In Japan, the activities of local assembly members in prefectural assembly sessions are crucial within the framework of representative democracy. This study provides a comprehensive review of existing research that quantifies the activities of local assembly members through topic modeling. Furthermore, the BERTopic model was implemented to perform semantic clustering and topic labeling by combining bidirectional encoder representations from transformer (BERT) embeddings and clustering algorithms. The researchers used a dataset of local assembly minutes transformed into a machine-readable format and selected a subset of general questions (159 general questions from 32 members between 2011 and 2015, and 122 general questions from 23 members between 2016 and 2020) in Fukushima Prefecture. The results demonstrate that BERTopic exhibits robust semantic representation and visualization capabilities, thereby improving the understanding of relationships among the main topics. Compared to the Latent Dirichlet Allocation (LDA) and the Structural Topic Model (STM), BERTopic outperformed in terms of identifying subtle topics and yielded a higher topic coherence score, indicating more meaningful and consistent thematic extraction. Furthermore, this research contributes to documenting the recovery process in disaster-affected prefectures after the Great East Japan Earthquake and provides reference material for recovery efforts.