S2ynRE: Two-stage Self-training with Synthetic data for Low-resource Relation Extraction

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 8186–8207 July 9-14, 2023 ©2023 Association for Computational Linguistics S2ynRE: Two-stage Self-training with Synthetic data for Low-resource Relation Extraction Benfeng Xu1∗, Quan Wang2, Yajuan Lyu3, Dai Dai3 Yongdong Zhang14 and Zhendong Mao14† 1University of Science and Technology of China 2Beijing University of Posts and Telecommunications, 3Baidu Inc. 4Institute of Artificial Intelligence, Hefei Comprehensive National Science Center benfeng@mail.ustc.edu.cn, zdmao@ustc.edu.cn Abstract Current relation extraction methods suffer from the inadequacy of large-scale annotated data. While distant supervision alleviates the prob- lem of data quantities, there still exists do- main disparity in data qualities due to its re- liance on domain-restrained knowledge bases. In this work, we propose S2ynRE, a frame- work of two-stage Self-training with Synthetic data for Relation Extraction. We first lever- age the capability of large language models to adapt to the target domain and automati- cally synthesize large quantities of coherent, realistic training data. We then propose an ac- companied two-stage self-training algorithm that iteratively and alternately learns from syn- thetic and golden data together. We conduct comprehensive experiments and detailed abla- tions on popular relation extraction datasets to demonstrate the effectiveness of the pro- posed framework. Code is available at https: //github.com/BenfengXu/S2ynRE. 1 Introduction Relation extraction systems aim at discovering rela- tional knowledge between entities by reading from unrestricted texts (Cardie, 1997). Although neural methods, especially pre-trained language models, have greatly advanced the state-of-the-art relation extraction capability (Zeng et al., 2014; Wu and He, 2019), they still require large quantities of training data (Han et al., 2020). However, high-quality an- notations are usually very expensive to obtain, mak- ing low-resource relation extraction a very practical challenge in many real-world scenarios. Distant supervision (Mintz et al., 2009), which automatically annotates relational statements by aligning entities with an existing knowledge bases (Bollacker et al., 2008; Vrandeˇci´c and Krötzsch, 2014), has been widely explored as an effective way to construct large scale relational ∗Joint work of USTC and Baidu Inc. †Corresponding author: Zhendong Mao dataset. Many recent works exploit such data in a pretraining stage to learn relational representa- tions (Baldini Soares et al., 2019; Peng et al., 2020; Qin et al., 2021). Although this line of meth- ods have seen certain improvements, they still in- evitably raise the concern that the distantly anno- tated data can vary considerably from downstream tasks both in target schema and in context distribu- tions, thus may not be able to offer optimal trans- ferability. For instance, due to the reliance on ex- isting knowledge bases, current works mostly re- sort to Wikidata as the source of relational triples and Wikipedia (Vrandeˇci´c and Krötzsch, 2014) as the corpus for distant supervision. This circum- scribes distant data to only factual knowledge be- tween world entities, while downstream tasks may be of other special interests involving various do- mains, ranging from semantic relation between nominals (Hendrickx et al., 2009) to chemical- protein interactions (Kringelum et al., 2016). Meanwhile, recent advances in …

Content truncated to 500 words