Named Entity Recognition of traditional architectural text based on BERT

Traditional architecture is an important component carrier of traditional culture. Through deep learning models, relevant entities can be automatically extracted from unstructured texts to provide data support for the protection and inheritance of traditional architecture. However, research on text information extraction oriented to this field has not been effectively carried out. In this paper, a data set of nearly 50,000 words in this field is collected, sorted out, and annotated, five types of entity labels are defined, annotation specifications are clarified, and a method of Named Entity Recognition based on pre-training model is proposed. BERT (Bidirectional Encoder Representations from Transformers) pre-training model is used to capture dynamic word vector information, Bi-directional Long Short-Term Memory (BiLSTM) module is used to capture bidirectional contextual information with positive and reverse sequences. Finally, classification mapping between labels is completed by the Conditional Random Field (CRF) module. The experiment shows that compared with other models, the BERT-BiLSTM-CRF model proposed in this experiment has a better recognition effect in this field, with F1 reaching 95.45%.