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

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Deep Learning

Background:

• Chinese Named Entity Recognition (NER) faces challenges with nested entities and ambiguous boundaries.
• Existing BiLSTM-CRF and Transformer models struggle with hierarchical structures and overlapping spans.
• Lack of morphological markers and geometric modeling hinder performance.

Purpose of the Study:

• To propose a novel neural architecture, the CEAF model, for improved Chinese NER.
• To address limitations in modeling nested entities and resolving boundary ambiguities.
• To leverage geometric deep learning for enhanced feature representation.

Main Methods:

• Developed the CEAF model, a multi-stage neural architecture for Chinese NER.
• Utilized BERT-derived subword embeddings and BiLSTM for contextual and sequential patterns.
• Introduced the Deep Context Feature Attention Module (DCAM) with capsule networks and position-aware attention.
• Incorporated the Adaptive Feature Fusion Network (AFFN) for feature integration.

Main Results:

• The CEAF model demonstrated superior performance over baseline models on multiple Chinese and English NER datasets.
• Experiments confirmed the model's effectiveness in handling nested entity structures and boundary ambiguities.
• Visualization analysis validated the model's geometric deep learning capabilities.

Conclusions:

• The CEAF model offers a significant advancement in Chinese NER, particularly for complex entity recognition.
• The integration of capsule networks and attention mechanisms provides robust solutions for hierarchical and ambiguous entity identification.
• The study highlights the potential of geometric deep learning in advancing NLP tasks.