Topical and emotional expressions regarding extreme weather disasters on social media: a comparison of posts from official media and the public

The study addresses how official media in China framed and reported the July 2021 Zhengzhou (Henan) extreme rainfall and flood disaster and how the public responded on social media. With extreme weather events increasing globally, understanding media frames and public perceptions is critical for agenda-setting and inclusive climate adaptation policy. Social media plays central roles for both authorities and the public during crises (information seeking, risk communication, sentiment detection, and coordination). Prior work shows mass media can bias narratives toward certain frames (e.g., emphasizing hazards over social vulnerability, or government triumph/solidarity), potentially shaping public opinion. However, the linkage and dynamics between official media posts and public resonance remain underexplored, especially outside COVID-19 contexts. Using Weibo interactions, the authors examine: RQ1 topics and emotions during emergency response; RQ2 during crisis learning after the accountability report; and RQ3 changes in topics/emotions and differences across periods.

The paper situates its inquiry within several strands: (1) rising frequency and impacts of extreme weather worldwide and the policy salience of disaster-driven discussion; (2) social media in disasters for information seeking, collective intelligence, situational awareness, and coordination by citizens and authorities; (3) media framing effects and disaster myths, including biases toward certain narratives (e.g., focusing on hazard rather than vulnerability, or promoting government success and social cohesion), and differences across political contexts; (4) limited evidence on direct media–public interaction dynamics on social platforms beyond COVID-19 communications; and (5) known inequalities in media attention to marginalized or rural areas during large-scale disasters, which can affect relief distribution. These literatures motivate analyzing topical frames and emotions of official media posts versus public comments over time.

Case: July 2021 Zhengzhou (Henan, China) extreme precipitation and floods leading to 398 deaths (380 in Zhengzhou) and large economic losses. A central government accountability investigation identified governance and preparedness failures; report released Jan 21, 2022, with penalties for 97 officials and actions against contractors. Data source: Sina Weibo. Identified 38 national official media accounts authorized to produce news online; 20 reported the event; 12 accounts met inclusion (≥3 posts on the event or ≥100 comments). Collected all relevant posts (manual) and public comments (programmatic crawling) related to the Henan floods. Temporal segmentation: Emergency response period (Jul 20, 2021–Jan 20, 2022) and crisis lessons learning period (Jan 21, 2022–Mar 31, 2022; one month after report release). Counts: Emergency response—570 official media posts and 20,553 public comments. Crisis learning—37 posts and 4,455 comments. Preprocessing: Removed non-Unicode chars, blanks, punctuation, and non-Chinese phrases via regex; de-duplicated sentences using mechanical compression; tokenized Chinese text with Jieba; removed stop-words. Topic modeling: Constructed bag-of-words for each post/comment. Applied Latent Dirichlet Allocation (gensim). Determined number of topics k via perplexity minimization; optimal k=5 in both periods. Extracted top-10 keywords per topic and topic proportions. Emotion analysis: Used DUTIR affective lexicon (27,466 Chinese emotion words) with 7 major categories: depression, like, dislike, fear, surprise, joy, anger (and 21 subcategories). Counted emotion word frequencies per corpus and aggregated to major categories; assigned each record an emotion label based on highest-frequency category. Cross-tabulated emotions by topics to produce distribution heat maps and compared media vs public across periods.

