RelO: An Overlapping Relation Extraction Dataset and Model

Code

• github.com/disquietbookshare/relo
  OfficialIn paperpytorch★ 0

Abstract

We introduce an overlapping relation extraction dataset(RelO) constructed from Wikipedia and Wikidata. RelO consists of 308,579 sentences, representing 40 relations and three different overlapping types. We evaluate the state-of-the-art relation extraction methods on RelO and results show that RelO is challenging for these relation extraction methods. We also use RelO as a transfer learning resources and the fine-tuned two models achieve state-of-the-art results on other related datasets. To handle the overlapping relation extraction task, We extend a pipeline framework by utilizing overlapping type information and semantic distance information of relation representations. Through careful experiments, we validate the importance of these information for overlapping relation extraction. All details and resources about the dataset and baselines are released on https://github.com/disquietBookShare/Relo.

Tasks

Reproductions