# What's New? Summarizing Contributions in Scientific Literature

**Authors:** Hiroaki Hayashi, Wojciech Kryściński, Bryan McCann, Nazneen Rajani, Caiming Xiong  
**arXiv:** [2011.03161](https://arxiv.org/abs/2011.03161)  
**Date:** 2020-11-06  
**Summary page:** https://www.summarizepaper.com/en/paper/2011.03161v2/

## Key points

- Authors propose a new task called disentangled paper summarization
- Task aims to generate separate summaries for contributions and context of scientific papers
- S2ORC corpus is extended with "contribution" and "context" reference labels
- Three baseline approaches for disentangled paper summarization are introduced
- Unified model controlled by input code prefixes
- Model with separate generation heads for disentangled outputs
- Training strategy incorporating additional supervision from citations
- Automatic evaluation protocol measures relevance, novelty, and disentanglement of generated summaries
- Human study involving expert annotators validates the approach
- New task considered more helpful than traditional scientific paper summarization methods in 79% of cases
- Approach shows promising results in helping researchers identify key findings quickly

## Summary

In their study titled "What's New? Summarizing Contributions in Scientific Literature," authors Hiroaki Hayashi, Wojciech Kryściński, Bryan McCann, Nazneen Rajani, and Caiming Xiong address the challenge of keeping up with the vast amount of academic articles being shared daily. They propose a new task called disentangled paper summarization which aims to generate separate summaries for the contributions and context of scientific papers. To support this task, the authors extend the S2ORC corpus by adding "contribution" and "context" reference labels across various domains. The authors introduce three baseline approaches for disentangled paper summarization: 1) a unified model controlled by input code prefixes; 2) a model with separate generation heads specialized in producing disentangled outputs; and 3) a training strategy that incorporates additional supervision from inbound and outbound citations. They also propose an automatic evaluation protocol that measures the relevance, novelty, and disentanglement of generated summaries. To validate their approach, the authors conduct a human study involving expert annotators. The results show that in 79% of cases their new task is considered more helpful than traditional scientific paper summarization methods. Overall, this study presents a novel approach to addressing the challenge of staying updated with scientific literature by generating separate summaries for contributions and context. The proposed methods show promising results in helping researchers quickly identify key findings in academic articles.