A Simple yet Effective Relation Information Guided Approach for Few-Shot Relation Extraction

Findings of the Association for Computational Linguistics: ACL 2022, pages 757 - 763 May 22-27, 2022 c⃝2022 Association for Computational Linguistics A Simple yet Effective Relation Information Guided Approach for Few-Shot Relation Extraction Yang Liu♠♢, Jinpeng Hu♠♢, Xiang Wan♢♣†, Tsung-Hui Chang♠♢† ♠The Chinese University of Hong Kong (Shenzhen) ♢Shenzhen Research Institute of Big Data ♣Pazhou Lab, Guangzhou, 510330, China ♠{yangliu5, jinpenghu}@link.cuhk.edu.cn ♢wanxiang@sribd.cn ♠changtsunghui@cuhk.edu.cn Abstract Few-Shot Relation Extraction aims at predict- ing the relation for a pair of entities in a sen- tence by training with a few labelled examples in each relation. Some recent works have intro- duced relation information (i.e., relation labels or descriptions) to assist model learning based on Prototype Network. However, most of them constrain the prototypes of each relation class implicitly with relation information, generally through designing complex network structures, like generating hybrid features, combining with contrastive learning or attention networks. We argue that relation information can be intro- duced more explicitly and effectively into the model. Thus, this paper proposes a direct addi- tion approach to introduce relation information. Specifically, for each relation class, the relation representation is first generated by concatenat- ing two views of relations (i.e., [CLS] token embedding and the mean value of embeddings of all tokens) and then directly added to the original prototype for both train and prediction. Experimental results on the benchmark dataset FewRel 1.0 show significant improvements and achieve comparable results to the state-of-the- art, which demonstrates the effectiveness of our proposed approach. Besides, further analyses verify that the direct addition is a much more effective way to integrate the relation represen- tations and the original prototypes. 1 2 1 Introduction Relation Extraction (RE) (Bach and Badaskar, 2007) is a fundamental task of Natural Language Processing (NLP), which aims to extract the re- lations between entities in sentences and can be †Corresponding author. 1The code is released at https://github.com/ lylylylylyly/SimpleFSRE. 2Main results in this paper can be found in the CodaLab competition (liuyang00) at https://competitions. codalab.org/competitions/27980#results. Figure 1: The figure is an intuitive illustration of the difference in ways to introduce relation information be- tween most existing works and our proposed approach. The orange vector and the blue vector denote represen- tations of relations and prototypes, respectively. applied to other advanced tasks (Li et al., 2021; Hu et al., 2021). However, RE usually suffers from la- beling difficulties and train data scarcity due to the massive cost of labour and time. In order to solve the problem of data scarcity, Few-Shot Relation Extraction (FSRE) (Han et al., 2018; Gao et al., 2019a; Qu et al., 2020; Yang et al., 2021) task has become a research hotspot in academia in recent years. The task is firstly to train on large-scale data on existing relation types, and then quickly migrate to a small amount of data on new relation types. Inspired by the success of few-shot learning in the computer vision (CV) community (Sung et al., 2018; Garcia and Bruna, 2018), various methods are introduced into FSRE. One of the …

Content truncated to 500 words