表情符號賦能:基於社群媒體情感分析的 災害心理復原力與管理策略研究;Emoji Empowerment: Disaster Psychological Resilience and Management Strategies Based on Social Media Sentiment Analysis

NCU Institutional Repository > 管理學院 > 資訊管理研究所 > 博碩士論文 > Item 987654321/98776

jsp.display-item.identifier=請使用永久網址來引用或連結此文件: https://ir.lib.ncu.edu.tw/handle/987654321/98776

题名:	表情符號賦能:基於社群媒體情感分析的災害心理復原力與管理策略研究;Emoji Empowerment: Disaster Psychological Resilience and Management Strategies Based on Social Media Sentiment Analysis
作者:	周世哲 Chou, Shih-che
贡献者:	資訊管理學系
关键词:	災害管理;社群媒體;情緒分析;主題建模;心理復原;disaster management;social media;sentiment analysis;topic modeling;psychological recovery
日期:	2025-07-22
上传时间:	2025-10-17 13:20:37 (UTC+8)
出版者:	國立中央大學
摘要:	氣候變遷推升極端氣候事件之強度與頻率，臺灣受地震與颱風雙重威脅，社群媒體已成民眾即時回報災情與抒發情緒的主要管道，現有研究多限於單一災害或災中單一時段，難以系統掌握災前、災中、災後連續情緒軌跡與心理復原歷程。本研究蒐集2019至2024年Dcard上與地震、颱風相關128,619筆貼文與留言，透過資料清洗、資料前處理及NTUSD情緒字典初標註後抽樣複核，完成三向情緒標籤，以LDA建構災前、災中、災後三階段之主題模型，提出整合ALBERT、TextCNN、GRU 與階層式注意力輕量模型，將 emoji 文字化特徵納入以提升分類效能。Kolmogorov-Smirnov 二樣本檢定結果顯示，災中貼文量與 emoji 密度皆顯著高於災前與災後(p < 0.01)，反映事件高峰期資訊交流與情緒互動最為活躍，採 Spearman與Kendal等級相關及Theil-Sen中位數斜率迴歸(無母數方法)檢驗社群特徵對心理復原時間??r影響，地震事件中，討論量??t與心理復原時間??r呈現顯著正相關 (Spearman ρ=0.791, ??S=0.002；Kendall τ=0.659, ??K= 0.005；Theil-Sen β = +0.00266)，負向情緒平均佔比??，向情緒平均佔比?? 同樣延長??r(Spearman ρ = 0.839, ??S= 0.001；β = +4.785)，而正未達顯著。颱風事件則僅觀察到??t與??r正向效應(Spearman ρ = 0.703, ??S= 0.002；τ = 0.553, ??K= 0.005；β =+0.00306)，??與??斜率皆不顯著。本研究建構可於一般設備即時部署輕量 Transformer-emoji 融合分析框架，量化貼文熱度與情緒組成對群體復原時間之雙重效應，證實資訊熱度延滯復原現象，這項研究為政府與政策管理者在災害期間，提供實證基礎，從而強化社會的韌性與反脆弱能力。;Climate-change-driven intensification of extreme events exposes Taiwan simultaneously to earthquakes and typhoons. While social media has become the primary channel for citizens to report damage and voice emotions in real time, most prior studies examine a single hazard or a single phase of the event cycle, leaving the full emotional trajectory before, during and after disasters unclear. This study harvested 128 619 earthquake- and typhoon-related posts and comments published on the Taiwanese platform Dcard between 2019 and 2024. After data cleaning and tokenisation, the NTUSD sentiment lexicon was applied and manually audited to obtain three-class sentiment labels (positive, neutral, negative). Latent Dirichlet Allocation was then used to construct topic models for the pre event, event-day and post-event stages. A lightweight sentiment classifier that integrates ALBERT, TextCNN, GRU and hierarchical attention, augmented with emoji-to-text features, is proposed to improve classification performance.Kolmogorov–Smirnov two sample tests reveal that both posting volume and emoji density rise significantly during the disaster compared with the pre- and post-stages (p < 0.01), indicating that information exchange and emotional interaction peak at the hazard apex. Non-parametric analyses Spearman and Kendall rank correlations combined with Theil-Sen median-slope regression were employed to examine how social-media features affect community psychological recovery time ??r. In the earthquake subset (n = 12), the discussion peak ??t is significantly positively correlated with the psychological recovery time ??r. (ρ = 0.791, PS = 0.002; τ = 0.659, ??K = 0.005; β = +0.00266 day · post?1) and a higher negative sentiment share ?? further lengthens recovery (ρ = 0.839, ??S = 0.001; β = +4.785 days %?1), whereas the positive-sentiment share ?? is non-significant. In the typhoon subset (n = 16), only ??t retains a significant positive effect on ??r (ρ = 0.703, ??S = 0.002; τ = 0.553, ??K = 0.005; β = +0.00306 day · post?1); neither ?? nor ?? is significant.The proposed Transformer-emoji framework can be deployed on commodity hardware and quantitatively captures the dual influence of posting heat and sentiment composition on collective recovery. The results corroborate an “information-heat delay” effect higher online attention slows emotional rebound and provide empirical evidence that can inform governmental decision-makers seeking to strengthen societal resilience and antifragility during compound disasters.
显示于类别:	[資訊管理研究所] 博碩士論文

文件中的档案:

档案	描述	大小	格式	浏览次数
index.html		0Kb	HTML	39	检视/开启

检视Licence

在NCUIR中所有的数据项都受到原著作权保护.

社群 sharing

数据加载中.....