Construction of Dual-Channel Intangible Cultural Heritage Knowledge Graph Based on Improved BERT Model

There are lots of long entities, nested structures and professional terms in texts in Intangible Cultural Heritage (ICH), which leads to insufficient accuracy in named entity recognition, thus restricting the quality and automation level of knowledge graph construction. This study proposes an improved BERT model that integrates dual-channel feature enhancement and boundary-aware attention mechanisms. Its core is a dual-channel architecture—specifically, the parallel combination of IDCNN and BiLSTM—that collaboratively capture local features and global dependencies, with a boundary-aware attention mechanism further enhancing entity boundary discrimination. The boundary-aware attention mechanism is introduced to accurately identify entity boundaries, and a knowledge graph construction model based on high-precision named entity recognition is designed. Experiments on the self-built ICH text dataset and the general MSRA dataset show that the precision, recall rate, and F1-score of the proposed model on the ICH dataset reach 94.50%, 94.98%, and 94.74%, which are 2.65%, 6.42%, and 4.33% higher than the baseline model. When dealing with long entities with a character length of not less than 15, the F1-score is improved by 12.82%, and the average degree median and clustering coefficient median of the graph constructed are improved by 5.9% and 14.5% compared with the baseline method. Meanwhile, the extraction accuracy of high-frequency and low-frequency relationships reaches a maximum of 99.2%, which is 16.12% higher than the baseline. The results demonstrate that the proposed method effectively enhances the accuracy of entity recognition and relationship extraction. Unlike the simple stacking of existing technologies, the innovation of this framework lies in its design of a collaborative optimization architecture integrating dual channels and attention mechanisms, which addresses the complex challenges of ambiguous entity boundaries and multi-scale features in intangible cultural heritage texts. It realizes the automatic construction of high-quality and scalable ICH knowledge graph, and can provide reliable support for the digital protection and knowledge mining of ICH.