CEAF: Capsule network enhanced feature fusion architecture for Chinese Named Entity Recognition

Chinese Named Entity Recognition (NER) is a fundamental task in the field of natural language processing, where achieving deep semantic mining of nested entities and accurate disambiguation of character-level boundary ambiguities stands as its core challenge. Existing methods, mostly based on the BiLSTM-CRF sequence labeling framework or Transformer attention mechanisms, have inherent limitations in modeling the hierarchical structural dependencies of nested entities and resolving semantic conflicts in overlapping character spans. To address challenges such as the lack of morphological markers, propagation of boundary ambiguities, and insufficient geometric modeling in the feature space, we propose a novel multi-stage neural architecture—the CEAF model, a specialized neural framework tailored for Chinese NER tasks. The architecture leverages BERT-derived subword embeddings to capture character-level contextual representation and incorporates BiLSTM to model position-sensitive sequential patterns. Meanwhile, to effectively tackle the complex challenges of boundary uncertainty and nested entity composition, the CEAF model innovatively introduces the Deep Context Feature Attention Module (DCAM). This module pioneeringly integrates capsule routing protocols with position-aware attention mechanisms, processing information through dual parallel paths: on one hand, it leverages the powerful spatial relationship modeling capability of capsule networks to clearly parse the hierarchical structure and part-whole relationships between entities; on the other hand, it utilizes position-aware attention to focus on key positional information, dynamically adjust the attention to different positional information, accurately locate entity boundaries, effectively resolve boundary ambiguity, and achieve efficient and accurate modeling of nested entity structures. In addition, the Adaptive Feature Fusion Network (AFFN) effectively bridges the semantic gap between global contextual coherence and local boundary precision by selecting more discriminative fusion features. Generalization experiments on three Chinese benchmark datasets and one English dataset demonstrate that the CEAF model outperforms baseline models. Visualization analysis further verifies the modeling capability of the CEAF model, providing new insights into geometric deep learning approaches for Chinese NER.