# Summ^N: A Multi-Stage Summarization Framework for Long Input Dialogues and Documents

**Authors:** Yusen Zhang, Ansong Ni, Ziming Mao, Chen Henry Wu, Chenguang Zhu, Budhaditya Deb, Ahmed H. Awadallah, Dragomir Radev, Rui Zhang  
**arXiv:** [2110.10150](https://arxiv.org/abs/2110.10150)  
**Date:** 2021-10-16  
**Summary page:** https://www.summarizepaper.com/en/paper/2110.10150v2/

## Key points

- Summ$^N$ is a multi-stage framework designed to improve summarization of long texts
- It splits input texts into multiple stages for generating coarse and fine-grained summaries
- Can handle single-source documents and dialogues using different abstractive summarization models
- Outperforms previous methods in terms of ROUGE scores on various datasets
- Data and code are available on GitHub for transparency and reproducibility
- Authors highlight challenges in longer dialogue and document summarization tasks due to limitations of training large pretrained language models

## Summary

, , , , 

The authors present Summ$^N$, a multi-stage framework designed to overcome the limitations of existing pretrained language models in processing long texts for summarization tasks. This approach splits input texts into multiple stages, allowing for a coarse summary to be generated before producing a final fine-grained summary. Summ$^N$ is versatile and can handle both single-source documents and dialogues, utilizing different abstractive summarization models. It outperforms previous methods in terms of ROUGE scores on various datasets, including meeting summaries from AMI, ICSI, and QMSum, TV series datasets from SummScreen, and a document summarization dataset GovReport. The authors provide transparency and reproducibility by making their data and code available on GitHub. They also discuss the challenges posed by longer dialogue and document summarization tasks and emphasize the need for innovative solutions due to the complexity of training large pretrained language models with limited input lengths.