LLM-Enhanced Theme Identification and Classification of Urban Housing Policy Texts

Urban housing development has always been a key area of focus in urban studies. Accurate understanding of housing policy texts is of great significance for the sustainable development of urban housing. Different from traditional text analysis methods, we have proposed a machine learning (ML) framework driven by a large language models (LLMs) to quickly and accurately extract policy text themes and classify them. We used Mengzi-BERT-Base (MBB) for semantic encoding of policy documents, used GPT-4o to extract policy keywords and grasp the macro logic, and finally integrated the features of the two using a random forest model to output classification results. The results show that this framework divides housing policy texts into 4 core themes, with an overall accuracy rate of 70.25%, and the accuracy rate of individual themes is 75.56%, indicating that this framework has certain application value in urban policy text analysis.