系统功能语言学理论视角下突发公共卫生事件谣言用户立场识别研究

王丹丹 杨艳妮 张瑞

摘 要：[目的/意义]突发公共卫生事件情境下，研究谣言传播中的用户立场识别，可为谣言真实性检测开辟新视角，为谣言治理与舆论引导提供新思路。[方法/过程]以COVID-19疫情期间微博上虚假谣言为研究对象，以提高谣言下评论帖子的立场分类准确性为研究目标，构建规范化用户立场检测分析建模框架，基于系统功能语言学（SFL）理论实现用户评论中特征的全面多维提取，通过统计和可视化分析筛选最优特征，比较分析不同模型组合实验结果，剖析影响分类准确性的样本因素。[结果/结论]研究提出的基于概念元功能、语篇元功能、人际元功能的特征组合对以往特征进行了有效梳理和补充;使用最优特征而非全体特征作为树形结构分类器输入，采用集成学习方法，可保证总体分类效果、缩短训练时间;数据分布不平衡性显著影响不同立场的评论识别准确性，“其他”立场识别效果更优，特征选取有力弥补了“询问”立场数据量上的分类劣势。

关键词：突发公共卫生事件;立场识别;系统功能语言学;分类器;集成学习;COVID-19疫情;微博;谣言

DOI：10.3969/j.issn.1008-0821.2021.02.003

〔中图分类号〕G206.2 〔文献标识码〕A 〔文章编号〕1008-0821（2021）02-0019-11

Abstract：[Purpose/Significance]In the context of public health emergencies，the research on user standpoint identification in rumor spreading can open up a new perspective for rumor authenticity detection，and provide new ideas for rumor governance and public opinion guidance.[Method/Process]This paper took the false rumors on the microblog during the COVID-19 pandemic as the research object，aimed at improving the accuracy of the standpoint classification of comments under rumors，constructed the standardized rumor standpoint detection and analysis modeling framework.Based on the theory of Systemic Functional Linguistics（SFL），the comprehensive multidimensional feature extraction of user comments was realized and the optimal features were screened out by statistical and visual analysis.The experimental results of different model combinations were compared and analyzed and the sample factors influencing the accuracy of classification were analyzed.[Result/Conclusion]The feature combination based on ideational meta function，textual meta function and interpersonal meta function effectively combed and complemented the previous features;the optimal features rather than all features being used as the input of tree structure classifier and adopting the ensemble learning method could ensure the overall classification effect and shorten the training time;and the imbalance of data distribution significantly affected the accuracy of the comment recognition of different standpoints.The result of“other”standpoint recognition was better，and feature selection was powerful，which maked up for the disadvantage of classification of“standpoint”position data.

Key words：public health emergency;standpoints recognition;SFL;classifier;ensemble learning;COVID-19 epidemic;microblog;rumor

突發公共卫生事件情境下，网络谣言甚嚣尘上，在线社交媒体凭借参与、公开、交流、对话、社区化特性，增加了谣言传播速度、广度与深度[1]。如COVID-19疫情期间“中部战区空军在武汉上空播撒消毒粉液”“武汉市民使用高浓度酒精室内消毒引发火灾”“新型冠状病毒可能诞生于人为设计的基因改造”等不实信息大肆传播，不仅加剧了恐慌情绪蔓延，也影响公众对科学防疫的正确认识，各种阴谋论更成为大国之间抗疫联盟形成的绊脚石。……