A Hierarchical Entity Graph Convolutional Network for Relation Extraction across Documents

Proceedings of Recent Advances in Natural Language Processing, pages 1022–1030 Sep 1–3, 2021. https://doi.org/10.26615/978-954-452-072-4_115 1022 A Hierarchical Entity Graph Convolutional Network for Relation Extraction across Documents Tapas Nayak∗ Department of Computer Science Indian Institute of Technology Kharagpur tnk02.05@gmail.com Hwee Tou Ng Department of Computer Science National University of Singapore nght@comp.nus.edu.sg Abstract Distantly supervised datasets for relation ex- traction mostly focus on sentence-level extrac- tion, and they cover very few relations. In this work, we propose cross-document relation extraction, where the two entities of a rela- tion tuple appear in two different documents that are connected via a chain of common enti- ties. Following this idea, we create a dataset for two-hop relation extraction, where each chain contains exactly two documents. Our proposed dataset covers a higher number of relations than the publicly available sentence- level datasets. We also propose a hierarchical entity graph convolutional network (HEGCN) model for this task that improves performance by 1.1% F1 score on our two-hop relation ex- traction dataset, compared to some strong neu- ral baselines. 1 Introduction The idea of distant supervision (Mintz et al., 2009) eliminates the need for manual annotation for ob- taining training data for relation extraction. Previ- ously, this idea is used mostly to create sentence- level datasets. However, the assumption of distant supervision, that the two entities of a tuple must ap- pear in the same sentence, is overly strict. We may not find an adequate number of evidence sentences for many relations as both entities do not appear in the same sentence. The relation extraction mod- els built on such data can find relations only for a small number of relations and the relations of most knowledge bases (KBs) will be out of the reach of such models. To address this issue, we propose a multi-hop re- lation extraction task where the subject and object entities of a tuple can appear in two different docu- ments, and these two documents are connected via ∗ This work was done when the first author was a PhD student at the National University of Singapore. some common entities. We can create a chain of entities from the subject entity to the object entity of a tuple via the common entities across multi- ple documents. Each link in this chain represents a relation between the entities located at the end- points of the link. We can determine the relation between the subject and object entities of a tuple by following this chain of relations. This approach can give training instances for more relations than sentence-level distant supervision. Following the proposed multi-hop approach, we create a two-hop relation extraction dataset for the task. Each in- stance of this dataset has two documents, where the first document contains the subject entity and the second document contains the object entity of a tuple. These two documents are connected via at least one common entity. This idea can be extended to create an N-hop dataset. We also propose a hierarchical entity graph con- volutional network (HEGCN) …

Content truncated to 500 words