基于易感人格语言特征的抑郁风险分析与预测  

Analysis and prediction of depression based on language characteristics of vulnerable personality

作　　者：胡晓俊 张鹏[2] 甘国兵 吴斌 张烁 HU Xiaojun;ZHANG Peng;GAN Guobing;WU Bin;ZHANG Shuo(School of Education,Tianjin University,Institute of Applied Psychology,Tianjin University,Tianjin 300350,China;College of Intelligence and Computing,Tianjin University;Artificial Intelligence Research&Development Department in Qishuo(Tianjin)Intelligent Technology Limited Company)

机构地区：[1]天津大学教育学院,应用心理研究所,300072 [2]天津大学智能与计算学部 [3]起硕(天津)智能科技有限公司人工智能研发事业部

出　　处：《中国健康心理学杂志》2023年第9期1281-1287,共7页China Journal of Health Psychology

基　　金：国家自然科学基金(编号:62276188);面向可持续发展的人工智能公益研究计划(编号:AI4SDGs)。

摘　　要：目的:基于社交媒体数据,探究易感人格语言特征与抑郁风险之间的关系,为抑郁预防与心理健康服务提供依据。方法:通过Python爬取新浪微博用户在2016-2021年的原创微博文本,并用易感人格与“文心”词典提取词类特征,构建多种机器学习模型进行抑郁分析与预测。结果:抑郁与非抑郁两类用户在易感人格语言的使用上存在显著差异,表现为抑郁用户在以下5维度的词频上显著高于非抑郁用户:封闭防御(F=700.32,P<0.001)、敏感好胜(F=671.50,P<0.001)、自我专注(F=590.09,P<0.001)、退让顺从(F=514.05,P<0.001)、严谨认真(F=48.57,P<0.001);其次,基于易感人格特征进行抑郁预测,在准确率、精确率及F1分数上比“文心”高出0.4%~6.5%。本文考虑到两类特征可能存在互补性,合并两词典中所有显著的特征,在梯度提升树分类器上预测效果最好,准确率达83.9%,F1分数达82.4%。结论:本文开发的抑郁易感人格词典性能良好,提取的语言特征有利于解释抑郁的成因,且能较为准确地对抑郁风险进行自动预测。Objective:To explore the relationship between linguistic characteristics of vulnerable personality and depression risk in social media,which provides evidence for depression prevention and mental health services.Methods:Python was used to crawl the microblog texts of Sina Weibo users from 2016 to 2021,and a variety of machine learning models were built to analyze and predict depression based on the feature of word class in vulnerable personality and TextMind.Results:There were significant differences between depressed and non-depressed users in the use of language of vulnerable personality.The word frequency of depressed users was significantly higher than that of non-depressed users in the following five dimensions:Such as defensiveness(F=700.32,P<0.001),emulousness(F=671.50,P<0.001),self-absorption(F=590.09,P<0.001),compliance(F=514.05,P<0.001),perfectionism(F=48.57,P<0.001).Secondly,compared with the features of TextMind,the prediction of depression based on the feature of vulnerable personality was better in accuracy,precision,and F1 score.Considering the possible complementarity of the two types of features,this paper combined all the significant features in the above dictionaries.The best prediction results were achieved on the Gradient Boosting Decision Tree,with an accuracy of 83.9%and an F1 score of 82.4%.Conclusion:The performance of the lexicon of personality vulnerability to depression in this paper is good.The linguistic features extracted from the lexicon are helpful to explain the causes of depression and predict the risk of depression automatically more accurately.

关 键 词：抑郁易感人格 抑郁风险识别 心理词典 社交媒体 

分 类 号：R395.4[哲学宗教—心理学]