A multi-layer soft lattice based model for Chinese clinical named entity recognition.

Area of Science:

• Natural Language Processing
• Medical Informatics
• Machine Learning

Background:

• Named Entity Recognition (NER) is crucial for clinical tasks like knowledge graphs and decision support.
• Long short-term memory (LSTM) models are common for clinical NER but require deep networks, long training times, and extensive data.
• These limitations hinder LSTM adoption in time-constrained clinical settings.

Purpose of the Study:

• To develop a novel Transformer-based model for improved clinical Named Entity Recognition (NER).
• To address the limitations of Long Short-Term Memory (LSTM) models in clinical NER, specifically regarding training time and data requirements.
• To enhance the model's ability to capture long-distance dependencies and complex linguistic features in clinical text.

Main Methods:

• A novel soft lattice structure Transformer model was developed, integrating BERT and a CRF module.
• The model incorporates a WordPiece module and a soft lattice structure Transformer module to process clinical text.
• This architecture is designed to effectively model long-distance dependencies and capture Chinese word lattice information.

Main Results:

• The soft lattice structure Transformer model achieved a 1-5% increase in F1 score on the CCKS NER task compared to LSTM with CRF models.
• The proposed model demonstrated reduced training time compared to existing LSTM-based approaches.
• High performance was observed in recognizing long medical terms (91.6% F-measure) and abbreviations/numbers (90.36% F-measure).

Conclusions:

• The soft lattice structure Transformer is effective for clinical NER, particularly for Chinese medical records.
• The model's ability to capture lattice information enhances its suitability for complex clinical text.
• This approach offers a more efficient and accurate solution for clinical Named Entity Recognition.