MediaSum

MediaSum，一个大规模的媒体采访数据集，由463.6k成绩单和抽象摘要组成。为了创建这个数据集，我们从NPR和CNN收集采访记录，并使用概述和主题描述作为摘要。与现有的用于对话摘要的公共语料库相比，我们的数据集要大一个数量级，并且包含来自多个领域的复杂多方对话。我们进行统计分析，以证明电视和广播采访的笔录中表现出的独特位置偏差。我们还表明，MediaSum可用于迁移学习，以提高模型在其他对话摘要任务上的性能。

MediaSum, a large-scale media interview dataset, comprises 463.6k transcripts and abstractive summaries. To construct this dataset, we collected interview records from NPR and CNN, and used their outlines and topic descriptions as the corresponding summaries. Compared with existing public corpora for conversational summarization, our dataset is an order of magnitude larger and contains complex multi-party dialogues across multiple domains. We conducted statistical analyses to demonstrate the unique positional bias exhibited in transcripts of television and radio interviews. We further show that MediaSum can be utilized for transfer learning to improve model performance on other conversational summarization tasks.

搜集汇总

背景与挑战

背景概述

MediaSum是一个大规模媒体采访数据集，包含46.3万条来自NPR和CNN采访的转录文本和摘要，其规模远超现有对话摘要语料库，并展现独特的位置偏差。该数据集可用于迁移学习，以提升其他对话摘要任务的模型性能。