Jointprop: Joint Semi-supervised Learning for Entity and Relation Extraction with Heterogeneous Graph-based Propagation

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 14541–14555 July 9-14, 2023 ©2023 Association for Computational Linguistics Jointprop: Joint Semi-supervised Learning for Entity and Relation Extraction with Heterogeneous Graph-based Propagation Zheng Yandan1,2, Hao Anran1 and Luu Anh Tuan1 1School of Computer Science and Engineering 2 Interdisciplinary Graduate Program-HealthTech Nanyang Technological University, Singapore {yandan002, s190003}@e.ntu.edu.sg anhtuan.luu@ntu.edu.sg Abstract Semi-supervised learning has been an impor- tant approach to address challenges in extract- ing entities and relations from limited data. However, current semi-supervised works han- dle the two tasks (i.e., Named Entity Recog- nition and Relation Extraction) separately and ignore the cross-correlation of entity and rela- tion instances as well as the existence of sim- ilar instances across unlabeled data. To alle- viate the issues, we propose Jointprop, a Het- erogeneous Graph-based Propagation frame- work for joint semi-supervised entity and re- lation extraction, which captures the global structure information between individual tasks and exploits interactions within unlabeled data. Specifically, we construct a unified span-based heterogeneous graph from entity and relation candidates and propagate class labels based on confidence scores. We then employ a propa- gation learning scheme to leverage the affini- ties between labelled and unlabeled samples. Experiments on benchmark datasets show that our framework outperforms the state-of-the-art semi-supervised approaches on NER and RE tasks. We show that the joint semi-supervised learning of the two tasks benefits from their codependency and validates the importance of utilizing the shared information between unla- beled data. 1 Introduction Named Entity Recognition (NER) and Relation Extraction (RE) are two crucial tasks in Informa- tion Extraction. Supervised learning schemes have made significant progress in NER and RE research by leveraging rich annotated data (e.g., Lin et al. (2020); Yamada et al. (2020); Baldini Soares et al. (2019). However, high-quality data annotation still involves extensive and expensive labor. Moreover, training NER and RE models in various domains and applications require diverse annotated data. Semi-supervised learning approaches (SSL) em- ploy a small amount of annotated data as a source (a) Before label propagation (b) After label propagation Figure 1: An example of label propagation. We repre- sent the sentence as a triplet (entity1, relation, entity2) which consists of an entity pair (circle) and a relation (tri- angle) in a graph structure. The colored nodes indicate labeled semantic units (entity or relation candidates), while the uncolored nodes represent the unlabeled se- mantic units. Purple denotes relation label Used-for, blue denotes for entity label Method, and orange denotes another entity label label Task. of supervision for learning powerful models at a lower cost. SSL in NER and RE have performed very well in recent years by employing bootstrapping, dis- tant supervision or graph-based approach (Batista et al., 2015; Zeng et al., 2015; Delalleau et al., 2005). However, they train a NER model (Yang et al., 2018; Chen et al., 2018; Lakshmi Narayan et al., 2019) or a RE model (Lin et al., 2019; Hu et al., 2021a; Li et al., 2021) separately. Therefore, they …

Content truncated to 500 words