Data volume: Emergency response period—570 posts, 20,553 comments; Crisis learning—37 posts, 4,455 comments. Emergency response topics (official media; Table 1): - Hazard (30.8%): extreme/heavy rain, precipitation, flood, warnings, weather, Henan, Zhengzhou. - Location (19.8%): specific sites with casualties (metro stations, tunnels, highways, hospital, weather station). - Rescue (17.1%): rescue, vehicles, firefighters, search, volunteers, trapped/escaped, flood prevention. - Metro (16.8%): subway Line 5 incident, rescue of trapped passengers, operational failure, driver/passengers. - Impacts (15.4%): deaths, statistics, emergency, recovery, trapped/perished, stress. Emergency response topics (public comments; Table 1): - Help-seeking (24.0%): appeals for rescue, reports of trapped/casualties, especially from rural/neglected areas (Xinxiang, Gongyi, Weihui). - Moral support (21.7%): encouragement for rescuers, solidarity, hope. - Donation (18.6%): charity, donations, Red Cross/foundations, trust/channels. - Condolence (18.1%): honoring victims and heroes, mourning. - Places-neglected (17.6%): highlighting underreported rural/township areas. Crisis learning topics (official media; Table 2): - Attribution (27.7%): risk awareness shortages, information hiding, deaths/missing, exposure, accountability. - Investigation (24.8%): processes and actors (mayor Xu Liyi, CPC officials, State Council), investigation report. - Punishment (24.2%): prosecutions, police actions, disciplinary measures, accountability for officials and related personnel. - Response lessons (14.7%): organizational shortcomings, duty fulfillment, flood prevention, housing/recovery. - Prevention lessons (8.7%): prevention, adaptation, accountability, risk eradication. Crisis learning topics (public comments; Table 2): - Condolence (24.6%). - Lessons (20.8%): learning from history, strong accountability, understanding disaster nature. - Condemn (20.5%): calling out information hiding, governance failures, resignations, disappointment. - Encourage (18.0%): praise heroes, accountability, hope for no future disasters. - Disaster (16.1%): revisiting impacts and mixed natural/man-made causes. Emotions (Fig. 3, Fig. 4): - Emergency response: Official media emotions ranked: like > anger > dislike > depression > fear > joy > surprise; Public: anger > like > joy > dislike > depression > fear > surprise. Largest divergence in anger (higher in public) and like (higher in media). Liking in media concentrated on hazard and impacts topics; public anger concentrated on help-seeking; public joy linked to moral support. - Crisis learning: Official media: depression > like > dislike > anger > fear > joy > surprise; Public: anger > dislike > like > depression > fear > joy > surprise. Media showed strong depression around attribution, response lessons, and punishment; some liking around investigation process. Public anger spread across all topics, especially high when expressing condolence; depression higher when revisiting disaster impacts. Overall: Media focused on factual reporting and later on investigation/punishment/lessons; public focused on help-seeking, moral support, condolence, praise/condemnation, and reiterating impacts. Emotions showed limited coherence: public anger dominated both periods; media shifted from like (July 2021) to depression (Jan 2022).

Findings address the research questions by revealing distinct topical frames and emotional expressions between official media and the public and how these evolved from immediate response to post-report learning. Official media emphasized event facts, high-profile urban infrastructure failures, and later institutional accountability and lessons, consistent with state-aligned framing. Public comments highlighted grassroots needs—especially neglected rural areas—using official media posts as venues for help-seeking and situational reporting, alongside moral support and mourning. Significance: The documented media attention bias toward urban, high-salience scenes risks overlooking marginalized regions, potentially skewing relief allocation. Public use of official media comment spaces demonstrates a practical channel for surfacing needs of underserved groups, suggesting operational value for emergency managers to monitor and integrate such inputs. Emotionally, persistent public anger—centered on unmet needs during response and grief during learning—indicates limits of media framing effects on public sentiment in this context and underscores the need for empathetic, responsive risk communication. Platform moderation and display policies may shape visible discourse, yet notable divergence remained. These insights inform crisis informatics, risk communication strategies, and policy learning after extreme events.

The study contributes evidence on divergent topical frames and emotional dynamics between official media and the public on Weibo surrounding the 2021 Henan floods. During the emergency response, official media emphasized hazards, locations, metro failure, rescue, and impacts, while the public focused on help-seeking from neglected rural areas, moral support, donations, and condolence. Following the accountability report, media centered on attribution, investigation, punishment, and lessons for response and prevention; public discourse turned to condolence, lessons, condemnation, encouragement, and reiteration of impacts. Emotionally, public anger dominated both stages; media shifted from liking to depression. Future research should compare central versus local official media to detect differential framing and expand beyond short social posts to full news articles and broadcasts to analyze narrative strategies. Integrating social media-derived needs assessments into emergency operations could improve equitable relief for marginalized communities.

- Sampling of media limited to national-level official outlets; patterns may differ for local media and other state-owned outlets not included. - Short length of social media posts constrains analysis of nuanced narrative strategies compared to full news reports. - Platform moderation and comment display controls may affect visibility and representativeness of public sentiment and topics observed.