Wikidata5M-SI

Wikidata5M-SI数据集是基于Wikidata5M构建的大规模基准，专注于知识图谱中的半归纳链接预测任务。该数据集包含5500条记录，主要用于评估模型对新实体的链接预测能力。数据集通过提供不同程度的上下文信息（从仅结构到包含文本提及和详细描述）来支持0-shot、few-shot和transductive任务。Wikidata5M-SI旨在解决大规模知识图谱中频繁出现新实体时的链接预测问题，避免模型重新训练的需求，并推动模型在未见实体上的泛化能力研究。

The Wikidata5M-SI dataset is a large-scale benchmark constructed based on Wikidata5M, focusing on the semi-inductive link prediction task in knowledge graphs. This dataset contains 5,500 records, and is primarily used to evaluate models' link prediction capabilities towards unseen entities. The dataset supports zero-shot, few-shot, and transductive tasks by providing varying levels of contextual information, ranging from structural-only data to text mentions and detailed descriptions. Wikidata5M-SI aims to address the link prediction problem when unseen entities frequently emerge in large-scale knowledge graphs, eliminate the need for model retraining, and advance research on the generalization capabilities of models towards unseen entities.