Enhancing named entity recognition in low-resource scenarios using diffusion variational autoencoder and knowledge graphs

In low-resource Named Entity Recognition (NER) tasks, existing deep learning methods often suffer from inaccurate entity boundary identification and insufficient contextual understanding due to scarce annotated data. To address this problem, this paper proposes a diffusion variational knowledge graph-enhanced NER model (DV-KG-NER). The method operates through four key steps. First, it utilizes knowledge graph embeddings and graph attention networks to construct structured entity representations and fuse them with textual features. Second, a variational autoencoder encodes the fused features into latent representations. Third, a diffusion mechanism performs multi-step denoising processes on these latent representations to generate diverse training samples. Finally, graph attention networks complete entity recognition and boundary determination. The entire process employs a joint optimization strategy for end-to-end training. Experiments on eight datasets, including BC5CDR, SciERC, and CoNLL-2003, demonstrate that DV-KG-NER outperforms existing methods across multiple evaluation metrics. The model shows particularly improved performance in low-resource scenarios, such as biomedical and scientific literature domains.