Hierarchical Label-Enhanced Contrastive Learning for Chinese NER.

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Machine Learning

Background:

• Character-word lattice structures show promise for Chinese Named Entity Recognition (NER) but are computationally expensive.
• Lattice-based models face challenges with inference speed and lexicon quality, potentially degrading performance.
• Noise words and limited lexicon coverage can negatively impact NER accuracy.

Purpose of the Study:

• To propose an alternative method, Hierarchical Label-Enhanced Contrastive Learning (HLCL), for Chinese NER.
• To improve entity boundary and type information integration without relying on lattice structures.
• To enhance both the efficiency and performance of Chinese NER models.

Main Methods:

• HLCL utilizes sentence-level contrastive learning (SCL) to model global mutual information between labels and sentences.
• Token-level contrastive learning (TCL) is employed to bridge the representation gap between original and label-enhanced characters.
• The method focuses on transferable label semantics and a concise model for efficient inference.

Main Results:

• HLCL demonstrates excellent efficiency and performance compared to existing lattice-based approaches.
• The proposed method effectively integrates entity boundary and type information.
• Experiments on four Chinese NER datasets validate the effectiveness of HLCL.

Conclusions:

• HLCL offers a robust and efficient alternative to lattice-based models for Chinese NER.
• The method successfully leverages label semantics and contrastive learning for improved performance.
• HLCL achieves a superb inference speed while maintaining high accuracy.