ZeroSCROLLS
收藏资源简介:
ZeroSCROLLS是一个针对长文本自然语言理解的零样本基准,包含10个测试任务,每个任务都需要对不同类型的长文本进行推理。该数据集由特拉维夫大学布劳恩宁计算机科学学院创建,旨在评估模型在无监督学习情况下的表现。数据集内容丰富,包括政府报告、电视剧本、会议记录等多种类型的长文本数据。创建过程中,研究人员精心设计了任务和评估指标,确保数据集的实用性和准确性。ZeroSCROLLS主要用于研究长文本理解,特别是在零样本学习场景下的应用,旨在推动模型在处理长序列文本时的性能提升。
ZeroSCROLLS is a zero-shot benchmark for long-text natural language understanding, comprising 10 test tasks each requiring reasoning over distinct types of long texts. This dataset was developed by the Braunning School of Computer Science, Tel Aviv University, with the objective of evaluating model performance under unsupervised learning scenarios. It features a rich collection of long-text data across multiple categories, including government reports, TV scripts, meeting transcripts, and more. During the development process, researchers meticulously designed the tasks and evaluation metrics to ensure the dataset's practicality and accuracy. Primarily utilized for research on long-text understanding, particularly in zero-shot learning applications, ZeroSCROLLS aims to promote the enhancement of model performance when handling long-sequence texts.
- 1ZeroSCROLLS: A Zero-Shot Benchmark for Long Text Understanding特拉维夫大学布劳恩宁计算机科学学院 · 2023年



