FB20K数据集

FB20K是基于FB15K构造了一个包含zero-shot entity的数据集，涉及4种类型的测试数据，其中，e-e表示头尾实体均出现在训练集中，d-e、e-d分别表示测试三元组中的头实体、尾实体为新实体（即没有出现在训练集中），d-d表示头尾实体均为新实体。数据集存储规模约45MB。

FB20K is a dataset constructed based on FB15K that includes zero-shot entities. It covers four types of test data: e-e means both the head and tail entities of the test triple appear in the training set; d-e indicates that the head entity of the test triple is a novel entity (i.e., not present in the training set); e-d denotes that the tail entity of the test triple is a novel entity (i.e., not present in the training set); and d-d means that both the head and tail entities are novel entities. The storage size of this dataset is approximately 45 MB.

搜集汇总

背景与挑战

背景概述

FB20K数据集是基于FB15K构建的，专注于零样本实体场景，包含四种测试类型（e-e、d-e、e-d、d-d）以区分实体在训练集中的出现情况，数据集存储规模约为45MB。该设计旨在评估知识图谱补全模型在新实体上的泛化能力。