Recurrent Interaction Network for Jointly Extracting Entities and Classifying Relations

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 3722–3732, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 3722 Recurrent Interaction Network for Jointly Extracting Entities and Classifying Relations Kai Sun BDBC and SKLSDE Beihang University,China sunkai@buaa.edu.cn Richong Zhang∗ BDBC and SKLSDE Beihang University, China zhangrc@act.buaa.edu.cn Samuel Mensah BDBC and SKLSDE Beihang University, China samensah@buaa.edu.cn Yongyi Mao School of EECS University of Ottawa,Canada ymao@uottawa.ca Xudong Liu BDBC and SKLSDE Beihang University, Beijing, China liuxd@act.buaa.edu.cn Abstract The idea of using multi-task learning ap- proaches to address the joint extraction of en- tity and relation is motivated by the related- ness between the entity recognition task and the relation classification task. Existing meth- ods using multi-task learning techniques to ad- dress the problem learn interactions among the two tasks through a shared network, where the shared information is passed into the task- specific networks for prediction. However, such an approach hinders the model from learning explicit interactions between the two tasks to improve the performance on the indi- vidual tasks. As a solution, we design a multi- task learning model which we refer to as re- current interaction network which allows the learning of interactions dynamically, to effec- tively model task-specific features for classi- fication. Empirical studies on two real-world datasets confirm the superiority of the pro- posed model. 1 Introduction The extraction of entities and relations from tex- tual data comprises of two sub-tasks: entity recog- nition (ER) and relation classification (RC). The ER task aims at extracting all entities in a given text. The RC task aims at classifying the re- lation between any pair of entities in the text. In practice, both tasks are required to be solved jointly, and have been observed to contribute sig- nificantly in extracting structured knowledge from unstructured text for several applications, includ- ing knowledge base construction (Komninos and Manandhar, 2017; Deng et al., 2019; Nathani et al., 2019). For instance, consider the sentence John was born in Sheffield, a city in the north of Eng- land. The goal of a joint entity and relation ex- traction task is to identify all the factual rela- tional triples (or relational facts) (Sheffield, ∗Corresponding author birth place of, John) and (England, contains, Sheffield). S A B S A B (a) Flat Structure (b) Graph Structure Figure 1: Two topological structures for multi-task learning. Here, A and B are related tasks, and S is the shared information of the two tasks. The directed edges define the information flow. The simplest approach to solve this joint task is to utilize a pipeline-based approach by firstly extracting all entities in the sentence and then clas- sifying the relation between all entity pairs (Ze- lenko et al., 2003; Zhou et al., 2005; Chan and Roth, 2011). However, pipeline-based approaches disregard the correlation between ER and RC tasks, leading to error propagation in these methods. Recently, researchers have exploited multi-task learning (MTL) (Collobert and Weston, 2008) tech- niques to capture the correlation between the ER and RC …

Content truncated to 500 words