[Dataset] XSum

1. Introduction

• The XSum dataset consists of 226,711 Wayback archived BBC articles ranging over almost a decade (2010 to 2017) and covering a wide variety of domains (e.g., News, Politics, Sports, Weather, Business, Technology, Science, Health, Family, Education, Entertainment and Arts).

• paperswithcode: https://paperswithcode.com/dataset/xsum
• Github: https://github.com/EdinburghNLP/XSum (Model 및 Dataset 전체 Repo.)
• Download: https://github.com/EdinburghNLP/XSum/tree/master/XSum-Dataset
• Paper: https://arxiv.org/pdf/1808.08745.pdf

2. 특징

• 언어: English

• XSum = Extreme Summarization Dataset

  • Single-document 에 대한 abstractive summarization 데이터셋
  • 하나의 document 가 존재하고, “What is the article about?” 이라는 물음에 대해 짧은 one-sentence summarization 하는 것이 목적
  • 또다른 대표 benchmark set인 CNN DailyMail 에 비해 gold summary 가 더욱 abstractive 함. 즉, extractive model 로 테스트 하면 성능이 낮게 나옴.

• Datasets 크기

  • train: 203,577
  • validation: 11,305
  • test: 11,301

• Datasets 평균 length

  • Document
    • word: 431.07
    • sentence: 19.77
  • Summary
    • word: 23.26
    • sentence: 1.00 (fixed)

• Vocab size

  • Document: 399,147
  • Summary: 81,092