BUMP

BUMP是一个由数据挖掘公司创建的数据集，包含889对人类编写的极小差异摘要对。该数据集通过在CNN/DailyMail数据集的摘要中引入单一错误来生成不忠实摘要，旨在评估自动忠实度度量的性能。BUMP数据集特别适用于测试忠实度度量的一致性、对人类编写文本的有效性以及对不同错误类型的敏感性，为解决摘要生成中的忠实度问题提供了重要工具。

BUMP is a dataset developed by a data mining company, consisting of 889 pairs of human-written summaries with minimal pairwise differences. This dataset generates unfaithful summaries by introducing a single error into the summaries from the CNN/DailyMail dataset, aiming to evaluate the performance of automatic faithfulness metrics. The BUMP dataset is particularly suitable for testing the consistency of faithfulness metrics, their effectiveness on human-written text, and their sensitivity to different error types, providing an important tool for addressing faithfulness issues in summarization.