基于社会媒体的自杀风险研究:自杀意念及污名化态度

PSYCH OpenIR > 社会与工程心理学研究室

	基于社会媒体的自杀风险研究:自杀意念及污名化态度
其他题名	Suicidal Risk Facts Identification Based on Analyses of Social Media:Suicide Ideation And Suicide Stigmatization
	田玮
导师	朱廷劭
	2017-05
摘要	自杀是一种全球性的公共健康问题，当个体遭遇感情失败、事业受挫、人生坎坷就可能会失去信心、进而产生心理问题，甚至有可能选择自杀，自杀率的逐年上升己经成为社会公共健康迫切需要解决的严重问题。鉴此，自杀预防工作对维护社会稳定，促进社会公共健康起着尤为重要的作用。随着互联网的迅猛发展，越来越多的社会媒体用户选用微博作为自我情绪情感的表达途径，这些情绪情感的表达也包含了自杀直播和针对自杀的个人态度。这就为自杀预防工作提供了新的契机，许多学者也已经开始尝试利用社会媒体信息进行与自杀相关的各种分析研究工作，但现有的基于微博博文的自杀预防的研究还处于起步阶段，成果较少。本研究基于新浪微博这个平台，从个体自杀意愿和社会针对自杀行为所形成的污名化态度两方面，分别从主观和客观因素对微博一博文进行分析，通过三个研究，试图达成两个科学目标:第一，使用深度学习算法建立社会媒体自杀和自杀无名化态度识别器，探讨通过社交平台实时评估个体用户自杀意愿和自杀污名化态度的有效性。第二，通过对直播自杀持有污名化态度的微博进行语言使用特征差异的研究，探讨利用语言使用特征差异过滤污名化网络信息的可能性。研究结果显示: (1)深度学习算法对自杀意愿的识别的准确率可以达到93. 5%，正确率较其它两种算法高出2. 5%左右;深度学习算法对自杀污名化识别的准确率可以达到79, 8%，正确率较其它两种算法高出10%。 (2)针对对直播自杀持有污名化态度微博的研究，非参数检验和效果量分析结果表明，“功能词“和“副词“等5个特征维度差异有统计学意义且效果较大(p<0. 05, Cohen's d>0. 5)9“代名词“和“朋友词“等21个特征维度差异有统计学意义且效果中等(p<0. 05, 0. 2<=Cohen's d<0. 5 )。其中，自杀污名化组的“朋友词“、“焦虑词“、“时间词“、“工作词“和“休闲词“的使用频率低于非自杀污名化组。自杀污名化组的“功能词“和“动词“、等21个语言维度的使用频率高于非自杀污名化组。 (3)针对对直播自杀持有污名化态度微博用户的研究，非参数检验和效果量分析结果表明，“副词‘夕、“语助词“、“应和词“和“’填充赘词“4个特征维度差异有统计学意义且效果较大(p<0. 05, Cohen's d>0. 5 ) 9“代名词“和“生气词“等34个特征维度差异有统计学意义且效果中等(p<0. 05, 0. 2<=Cohen's d<0. 5 )。其中，自杀污名化组“工作词“和“体闲词“的使用频率低于非自杀污名化组。‘自杀污名化组的“代名词“和“副词“等36个语言维度的使用频率高于非自杀污名化组。本文的实验结果表明利用深度学习识别器对微博用户自杀意愿和自杀污名化态度的预测是有效可行的，同时对微博直播自杀的污名化态度的语言使用特征差异分析结果也可以为有效从社会客观环境预防自杀提供有效支持。但针对微博用户自杀污名化态度的语言使用特征差异分析由于现有数据量的不足有待进一步研究。
其他摘要	Suicide is a worldwide public health problem. When individuals experience their life setback, demise of relationship and failure of business, they may lose confidence and fell into the psychological crisis. Some even choose to end their life by suicide, and the increasing of suicide rate has endangered social public health. Therefore, suicide prevention plays an important role in maintaining social stability and improving social public health. With the development of Internet, more and more people would like to express their emotion and feeling in social media, including suicidal ideation and attitude towards suicide. rt is a new opportunity for suicide prevention. Researchers have begun to use social media data to analyze the information related with suicide; Nevertheless, the current study of suicide prevention basing on microblog is still in the early stage, and there is little research achievements. This project is focused on individual suicide ideation and suicide stigmatization of social to analyze the microblog base on the platform of Sina-microblog, and thinking from both subjective and objective facts. It has achieved two scientific objectives via three studies. The 1st objective is to set up the suicidal ideation and suicide stigmatization recognizer with deep learning, and address the effectiveness of suicidal risk and suicide stigmatization assessment in social media. The 2nd objective is to research the differences in linguistic features between suicide stigmatization and non-stigmatization of suicide, and explore the possibility of suicide stigmatization identification base on linguistic features. The results of the three studies are shown as follows: (1) The accuracy rate of suicide ideation base on deep learning is 93.5%, which is 2.5% more than other two algorithms. The accuracy rate of suicide stigmatization base on dcep learning is 79%, which is 10% more than other two algorithms. (2) Considering the i0esearch of microblog, the results showed that "Function word" and "Adverb", etl have obvious dissimilarity with good effect (P<0.05, Cohen's d>0.5). There are }l linguistic features have obvious dissimilarity with medium effect (P<0:05, 0.2<=Cohen's d<0.5), including "Pronoun" and "Friends words", etl. Stigmatized Weibo expressed less in "Friends words", "Anxiety words", "Tlme words", "Word words" and "Leisure words", but more frequently use "Function words" and "Verb", etl than non-stigmatized Weibo. (3) Considering the research of microblog user, the results showed that "Adverb", "Interjunction words", "Assent wolds" and "Filler words" have obvious dissimilarity with good effect (P<0.05, Cohcn's d>0.5). There are 34 linguistic features have obvious dissimilarity with medium effect (P<0.05, 0.2<=Cohen's d<0.5), including "Pronoun" and "Anger Words", etl. Stigmatized Weibo expressed less in "Work words" and "Leisure words", but more frequently use "Function Words", "Pronoun" and "Adverb", etl than non-Stigmatized Welbo. The results indicate that deep learning algorithm works more effectively for suicide risk and suicide stigmatization identification. And also, the differencesresults in linguistic features between suicide stigmatization and non-stigmatization of suicide against microblog can effectively support the suicide prevention from society objective circumstances. However, the differences in linguistic features of microblog users need further investigation due to lack of data.
学科领域	健康心理学
关键词	自杀风险因素微博文本分析深度学习污名化
学位类型	硕士
语种	中文
学位名称	理学硕士
学位专业	健康心理学
学位授予单位	中国科学院大学;中国科学院心理研究所
学位授予地点	北京
统计方法	卡方检验;独立样本t检验;回归分析
取样对象	人类
性别	男;女
病症	自杀行为
被试数量	实验一：7301;实验二：4967;实验三：4582
国家或地区	中国
统计软件	SPSS 19.0
文献类型	学位论文
条目标识符	https://ir.psych.ac.cn/handle/311026/28668
专题	社会与工程心理学研究室
作者单位	1.中国科学院大学; 2.中国科学院心理研究所
推荐引用方式 GB/T 7714	田玮. 基于社会媒体的自杀风险研究:自杀意念及污名化态度[D]. 北京. 中国科学院大学;中国科学院心理研究所,2017.

条目包含的文件
文件名称/大小	文献类型	版本类型	开放类型	使用许可
田玮-硕士学位论文.pdf（5117KB）	学位论文		开放获取	CC BY-NC-SA	请求全文