Cracking Chinese Named Entity Recognition (NER) remains challenging, largely from semantic ambiguity. To tackle this, we’ve developed a novel approach, BTBGC, which incorporates a Graph Convolutional Network (GCN). First, in the embedding layer, BERT generates combined embeddings for token, segment, and position to represent the input text. Second, in the encoding layer, BERT is initially deployed to derive context-specific features from the text. Then, the self-attention weights from the Transformer in BERT are used to initialize a graph structure, which is encoded using a GCN to capture more effective global features. The outputs from BERT encoding and GCN encoding are combined to enhance text feature representation. Finally, in the decoding layer, a Conditional Random Field (CRF) is used for decoding to output label sequence with the highest probability. The proposed method achieves F1 scores of 65.4%, 90.0%, and 91.4% on datasets from three different domains: Weibo, cMedQANER, and EduNER respectively, outperforming other baseline models. Empirical data validate BTBGC’s efficacy and adaptability.