Pu ZHANG (张谱)

As extreme heatwaves intensify into a primary public health threat under climate change, understanding the dynamics of public perception is critical for effective risk communication. While traditional methods are limited in scale and timeliness, social media offers an unprecedented real-time window into public discourse. This study provides the first large-scale, computational analysis of Chinese public discourse on extreme heat, leveraging 348,417 user comments from the platform Douyin. After addressing multicollinearity through a rigorous variable reduction process, our novel framework combines a Large Language Model (LLM) for thematic classification with an FDR-corrected regression model and an explainable AI (SHAP) analysis to identify and interpret the drivers of public concern. Thematically, we find a stark divergence between the most frequent topics (e.g., regional comparisons) and the most publicly endorsed topics (e.g., fatalism and mutual care), revealing latent community values. Our regression analysis, corrected for multiple comparisons, reveals a powerful and statistically robust link: higher maximum temperatures directly drive increased public discourse on both “Health Impacts” and “Entertainment and Emotional Venting.” The SHAP analysis provides a deeper layer of insight, revealing that underlying geographical context (i.e., average elevation) and socioeconomic status exert the greatest overall influence on the nature of the conversation, shaping everything from tourism-related discussions to expressions of social vulnerability. These findings demonstrate that public discourse on heat is shaped by a hierarchy of factors, from statistically robust climate triggers to highly impactful socioeconomic and geographical contexts. This work provides a scalable framework for a dynamic Social and Health Impact Assessment (SIA/HIA), underscoring the need for tailored communication strategies that account for both environmental exposure and social vulnerability.

Sorting out the changing characteristics of online public opinion triggered by a series of events in the investigation and assessment of major natural disasters is of great practical significance for optimizing the work of disaster investigation and assessment, governing the ecology of online public opinion, and enhancing the effect of comprehensive disaster reduction. In this paper, we collected relevant comments from several official media accounts, such as People’s Daily, and evaluated their emotional color using a sentiment analysis method based on the BERT fine-tuning model. Furthermore, keyword co-occurrence semantic network theme analysis is conducted for texts presenting negative emotional overtones to assess the changes in public opinion hotspots. The impact of the relevant online public opinion characteristics and the release of the disaster investigation report on them was assessed in the context of the investigation report itself. Based on sorting out the characteristics of online public opinion on several related topics, targeted public opinion governance initiatives and relevant suggestions for improving the disaster investigation system are proposed. This paper is of positive significance for studying disaster public opinion and improving the effectiveness of emergency management of rainstorms and flooding disasters.

The escalating conflict between doctors and patients in China has become a critical societal issue,with social media emerging as a primary arena for …

In the era of digital governance, social media platforms have become critical arenas for citizen-state interaction and policy feedback. However, governments often struggle to translate massive, unstructured public discourse into actionable intelligence, frequently relying on aggregated metrics that obscure critical regional nuances. This study introduces a spatially-aware computational framework to bridge this gap. Integrating Large Language Model (LLM) based few-shot learning with SHAP (SHapley Additive exPlanations) and GIS analysis, we examine the public reception of China’s 2025 national childcare subsidy policy using a dataset of 352,448 comments. Our analysis reveals a stark divergence between information-seeking behaviors (high volume) and substantive legitimacy debates (high engagement). Crucially, we demonstrate that the socioeconomic drivers of this discourse—such as gender ratios and educational attainment—are highly spatially heterogeneous, necessitating distinct governance responses across different regions. This research contributes to the field of government information systems by validating a novel, replicable methodology for precision policy analytics, offering policymakers a tool to move beyond national averages and achieve more responsive, context-specific governance.

China’s industrialization threatens water systems and public health. This study uses a large language model to classify 170,000 Douyin comments into nine topics and seven emotions, identifying their provincial heterogeneity. We then applied a robust Beta Regression framework—using twelve provincial predictors and controlling for spatial autocorrelation and FDR to model the drivers of this heterogeneity. The analysis reveals a striking divergence: provincial-level factors are strong predictors of what citizens discuss, but not how they feel. We identified 10 robust drivers for topics; most notably, higher provincial ’Average Schooling Years’ strongly predicts "Public Health" discourse (beta = 1.638, q < 0.001). In stark contrast, after the same rigorous correction, zero provincial variables significantly predict emotional expression (all q > 0.1). This "topic-emotion split" provides a novel solution for environmental governance, enabling agencies to tailor strategies to specific, context-driven topics rather than just generalized public emotion.