Public feedback analysis on multi-stage emergency management policies using BERTopic-SKEP integrated model

Understanding public feedback across multi-stage emergency manaquerygement policies is critical for enhancing disaster governance efficacy. This study proposes an integrated deep learning framework combining BERTopic topic modeling and SKEP sentiment analysis to quantify public perceptions of government actions during the unprecedented "23.7" Beijing Rainstorm. Leveraging 50,015 social media posts, we systematically analyze policy-related public concerns and emotional responses through the four-stage emergency management lifecycle: prevention, preparedness, response, and recovery. Key innovations include: 1.BERTopic-SKEP fusion: Overcoming limitations of traditional LDA in short-text processing by integrating semantic embeddings and structured sentiment knowledge; 2.Multi-stage dynamics: Revealing significant disparities in public attention (response/recovery stages > 76% vs. prevention/preparedness < 2.5%) and sentiment polarity (recovery-stage positivity: 78.18%); 3. integration of Narrative Policy Framework (NPF) to interpret topic modeling results as structured policy narratives, identifying elements such as characters (e.g., government as hero/planner), plots, and morals, revealing public perceptions of government roles in emergency management. Results demonstrate that the proposed framework effectively captures nuanced public demands while providing actionable insights for optimizing policy legitimacy and operational efficacy.