# Generating Wikipedia by Summarizing Long Sequences

**Authors:** Peter J. Liu, Mohammad Saleh, Etienne Pot, Ben Goodrich, Ryan Sepassi, Lukasz Kaiser, Noam Shazeer  
**arXiv:** [1801.10198](https://arxiv.org/abs/1801.10198)  
**Date:** 2018-01-30  
**Summary page:** https://www.summarizepaper.com/en/paper/1801.10198v1/

## Key points

- Authors propose a method for generating English Wikipedia articles
- Approach treats it as a multi-document summarization task
- Extractive summarization used to identify important information from source documents
- Neural abstractive model used to generate the article
- Introduce decoder-only architecture for abstractive model to handle long sequences effectively
- Model capable of generating coherent and fluent multi-sentence paragraphs and complete Wikipedia articles
- Performance evaluated using perplexity scores, ROUGE scores, and human evaluations
- Method demonstrates ability to extract relevant factual information
- Decoder-only architecture allows for efficient processing of long sequences resulting in high-quality content

## Summary

In the paper titled "Generating Wikipedia by Summarizing Long Sequences," authors Peter J. Liu, Mohammad Saleh, Etienne Pot, Ben Goodrich, Ryan Sepassi, Lukasz Kaiser and Noam Shazeer propose a method for generating English Wikipedia articles by treating it as a multi-document summarization task. They employ extractive summarization to identify important information from source documents and use a neural abstractive model to generate the article. The authors introduce a decoder-only architecture for the abstractive model that can effectively handle very long sequences and surpasses the limitations of typical encoder-decoder architectures used in sequence transduction. This model is capable of generating coherent and fluent multi-sentence paragraphs and even complete Wikipedia articles. To evaluate the performance of their approach, the authors conduct experiments using reference documents. They demonstrate that their model can extract relevant factual information based on perplexity scores, ROUGE scores (a metric for evaluating text summarization) and human evaluations. Overall, this paper presents an innovative method for generating Wikipedia articles by summarizing multiple source documents. The proposed decoder-only architecture allows for efficient processing of long sequences resulting in high quality generated content.