DREEAM: Guiding Attention with Evidence for Improving Document-Level Relation Extraction

Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 1971–1983 May 2-6, 2023 ©2023 Association for Computational Linguistics DREEAM: Guiding Attention with Evidence for Improving Document-Level Relation Extraction Youmi Ma and An Wang and Naoaki Okazaki Tokyo Institute of Technology {youmi.ma@nlp., an.wang@nlp., okazaki@}c.titech.ac.jp Abstract Document-level relation extraction (DocRE) is the task of identifying all relations between each entity pair in a document. Evidence, de- fined as sentences containing clues for the re- lationship between an entity pair, has been shown to help DocRE systems focus on rel- evant texts, thus improving relation extraction. However, evidence retrieval (ER) in DocRE faces two major issues: high memory con- sumption and limited availability of annota- tions. This work aims at addressing these issues to improve the usage of ER in DocRE. First, we propose DREEAM, a memory-efficient ap- proach that adopts evidence information as the supervisory signal, thereby guiding the atten- tion modules of the DocRE system to assign high weights to evidence. Second, we propose a self-training strategy for DREEAM to learn ER from automatically-generated evidence on massive data without evidence annotations. Ex- perimental results reveal that our approach ex- hibits state-of-the-art performance on the Do- cRED benchmark for both DocRE and ER. To the best of our knowledge, DREEAM is the first approach to employ ER self-training1. 1 Introduction Document-level relation extraction (DocRE) has been recognized as a more realistic and challeng- ing task compared with its sentence-level coun- terpart (Peng et al., 2017; Verga et al., 2018; Yao et al., 2019). In DocRE, an entity can have multi- ple mentions scattered throughout a document, and relationships can exist between entities in different sentences. Therefore, DocRE models are expected to apply information filtering to long texts by fo- cusing more on sentences relevant to the current decision of relation extraction (RE) and less on ir- relevant ones. To this end, existing studies retrieve supporting evidence (evidence hereafter, Yao et al., 1The source code is available at https://github.com/ YoumiMa/dreeam [1] "The Archbishop" is the third episode of the first series of the BBC sitcom Blackadder ( The Black Adder ). [2] It is set in England in the late 15th century, and follows the exploits of the fictitious Prince Edmund as he is invested as Archbishop of Canterbury amid a Machiavellian plot by the King to acquire lands from the Catholic Church. [3] ... [5] Edmund, faced with the threat of assassination, attempts to escape to France into self-imposed exile; and in a later scene, two drunk knights overhear King Richard IV exclaiming "Who will rid me of this turbulent priest?" [6] The words attributed to King Henry II which led to Becket's death in 1170, and embark on a mission to murder Edmund. [7] … The Archbishop Subject: Prince Edmund Object: Blackadder Relation: present in work Evidence: 1,2 Figure 1: Example document and one of the relation triples from DocRED, where the i-th sentence is marked with [i] in the beginning. Mentions in bold …

Content truncated to 500 words