Re-TACRED (Revised-TACRED)

Re-TACRED 数据集是用于关系提取的 TACRED 数据集的显着改进版本。使用新的众包标签，Re-TACRED 修剪了注释不佳的句子并解决了 TACRED 关系定义的歧义，最终纠正了 23.9% 的 TACRED 标签。该数据集包含分布在 40 个关系中的超过 91,000 个句子。在 AAAI 2021 上展示的数据集。论文 (arXiv)：https://arxiv.org/abs/2104.08398

Re-TACRED is a significantly improved version of the TACRED dataset for relation extraction. Using new crowdsourced annotations, Re-TACRED trims poorly annotated sentences and resolves ambiguities in the relational definitions of TACRED, ultimately correcting 23.9% of TACRED's original labels. This dataset contains over 91,000 sentences distributed across 40 distinct relation categories. It was presented at AAAI 2021. The associated arXiv paper: https://arxiv.org/abs/2104.08398

搜集汇总

背景与挑战

背景概述

Re-TACRED是TACRED关系抽取数据集的改进版本，修正了23.9%的原标签错误，包含91,000多个句子覆盖40种关系，通过众包标签解决了原数据集的质量问题。该数据集由卡内基梅隆大学和微软于2021年发布。