Abstract: In low-resource Named Entity Recognition (NER) tasks, many transfer learning-based methods can alleviate data scarcity but often lead to the omission or incorrect recognition of some valid information, affecting the model's performance in resource-constrained environments. To address this issue, this paper proposes a multi-module collaborative NER model, Diff-Cascade-NER. First, a Variational Autoencoder (VAE) is used to learn data representations in the latent space and generate diverse samples. Then, context information, syntactic analysis, and VAE-reconstructed data are fed as conditional inputs to a Conditional Encoder (CE) for encoding. The encoded data is passed to a Cascade Diffusion Mechanism (CDM), which generates high-quality samples through multi-stage denoising and generation processes. Finally, Adversarial Learning (AL) optimizes the quality and diversity of generated samples. Experimental results show that, compared with existing models, Diff-Cascade-NER outperforms on eight low-resource datasets, particularly achieving F1 scores of 85.44% and 56.38% on the BC2GM and WNUT-16 datasets, respectively, demonstrating the effectiveness of the collaborative modules in low-resource NER tasks.

Key words: Low-resource Named Entity Recognition, Variational Autoencoder, Conditional Encoder, Cascade Diffusion Mechanism, Adversarial Learning

摘要： 在低资源命名实体识别（Named Entity Recognition ,NER）任务中，目前许多基于迁移学习的方法虽然能够缓解数据稀缺问题，但往往会导致句子中部分正确信息的遗漏或识别错误，从而影响模型在低资源环境中的效果。针对此问题，本文提出了一种基于多模块协同的NER模型——Diff-Cascade-NER。首先，利用变分自编码器（Variational Autoencoder, VAE）在潜在空间中学习数据表示，并生成多样化的样本；接着，将上下文信息、句法分析和VAE重构数据作为条件输入到条件编码器（Conditional Encoder, CE）进行编码；然后，将编码后的数据传递给级联扩散模型（Cascade Diffusion Mechanism, CDM），通过多阶段的去噪和生成过程产生高质量样本；最后，通过对抗学习阶段（Adversarial Learning, AL）优化生成样本的质量和多样性。实验结果表明，对比现有模型，Diff-Cascade-NER在8个低资源数据集上表现优越，特别是在BC2GM和WNUT-16数据集上，F1值分别达到85.44%和56.38%，验证了各模块协同作用在低资源NER任务中的有效性。

关键词: 低资源命名实体识别, 变分自编码器, 条件编码器, 级联扩散模型, 对抗学习