T-Rex : A Large Scale Alignment of Natural Language with Knowledge Base Triples

Several datasets with alignments between knowledge base triples and free text have been built, for several independent tasks, such as Relation Extraction, Knowledge base population, Relation Discovery…But where the others datasets have a small number of documents (<em>TAC-KBP</em>), only have the relations without the original documents (<em>FB15K-237</em>), have a little amount of relations (<em>Google-RE</em>), we present a dataset containing large-scale high-quality alignments between DBpedia abstracts and Wikidata triples, <em>T-REx</em>.

目前已针对关系抽取（Relation Extraction）、知识库填充（Knowledge base population）、关系发现（Relation Discovery）等多项独立任务，构建了若干实现知识库三元组（knowledge base triples）与自由文本（free text）对齐的数据集。然而现有同类数据集均存在不同程度的局限：TAC-KBP的文档数量有限，FB15K-237仅包含关系而无原始文本，Google-RE的关系体量偏小。为此，我们提出了T-REx数据集，该数据集包含DBpedia摘要与Wikidata三元组之间的大规模高质量对齐数据。