Analysis of public response during health crises is crucial for policy-making. While clustering method is commonly used for analyzing public reactions, traditional topic modeling approaches like LDA face scalability and interpretability challenges when applied to large-scale social media data. This study presents an AI-enhanced two-stage clustering framework that combines graph-based micro-clustering with temporal pattern analysis, augmented by multi-AI interpretation using Claude, GPT-4, DeepSeek, and offline baseline models. We define “AI” as neural language model-based systems (LLMs), while conventional statistical models, such as TF-IDF and LDA, serve as baselines. Applied to complete datasets (32 million original tweets), our AI-enhanced approach demonstrates enhanced specificity in topic identification and stronger correlation with temporal events compared to sampling-based LDA approaches, while achieving automated interpretation through multi-AI consensus without requiring manual coding. This study addresses the core question of whether multi-AI interpretation and multi-faceted similarity measurement can provide stable, scalable, and reliable analysis for large-scale Japanese social media discourse.
