面向社交媒体立场检测的数据增强方法  

Data Augmentation Approach for Social Media Stance Detection

作　　者：苏致中 席耀一[1] 陈宇飞 曹蓉 马洁琼 SU Zhizhong;XI Yaoyi;CHEN Yufei;CAO Rong;MA Jieqiong(Information Engineering University, Zhengzhou 450001, China;Unit 92020, Qingdao 266000, China)

机构地区：[1]信息工程大学,河南郑州450001 [2]92020部队,山东青岛266000

出　　处：《信息工程大学学报》2022年第1期58-65,共8页Journal of Information Engineering University

基　　金：国家社会科学基金青年科学基金资助项目(19CXW027)。

摘　　要：现有的公开立场检测评测中标签语料规模较小,制约深度学习立场检测模型的训练效果。数据增强技术能一定程度缓解训练数据不足带来的困扰。对于现有的数据增强技术易将社交媒体语言中的冗余信息进行扩充,无法有针对性增强的问题,提出一种基于层次注意力的数据混合增强方法。该方法首先从粗粒度段句出发,将决定立场判断的关键信息筛选出来;然后再进行细粒度词语变换,从而实现数据增强。实验表明,相较于另外两种数据增强方法,所提方法对立场检测模型的检测效果提升更加明显,证实了该方法的有效性。The small size of the label corpus in the existing public stance detection evaluation restricts the training effect of the deep learning stance detection model.Data augmentation can alleviate the trouble caused by insufficient training data to a certain extent.To address the problem that the existing data enhancement technology is easy to expand the redundant information in the social media language,and cannot be enhanced in a targeted manner,this paper proposes a data mixing augmentation approach based on hierarchical attention.Starting from the coarse-grained paragraph sentences,the algorithm is designed to filter out the key information that determines the stance and then the fine-grained word transformation is augmented.Experiments show that compared with two other data augmentation methods,the proposed method improves the detection effect of the stance detection model more significantly,which confirms the effectiveness of the method in this paper.

关 键 词：社交媒体 立场检测 深度学习 数据增强 层次注意力 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]