STAD: Self-Training with Ambiguous Data for Low-Resource Relation Extraction

Proceedings of the 29th International Conference on Computational Linguistics, pages 2044–2054 October 12–17, 2022. 2044 STAD: Self-Training with Ambiguous Data for Low-Resource Relation Extraction Junjie Yu1 Xing Wang2 Jiangjiang Zhao3 Chunjie Yang3 Wenliang Chen1 1Institute of Artificial Intelligence, School of Computer Science and Technology, Soochow University, China 2Tencent AI Lab, Shenzhen, China 3China Mobile Online Marketing and Services Center, China jjyu@stu.suda.edu.cn, brightxwang@tencent.com, wlchen@suda.edu.cn {zhaojiangjiang, yangchunjie}@cmos.chinamobile.com Abstract We present a simple yet effective self-training approach, named as STAD, for low-resource relation extraction. The approach first clas- sifies the auto-annotated instances into two groups: confident instances and uncertain in- stances, according to the probabilities predicted by a teacher model. In contrast to most previ- ous studies, which mainly only use the confi- dent instances for self-training, we make use of the uncertain instances. To this end, we pro- pose a method to identify ambiguous but use- ful instances from the uncertain instances and then divide the relations into candidate-label set and negative-label set for each ambiguous instance. Next, we propose a set-negative train- ing method on the negative-label sets for the ambiguous instances and a positive training method for the confident instances. Finally, a joint-training method is proposed to build the final relation extraction system on all data. Ex- perimental results on two widely used datasets SemEval2010 Task-8 and Re-TACRED with low-resource settings demonstrate that this new self-training approach indeed achieves signifi- cant and consistent improvements when com- paring to several competitive self-training sys- tems.1 1 Introduction Relation Extraction (RE) is a fundamental task in Information Extraction, which aims to obtain a pre- defined semantic relation between two entities in a given sentence (Zhou et al., 2005). In recent years, fine-tuning on the downstream RE tasks with pre- trained models (Soares et al., 2019) has achieved significant progress with the rapid development of the “Pre-train and Fine-tune” Paradigm (De- vlin et al., 2019) which leverages large-scale un- labeled data. However, RE still suffers from the data scarcity problem. For most RE tasks, due to 1 Code is publicly available at https://github.com/jjyunlp/ STAD The taste is from the ginseng , not alcohol . Probability 0.0 0.5 1.0 entity-origin(e1,e2) cause-effect(e2,e1) others Most of birdtown's earliest residents hailed from eastern european countries. Probability 0.0 0.5 1.0 entity-origin(e1,e2) cause-effect(e2,e1) others (a) Confident instance The taste is from the ginseng , not alcohol . Probability 0.0 0.5 1.0 entity-origin(e1,e2) cause-effect(e2,e1) others Most of birdtown's earliest residents hailed from eastern european countries. Probability 0.0 0.5 1.0 entity-origin(e1,e2) cause-effect(e2,e1) others (b) Uncertain instance Figure 1: Two examples of auto-annotated instances. For simplicity, we list two detailed relations and use “others” to represent the other relations. the task-specific definition of relations, the lack of customized annotation data poses great challenge for the supervised RE (Hendrickx et al., 2010; Zhang et al., 2017). Meanwhile, manually labeling large-scale RE data is extremely time-consuming, expensive, and laborious. As an alternative, auto- matically building annotated data for RE attracts a lot of attention in the research community (Mintz et al., 2010; Luo et al., …

Content truncated to 500 words