Named Entity Recognition (NER) in texts about Chinese traditional musical instruments faces challenges due to composite structures that blur entity boundaries and difficulties like dialect transliterations. To address these issues, this paper proposes a novel NER model that integrates adversarial training and a cross-attention mechanism. During the fine-tuning stage of Bidirectional Encoder Representations from Transformers (BERT) word embeddings, the model introduces adversarial training. It enhances the model’s ability to learn robust features by applying perturbations. Next, the model constructs a parallel feature extraction network comprising Iterated Dilated Convolutional Neural Network (IDCNN) and Bidirectional Gated Recurrent Unit (BiGRU). It also designs a cross-attention fusion module. This module enables the dynamic weighted fusion of context-aware features. Finally, the model employs a Conditional Random Field (CRF) decoder to constrain label transition logic. Experiments conducted on our self-built Chinese musical instrument text dataset show that the proposed model achieves an F1 score of 83.27%. This performance outperforms baseline models.