Boosting Document-Level Relation Extraction by Mining and Injecting Logical Rules

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 10311 - 10323 December 7-11, 2022 ©2022 Association for Computational Linguistics Boosting Document-Level Relation Extraction by Mining and Injecting Logical Rules Shengda Fan1∗, Shasha Mo1†, Jianwei Niu2,3 1 School of Cyber Science and Technology, Beihang University, Beijing 100191, China 2 Zhongguancun Laboratory, Beijing 100191, China 3 State Key Laboratory of Virtual Reality Technology and Systems, School of Computer Science and Engineering, Beihang University, Beijing 100191, China {fanshengda, moshasha, niujianwei}@buaa.edu.cn Abstract Document-level relation extraction (DocRE) aims at extracting relations of all entity pairs in a document. A key challenge to DocRE lies in the complex interdependency between the relations of entity pairs. Unlike most prior efforts focusing on implicitly powerful repre- sentations, the recently proposed LogiRE (Ru et al., 2021) explicitly captures the interdepen- dency by learning logical rules. However, Lo- giRE requires extra parameterized modules to reason merely after training backbones, and this disjointed optimization of backbones and extra modules may lead to sub-optimal results. In this paper, we propose MILR, a logic enhanced framework that boosts DocRE by Mining and Injecting Logical Rules. MILR first mines logical rules from annotations based on fre- quencies. Then in training, consistency reg- ularization is leveraged as an auxiliary loss to penalize instances that violate mined rules. Finally, MILR infers from a global perspec- tive based on integer programming. Compared with LogiRE, MILR does not introduce ex- tra parameters and injects logical rules dur- ing both training and inference. Extensive experiments on two benchmarks demonstrate that MILR not only improves the relation ex- traction performance (1.1%-3.8% F1) but also makes predictions more logically consistent (over 4.5% Logic). More importantly, MILR also consistently outperforms LogiRE on both counts. Code is available at https:// github.com/XingYing-stack/MILR. 1 Introduction Document-level relation extraction (DocRE) aims to identify relations of all entity pairs in a docu- ment, playing an essential role in knowledge graph construction (Luan et al., 2018), question answer- ing (Sorokin and Gurevych, 2017), etc. A key challenge to DocRE lies in the fact that relations ∗The first two authors contributed equally. †Corresponding author. Rules: spouse_of(𝑣0, 𝑣1) ↔ spouse_of(𝑣1, 𝑣0), parent_of(𝑣0, 𝑣1) ↔ child_of(𝑣1, 𝑣0), parent_of(𝒗𝟎, 𝒗𝟐) ← spouse_of(𝒗𝟎, 𝒗𝟏) ∧ parent_of(𝒗𝟏, 𝒗𝟐) Entities: Chusovitina, Alisher, Bakhodir Kurpanov [1] In 2002, Chusovitina’s son Alisher, then three years old, was diagnosed with leukemia. [2] … she moved to Germany with her husband Bakhodir Kurpanov, a former successful wrestler, and their son. [3] … Annotations & Predictions : child_of Alisher Bakhodir Kurpanov Chusovitina parent_of Figure 1: An example of the relational complex in DocRE and corresponding predictions produced by AT- LOP (Zhou et al., 2021a). Both solid and dotted arrows represent gold annotations of relations. Nevertheless, dotted arrows represent relation facts that ATLOP can- not identify. These missing facts can be directly ob- tained by the bold rule together with found facts. of entity pairs are not isolated. Rather, there exists complex interdependency between them. Consider the example in Fig. 1, where the text just explic- …

Content truncated to 500 words