business_person_company

一种用于远程监督关系提取的模型不可知去噪方法。我们提出了一种基于影响函数的干净实例选择新标准。它收集用于识别良好实例的样本级证据 (比损失级证据更具信息性)。我们还提出了一种师生机制，用于在引导清洁集时控制中间结果的纯度。

A model-agnostic denoising method for distant supervision relation extraction is proposed. We introduce a novel criterion for clean instance selection based on influence functions, which collects sample-level evidence (more informative than loss-level evidence) to identify high-quality instances. Furthermore, we present a teacher-student mechanism to control the purity of intermediate results during the bootstrapping of the clean dataset.

搜集汇总

背景与挑战

背景概述

该数据集旨在支持远程监督关系提取，提出了一种基于影响函数的模型不可知去噪方法，通过师生机制提升数据纯度。它由上海交通大学、字节跳动人工智能实验室和华东师范大学于2022年联合发布。