A Novel Table-to-Graph Generation Approach for Document-Level Joint Entity and Relation Extraction

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 10853–10865 July 9-14, 2023 ©2023 Association for Computational Linguistics A Novel Table-to-Graph Generation Approach for Document-Level Joint Entity and Relation Extraction Ruoyu Zhang1, Yanzeng Li1, Lei Zou1,2∗ 1Wangxuan Institute of Computer Technology, Peking University. Beijing, China 2TopGraph.AI {ry_zhang, zoulei}@pku.edu.cn liyanzeng@stu.pku.edu.cn Abstract Document-level relation extraction (DocRE) aims to extract relations among entities within a document, which is crucial for applications like knowledge graph construction. Existing methods usually assume that entities and their mentions are identified beforehand, which falls short of real-world applications. To overcome this limitation, we propose TAG, a novel table- to-graph generation model for joint extraction of entities and relations at document-level. To enhance the learning of task dependencies, TAG induces a latent graph among mentions, with different types of edges indicating differ- ent task information, which is further broadcast with a relational graph convolutional network. To alleviate the error propagation problem, we adapt the hierarchical agglomerative clustering algorithm to back-propagate task information at decoding stage. Experiments on the bench- mark dataset, DocRED, demonstrate that TAG surpasses previous methods by a large margin and achieves state-of-the-art results1. 1 Introduction Relation extraction (RE) is the task to extract re- lational facts from natural language text, which plays a crucial role in various downstream tasks, e.g. knowledge graph construction and question answering (Yih et al., 2015; Trisedya et al., 2019; Li and Zou, 2022). Early studies mostly focus on sentence-level RE, i.e. predicting relations among entities in one single sentence. However, in real- world scenarios such as Wikipedia articles or sci- entific papers, large amounts of relational facts are expressed across multiple sentences, which neces- sitate inter-sentence reasoning skills. Hence, recent efforts have been moving towards the more realis- tic document-level RE (DocRE) (Yao et al., 2019; Nan et al., 2020; Zhou et al., 2021). ∗Corresponding author. 1https://github.com/ridiculouz/TaG Subject: Balboa Boneke Object: Equatorial Guinean Relation: country of citizenship Subject: Balboa Boneke Object: Valencia Relation: place of death Subject: Valencia Object: Spain Relation: country Juan Balboa Boneke (9 June 1938 – 10 March 2014) was an Equatorial Guinean politician and writer. … After his exile, he settled down in Valencia with his second wife and her family. Balboa Boneke died from renal problems, coupled with a three-year depression caused by the death of his wife, on 10 March 2014 in Valencia , Spain . Figure 1: An example adapted from the DocRED dataset. Mentions refer to the same entity are in same color. We omit some relations and denote some entities with underline for clarity. Despite the rapid progress, most previous DocRE methods solely focus on the task of relation extraction, which assumes that entities and their corresponding mentions are given beforehand. As shown by Figure 1, to extract both of entities and relations at document-level, a natural idea is to use a pipeline approach. Traditionally, it first divide the whole task into subtasks of mention extrac- tion (ME), coreference resolution (COREF) and relation extraction …

Content truncated to 500 words