PALRACE

PALRACE是一个包含800篇阅读理解文章的数据集，由浙江大学创建，旨在提升机器阅读理解(MRC)任务的可解释性。该数据集的特点是每篇文章都附有人类标注的词级理由(rationale)，这些理由由至少26名读者标注，用于支持答案的选择。数据集中的文章来自RACE数据集，涵盖多种问题类型，如因果、事实、推理等。PALRACE的创建不仅支持了理由的监督学习，还为评估模型在定位和解释信息方面的能力提供了标准。此数据集的应用领域主要集中在提高MRC模型的解释性和性能，尤其是在模型未达到最先进性能时，通过人类理由的辅助，可以显著提升模型的表现。

PALRACE is a dataset consisting of 800 reading comprehension articles, created by Zhejiang University, with the goal of enhancing the interpretability of machine reading comprehension (MRC) tasks. A distinctive feature of this dataset is that each article is paired with human-annotated word-level rationales, which are labeled by at least 26 readers to justify the selection of answers. The articles in the PALRACE dataset are sourced from the RACE dataset, covering various question types including causal, factual, inferential, and others. The development of PALRACE not only supports supervised learning for rationales, but also establishes a standard benchmark for evaluating models' capabilities in locating and interpreting information. The primary application scenarios of this dataset focus on improving the interpretability and performance of MRC models; particularly when the model has not yet achieved state-of-the-art performance, the assistance of human rationales can significantly boost the model's performance.