Abstract: Objectives: POI data are widely used in research fields such as urban functional area identification, point-ofinterest recommendation, service facility layout and urban vitality evaluation. Methods: To address the issue of inaccurate types in raw POI data, this paper constructs a POI classification mapping framework for building function classification tasks and proposes a BERT-DPCNN-based POI classification mapping method. The model employs the BERT pre-trained model to generate textual representations of POI names and utilizes a Deep Pyramid Convolutional Neural Network (DPCNN) to extract thematic text features for POI reclassification, thereby providing high-quality semantic information for building function classification. To verify the performance of the proposed model, we constructed baseline models by combining Word2Vec and BERT word embedding models with classical deep learning models such as Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN) to assess their classification effectiveness. Furthermore, we designated a section of Chengdu's third ring road as our test area to validate the necessity of our research. The procedure involved: ①applying BERT-DPCNN-based classification mapping to the area's POI data, ②performing building function classification using the remapped and raw POI data, and ③conducting a comparative evaluation against classifications result of the two types of data. Results: The results show that the BERT-DPCNN model performs optimally in terms of precision, recall and F1 value, with a classification mapping accuracy of 93.20%, a recall of 93.51%, and an F1 value of 93.57%; moreover, the BERT-DPCNN model performs the most stable mapping of all types of POIs, with mapping accuracies of all five types of POIs of more than 92%, and F1 values all are above 91%. The classification mapping of the raw POI data effectively reduces the probability of other types of buildings being misclassified as scientific and educational and medical, and the overall classification accuracy of buildings is improved from 65.63% to 80.21%, with an overall improvement of 14.58%. Conclusions: The constructed model can well extract text features, realize classification mapping of POI data, and provide high-quality data source for building function classification.