CDD-NER: A Parameter-Efficient Cascade Dual-Stage Framework for Low-Resource Named Entity Recognition

A Dual-Granularity, Two-Stage Pointer Network for Named Entity Recognition in Chinese Social Media Short Text Named Entity Recognition (NER) constitutes a core foundational task in natural language processing. In recent years, the rapid proliferation of Chinese social media text—exemplified by micro-blogging platforms such as Weibo—has introduced distinctive challenges to conventional NER systems. Short-text scenarios are characterized by ambiguous entity boundaries, sparse contextual information, and severely imbalanced class distributions, all of which significantly degrade the recognition accuracy of existing models. To address these challenges, this work proposes a Dual-Granularity, Two-Stage Pointer Network (WBSS) that systematically improves named entity recognition performance in Chinese short-text domains. The method employs a pre-trained language model (RoBERTa-wwm-ext) as a shared feature encoder and sequentially constructs two layers of entity parsing heads upon its output representation: a top-level (coarse-grained) head covering four entity categories—Location (LOC), Person (PER), Organization (ORG), and Geo-Political Entity (GPE)—and a fine-level head that further distinguishes between Named entities (NAM) and Nominal entities (NOM) within each coarse category. At the architectural level, the first stage employs a Biaffine Span Head that explicitly models start-end dependencies through a learnable third-order tensor, formulated as $h_{s}^{\top} U h_{e}$, where $U \in \mathbb{R}^{C \times d \times d}$ is the bilinear weight tensor. This design effectively mitigates the difficulty of precisely delineating entity boundaries in short texts. The second stage incorporates an Adaptive Gated Fusion Module that soft-fuses the first-stage prediction probabilities into the second-stage contextual representation via a learnable gating network, thereby circumventing the contextual information truncation inherent in conventional Hard Mask strategies. Formally, the fused representation is computed as $h_{\text{fused}} = \alpha \odot h_{\text{BERT}} + (1 - \alpha) \odot h_{\text{prior}}$, where $\alpha$ is adaptively determined by the gate network conditioned on both the BERT features and the Stage-1 prediction probabilities. At the optimization level, the model jointly minimizes four loss components: Focal Loss, which mitigates the severe class imbalance between named and nominal entities; DIoU Loss, a differentiable boundary regression loss incorporating center-distance penalty for improved localization accuracy; and their coarse/fine-grained counterparts, combined through a weighted objective $\mathcal{L} = \lambda_1 \mathcal{L}{\text{top}}^{\text{focal}} + \lambda_2 \mathcal{L}{\text{top}}^{\text{DIoU}} + \lambda_3 \mathcal{L}{\text{fine}}^{\text{focal}} + \lambda_4 \mathcal{L}{\text{fine}}^{\text{DIoU}}$, where hyperparameter weights are determined via grid search and cross-dataset validation. In terms of training strategy, the method integrates hierarchical learning rates (pre-trained encoder: 2×10⁻⁵; newly initialized heads: 1×10⁻³), Exponential Moving Average (EMA) weight smoothing, and Feature Gradient Modification (FGM) adversarial training—applying perturbation to the embedding layer to enhance model robustness. The approach is evaluated on three publicly available Chinese NER datasets: the Weibo dataset (short-text scenario), the MSRA dataset (general news-domain NER), and the Resume dataset (structured resume text entity recognition), thereby spanning diverse domains ranging from informal social media posts to formally structured textual documents.