FewRel 2.0

该数据集是一个大规模的少样本关系分类数据集，它包含了来自维基百科的70,000个句子，覆盖了100个关系类别，并由众包工作者进行标注。具体来说，该数据集被划分为64个基础类别用于训练，16个类别用于验证，以及20个新颖类别用于测试。数据集的规模达到了70,000个句子和100个关系类别，其任务是进行少样本关系分类。

This dataset is a large-scale few-shot relational classification dataset. It contains 70,000 sentences sourced from Wikipedia, spanning 100 relational categories, and was annotated by crowd workers. Specifically, the dataset is partitioned into three subsets: 64 base categories for training, 16 categories for validation, and 20 novel categories for testing. With a scale of 70,000 sentences across 100 relational categories, its core task is few-shot relational classification.

搜集汇总

背景与挑战

背景概述

FewRel 2.0是一个大规模小样本关系分类数据集，专注于少样本学习场景下的关系分类任务。其评估基于N-way K-shot方法，要求模型在给定少量支持示例的情况下对查询示例进行分类。详细信息可参考EMNLP18论文。