WIKIHOP, MEDHOP

WIKIHOP和MEDHOP是两个跨文档的多跳阅读理解数据集，由伦敦大学学院的研究人员创建。WIKIHOP基于维基百科文章，旨在通过多个文档中的信息来回答关于特定实体属性的问题。MEDHOP则基于MEDLINE摘要，目标是基于科学发现关于药物和蛋白质及其相互作用的信息，来确定药物-药物相互作用。这两个数据集都利用了现有的知识库（如WIKIDATA和DRUGBANK）作为事实依据，通过远距离监督来生成数据。数据集的应用领域包括信息提取、搜索和问答系统，旨在解决需要从分散的文本证据中进行多步推理的问题。

WIKIHOP and MEDHOP are two cross-document multi-hop reading comprehension datasets created by researchers from University College London. WIKIHOP is based on Wikipedia articles, aiming to answer questions regarding specific entity attributes by leveraging information from multiple documents. MEDHOP, by contrast, is built upon MEDLINE abstracts, with the goal of identifying drug-drug interactions based on scientific findings about medications, proteins and their interactions. Both datasets utilize existing knowledge bases such as WIKIDATA and DRUGBANK as factual foundations, and generate data via distant supervision. The application fields of these datasets cover information extraction, search and question answering systems, targeting problems that require multi-step reasoning from scattered textual evidence.