A Multigranularity Text Driven Named Entity Recognition CGAN Model for Traditional Chinese Medicine Literatures

Area of Science:

• Computational linguistics
• Bioinformatics
• Traditional Chinese Medicine (TCM) knowledge extraction

Background:

• Extracting Traditional Chinese Medicine (TCM) knowledge from unstructured literature is crucial but challenging.
• Limited large-scale annotated data hinders conventional deep learning models for TCM Named Entity Recognition (NER).
• Existing unsupervised methods often require auxiliary data like domain dictionaries.

Purpose of the Study:

• To develop an effective Traditional Chinese Medicine Named Entity Recognition (TCM NER) model using a small-scale annotated corpus.
• To address the limitations of data scarcity in TCM text knowledge extraction.
• To improve the precision and recall of entity recognition in diverse TCM literature.

Main Methods:

• Proposed a multigranularity text-driven Named Entity Recognition (NER) model named MT-CGAN (Conditional Generation Adversarial Network).
• Designed a multigranularity text features encoder (MTFE) to capture rich semantic and grammatical information.
• Incorporated conditional constraints in the generator and discriminator and introduced seeds from different TCM text types to enhance NER precision.

Main Results:

• The MT-CGAN model demonstrated superior performance compared to baseline methods on four gold-standard datasets.
• Achieved significant improvements in precision (0.24–8.97%), recall (0.89–12.74%), and F1 score (0.01–10.84%).
• The model effectively extracts entities from various TCM literature types, especially those with more entity types and sparsity.

Conclusions:

• The proposed MT-CGAN approach offers a significant advantage for TCM NER, particularly with small-scale and diverse datasets.
• The model's ability to handle texts with higher sparsity and less regular features makes it robust.
• MT-CGAN provides an effective solution for structured TCM knowledge extraction from unstructured texts.