# NeuS: Neutral Multi-News Summarization for Mitigating Framing Bias

**Authors:** Nayeon Lee, Yejin Bang, Tiezheng Yu, Andrea Madotto, Pascale Fung  
**arXiv:** [2204.04902](https://arxiv.org/abs/2204.04902)  
**Date:** 2022-04-11  
**Summary page:** https://www.summarizepaper.com/en/paper/2204.04902v1/

## Key points

- Media framing bias can significantly impact political polarization
- A new task of generating neutral summaries from multiple news headlines has been proposed to address this issue and promote balanced news consumption
- The paper presents a new dataset, insights about framing bias through a case study, and effective metrics and models for the task
- Multi-document summarization (MDS) is identified as a challenging subtask that aims to condense multiple documents into concise yet informative summaries
- Deep neural models are increasingly applied to MDS tasks for improved performance
- The Neutral multi-news Summarization (NEUS) task is introduced to generate neutral headline summaries from news headlines with varying degrees of political bias
- The ALLSIDES dataset provides triplets of news reports from left, right, and center American publishers along with expert-written neutral summaries focusing on U.S. political topics prone to media bias
- Automating roundup generation using MDS models capable of providing a neutral understanding of issues can help mitigate framing biases in media sources
- This research contributes to advancing the field by exploring how framing bias manifests in news headlines and providing insights into mitigating biases through automated summarization techniques

## Summary

Media framing bias can have a significant impact on political polarization. To address this issue and promote balanced news consumption, a new task of generating neutral summaries from multiple news headlines of varying political spectrums has been proposed. This paper presents a new dataset, insights about framing bias through a case study, and effective metrics and models for the task. Experimental analyses provide insights into remaining challenges and future directions. Multi-document summarization (MDS) is identified as a challenging subtask of automatic text summarization that aims to condense multiple documents into a concise yet informative summary. Deep neural models have been increasingly applied to the MDS task, utilizing large-scale datasets and pre-trained language models for improved performance. While various subtopics of MDS have been explored in previous works, there is limited research on generating framing bias-free summaries from multiple news articles. To fill this gap, the Neutral multi-news Summarization (NEUS) task is introduced in this work. Its objective is to generate neutral headline summaries from news headlines with varying degrees and orientations of political bias. The ALLSIDES dataset from Allsides.com provides triplets of news reports from left, right, and center American publishers along with expert-written neutral summaries. The dataset primarily focuses on U.S. political topics prone to media bias such as elections, the White House, and politics. By crawling headline triplets and neutral headline summaries from Allsides.com, the source inputs {HL, HR, HC} are used to generate the target output Hneu for the NEUS task. It's worth noting that even "center" media sources may exhibit framing bias due to editorial judgment biases. Therefore, automating roundup generation using MDS models capable of providing a neutral understanding of issues can help mitigate this gap. Overall, this research contributes to advancing the field by exploring how framing bias manifests in news headlines and providing insights into mitigating biases through automated summarization techniques. By addressing the challenge of generating unbiased summaries from politically diverse sources, this work paves the way for more balanced and informed news consumption practices in an era marked by increasing media polarization.