised objective for pre-training large Transformer-based encoder-decoder models on massive text corpora.
- The objective is tailored for abstractive text summarization and involves removing/masking important sentences from an input document and generating them together as one output sequence from the remaining sentences, similar to an extractive summary.
- The authors evaluate their best PEGASUS model on 12 downstream summarization tasks spanning news, science, stories, instructions, emails, patents and legislative bills.
- The PEGASUS model achieves state-of-the-art performance on all 12 downstream datasets measured by ROUGE scores.
- The model shows surprising performance on low-resource summarization by surpassing previous state-of-the art results on six datasets with only 1000 examples.
- The authors validate their results using human evaluation and show that their model summaries achieve human performance on multiple datasets.
- Their work addresses the lack of systematic evaluation across diverse domains in previous pre-training objectives for text summarization and provides a promising approach for improving abstractive text summarization.

## Summary

In their paper titled "PEGASUS: Pre-training with Extracted Gap-sentences for Abstractive Summarization," authors Jingqing Zhang, Yao Zhao, Mohammad Saleh, and Peter J. Liu propose a new self-supervised objective for pre-training large Transformer-based encoder-decoder models on massive text corpora. This objective is tailored for abstractive text summarization and involves removing/masking important sentences from an input document and generating them together as one output sequence from the remaining sentences, similar to an extractive summary. The authors evaluate their best PEGASUS model on 12 downstream summarization tasks spanning news, science, stories, instructions, emails, patents and legislative bills. They demonstrate that it achieves state-of-the-art performance on all 12 downstream datasets measured by ROUGE scores. Additionally, the model shows surprising performance on low-resource summarization by surpassing previous state-of-the art results on six datasets with only 1000 examples. The authors validate their results using human evaluation and show that their model summaries achieve human performance on multiple datasets. Their work addresses the lack of systematic evaluation across diverse domains in previous pre-training objectives for text summarization and provides a promising approach for improving abstractive text summarization.