Enhanced Clustering Algorithm of Public Opinion Texts in Public Emergencies Network Considering Sentence-Level Similarity

To improve the clustering performance of sentence-level textual data in public emergency network opinion analysis, this study proposes a deep learning-driven grid-intensive clustering algorithm tailored for similarity-based text processing. A heterogeneous directed graph framework is designed to optimize the storage and retrieval architecture of public opinion data, while feature space reconstruction techniques are employed to restructure textual representations. The proposed method incorporates semantic relevance fusion for adaptive feature extraction and dynamic scheduling of similarity features in emergency-related discourse. By integrating fuzzy C-means clustering with deep learning-enhanced fragment peak fusion, the algorithm enables multi-dimensional attribute classification and grid-reinforced clustering. Experimental results demonstrate that the optimized algorithm significantly enhances convergence performance — reducing error rates by 18.7% compared to baseline methods — and exhibits strong adaptive learning capabilities, particularly in textual pattern recognition and public opinion evolution detection. Moreover, the hybrid architecture achieves a balance between computational efficiency and clustering precision, reducing processing time by 32% and improving the F1-score by 15.2%. These results confirm the algorithm’s effectiveness for real-time analysis of large-scale public opinion data during crisis events.