BillSum

BillSum 是第一个汇总美国国会和加利福尼亚州法案的数据集。 BillSum 数据集由三部分组成：美国培训账单、美国测试账单和加州测试账单。美国账单是从美国政府出版局 (GPO) 提供的 Govinfo 服务收集的。该语料库由国会第 103-115 届（1993-2018 年）会议的法案组成。数据分为 18,949 个火车账单和 3,269 个测试账单。对于加利福尼亚州，2015-2016 年会议的法案是直接从立法机构的网站上抓取的；这些摘要是由他们的立法顾问撰写的。 BillSum 语料库侧重于长度为 5,000 到 20,000 个字符的中型立法。作者选择以字符而不是单词或句子来测量文本长度，因为文本结构复杂，难以一致地测量单词。选择该范围是因为一方面，短期票据引入了微小的变化并且不需要摘要。虽然 CRS 为他们生成摘要，但它们通常包含法案的大部分文本。另一方面，很长的立法通常由几个大的部分组成。

BillSum is the first dataset focused on summarizing U.S. congressional and California state legislative bills. The BillSum dataset is composed of three parts: U.S. training bills, U.S. test bills, and California test bills. U.S. bills were collected from the Govinfo service provided by the U.S. Government Publishing Office (GPO). This corpus consists of bills from the 103rd to 115th sessions of the U.S. Congress (1993–2018), with the data split into 18,949 training bills and 3,269 test bills. For California, bills from the 2015–2016 legislative sessions were directly scraped from the legislature’s official website, and their summaries were written by the state’s legislative advisors. The BillSum corpus targets medium-sized legislative texts with lengths ranging from 5,000 to 20,000 characters. The authors opted to measure text length by character count instead of word or sentence count, as the complex structure of legislative texts makes consistent word count measurement challenging. This length range was chosen due to two considerations: on one hand, short bills exhibit minimal variability and do not necessitate summarization—while the Congressional Research Service (CRS) generates summaries for these short bills, such summaries often include most of the bill’s full text. On the other hand, extremely long legislative texts typically comprise multiple large sections.

搜集汇总

背景与挑战

背景概述

BillSum是首个专门汇总美国国会及加州法案的数据集，包含来自Govinfo的美国法案和加州立法机构网站的法案，总计约2.2万条数据。该数据集聚焦于长度为5000至20000字符的中型立法文本，适用于自动摘要任务。