Research on Chinese Named Entity Recognition Based on BERT-CNN-BiLSTM-CRF Model with Fusion Multi-Head Attention Mechanism

In the domain of natural language processing, Named Entity Recognition (NER) has emerged as a pivotal task for knowledge graph construction and information extraction. Due to the characteristics of polysemy and ambiguous word boundaries, the identification and extraction of Chinese entities pose greater challenges. In response to the shortcomings of traditional language processing models in extracting local semantic features of Chinese entities and effectively obtaining information from different positions in text statements, this paper integrates the multi-head attention mechanism with BERT, CNN, BiLSTM, and CRF components to propose a new entity recognition model. The BERT pre-trainedmodel generates word vectors that encapsulate contextual semantic information. A Convolutional Neural Network (CNN) then combines and extracts multi-level semantic information to enrich the feature set. The Bidirectional Long Short-Term Memory Network (BiLSTM) further processes these vectors to extract features, while the multi-head attention mechanism enhances the model's capability to capture long-distance dependencies within the text. Finally, a Conditional Random Field (CRF) decodes and generates the sequence of entity labels. Experimental results demonstrate that the performance of this model outperforms other language processing models on both the low-sample Dgre dataset and the high-sample Duie dataset, showing more excellent performance on the latter.