Towards Integration of Discriminability and Robustness for Document-Level Relation Extraction

Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 2606–2617 May 2-6, 2023 ©2023 Association for Computational Linguistics Towards Integration of Discriminability and Robustness for Document-Level Relation Extraction Jia Guo∗1,2, Stanley Kok1, Lidong Bing2 1School of Computing, National University of Singapore 2DAMO Academy, Alibaba Group guojia@u.nus.edu, skok@comp.nus.edu.sg l.bing@alibaba-inc.com Abstract Document-level relation extraction (DocRE) predicts relations for entity pairs that rely on long-range context-dependent reasoning in a document. As a typical multi-label classifica- tion problem, DocRE faces the challenge of effectively distinguishing a small set of pos- itive relations from the majority of negative ones. This challenge becomes even more diffi- cult to overcome when there exists a significant number of annotation errors in the dataset. In this work, we aim to achieve better integra- tion of both the discriminability and robust- ness for the DocRE problem. Specifically, we first design an effective loss function to endow high discriminability to both probabilistic out- puts and internal representations. We innova- tively customize entropy minimization and su- pervised contrastive learning for the challeng- ing multi-label and long-tailed learning prob- lems. To ameliorate the impact of label errors, we equipped our method with a novel nega- tive label sampling strategy to strengthen the model robustness. In addition, we introduce two new data regimes to mimic more realistic scenarios with annotation errors and evaluate our sampling strategy. Experimental results verify the effectiveness of each component and show that our method achieves new state-of- the-art results on the DocRED dataset, its re- cently cleaned version, Re-DocRED, and the proposed data regimes. 1 1 Introduction The problem of document-level relation extraction (DocRE) has garnered increasing attention from the research community (Quirk and Poon, 2017; Peng et al., 2017; Yao et al., 2019) due to its importance to real-world applications. DocRE is inherently a multi-label problem, in which we have to predict ∗This work was partially done when Jia Guo was an intern at DAMO Academy, Alibaba Group. 1Our codes and datasets are available at https:// github.com/guojiapub/PEMSCL. a set of relations from the pre-defined label set for every entity pair in a document. Thus, it is crucial for DocRE models to adopt an effective learning objective that can clearly distinguish massive se- mantically close relations. Recently, several works have proposed new loss functions to learn an adaptive threshold for better separating positive and negative relations. How- ever, these approaches (Zhou et al., 2021; Tan et al., 2022a) either enforce learning a total order among all relations that leads to superfluous comparisons and diminishing differences among them or im- properly penalize all pre-defined labels of positive entity pairs if their average margins are lower than the threshold when addressing the label imbalance problem (Zhou and Lee, 2022). In contrast, we propose an approach that learns a partial order, ranking all positive relations above a threshold indi- vidually, which is in turn ranked above all negative relations. Our approach does not waste precious data and probability mass in modeling the ordering among positive …

Content truncated to 500 words